DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Applicant' s arguments, filed 3/10/2026, have been fully considered. The following rejections and/or objections are either reiterated or newly applied. They constitute the complete set presently being applied to the instant application.
Applicants have amended their claims, filed 3/10/2026, and therefore rejections newly made in the instant office action have been necessitated by amendment.
Claims 1-4, 8-15, and 18-25 are the currently pending claims. Claims 1-4, 8, 10, 12-15, and 18-20 have been amended. Claims 5-7 and 16-17 have been canceled. Claims 21-25 have been newly added. Claims 1-4, 8-15, 18-20, and 21-25 are hereby under examination.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 10-12, and 18 are rejected under 35 U.S.C. 103 as obvious over Hadley et al. (US 20220061710 A1), hereto referred as Hadley, and further in view of Jimenez et al. (US 20210216894 A1), hereto referred as Jimenez, and further in view of Apostolova et al. (US 20200381090 A1), hereto referred as Apostolova.
Regarding claim 1, Hadley teaches that a computer-implemented method comprises: (Hadley, ¶[0032]: “As is known in the art, the computer 200 is adapted to execute computer program modules for providing functionality described herein. A module can be implemented in hardware, firmware, and/or software. In one embodiment, program modules are stored on the storage device 230, loaded into the memory 215, and executed by the processor 205”, this shows that Hadley discloses a computer system executing software modules using a processor and memory, thereby teaching a computer-implemented method); receiving, by one or more processors, a limited set of physiological features for a user that are based on a plurality of recorded sensor values for the user, recorded during a first defined time period (Hadley, ¶[0029]: "The memory 215 holds instructions and data used by the processor 205", this shows that sensor-based biological data stored in memory is directly accessed and used by processor 205, thereby supporting that the processor receives and uses physiological features computed from recorded sensor values; ¶[0046]: "The patient health management platform 130 receives biological data 310 recorded by a variety of technical sources. Biological data 310 includes sensor data comprising biosignals recorded by one or more sensors worn or implemented by a patient", this shows that Hadley receives biosignals measured by sensors, which constitute physiological features based on a plurality of recorded sensor values; ¶[0056]: "the digital twin module 350 retrieves all biological data 310 recorded within that time period (e.g., heart rate, exercise, continuous blood glucose, ketones, blood pressure, weight) to determine the patient's actual metabolic state", this explains that Hadley derives physiological features from a defined collection time period, which corresponds to the recited first defined time period; ¶[0051]: "The digital twin module 350 implements one or more machine-learned, metabolic models to analyze the patient data 320 recorded over a given time period to generate a prediction of the patient's metabolic state for that time period", explaining that Hadley uses a defined initial collection period of physiological data and then predicts future physiological values based on the collected data; ¶[0034]: "the platform 130 records measurements of various factors... include blood sugar, triglycerides, good cholesterol (high-density lipoprotein), blood pressure, and waist circumference", this shows that a plurality of sensor types are used to obtain physiological measurement values that are collected and used as physiological features for the user (see ¶[0040] for sensors); Collectively, this shows that a plurality of sensor types are used to obtain physiological measurement values that are collected and used as physiological features for the user; to the extent the recitation of a "limited" set of physiological features is argued to impose a constraint beyond what Hadley expressly teaches, it would have been prima facie obvious to one of ordinary skill in the art to apply Hadley's predictive modeling technique to circumstances in which only a bounded initial window of physiological sensor data is available for a given user, because such circumstances are routine in clinical monitoring contexts where a patient is newly enrolled in a monitoring program, has experienced gaps in sensor wear, or where predictions are required early in a monitoring period before a full longitudinal data history has accumulated; accommodating a limited initial physiological dataset within a predictive framework is a recognized design goal in the field, and one of ordinary skill in the art would have been motivated to configure Hadley's system to operate on whatever bounded set of sensor features was available within the defined collection period); generating, by the one or more processors, one or more activity encodings for the user based on an interaction data object for the user (Hadley, ¶[0047]: “platform 130 also receives patient data 320 that is recorded manually by a patient via an application interface on a patient device 110. Patient data 320 includes nutrition data, medication data, symptom data, and lifestyle data...”, this shows that Hadley processes multiple structured patient-recorded interaction entries (nutrition, medication, symptoms, lifestyle) which serve as interaction data objects used to generate activity encodings; ¶[0080]: “The digital twin module 350 may include a combination of machine-learned models to generate various representations of a metabolic state, for example metabolic models trained to predictively model a patient’s metabolic state based on recorded nutrition data, medication data, symptom data and lifestyle data, and to model a patient’s true metabolic state based on sensor data and lab test data”, this shows that Hadley converts interaction data objects (nutrition, medication, symptoms, lifestyle) into inputs for machine-learned models, which corresponds to generating activity encodings); and inputting, by the one or more processors, the combined input feature vector to the machine learning model to produce a physiological prediction for the user based on the combined input feature vector (Hadley, ¶[0035]: “The platform determines a current metabolic state of a human body by analyzing a unique combination of continuous biosignals ... including, but not limited to, near-real-time data from wearable sensors ... periodic lab tests ... nutrition data, medicine data, and symptom data ...”; ¶[0082]: “Inputs measured by wearable sensors and lab tests or recorded manually by a patient may be encoded into a vector representation, for example a feature vector, that a machine learned model is configured to receive. A feature vector comprises an array of feature values each of which represents a measured or recorded value of an input biosignal”; ¶[0089]: “Briefly, a representation of a patient's metabolic state is generated by inputting wearable sensor data, lab test, and recorded patient data as input values to the model's function and parameters ...”; Abstract: “The platform implements a short-term prediction model to generate a daily prediction of the patient's glucose level based on nutrition data reported by the patient and sensor data and lab test data collected for the patient”, Hadley teaches inputting a feature vector comprising physiological and patient-recorded data into a machine learned model to generate a physiological prediction for the patient).
Also regarding claim 1, Hadley does not explicitly teach that the interaction data object corresponds to a second defined time period preceding the first defined time period, and the one or more activity encodings comprises one or more of: (i) a one-hot encoding of a historical medication usage of the user during the second defined time period, or (ii) a feature embedding encoding a presence of one or more activity codes within the interaction data object. Rather, Hadley teaches that patient data, including medication data, is stored as part of an ongoing timeline of entries for a current time period, such that the system retains medications taken by the patient at recorded times over that current time period (Hadley, ¶[0048]: “the patient data store 330 stores biological data 310 and patient data 330 as an ongoing recorded timeline of entries for a current time period… Accordingly, the timeline of entries stored in the patient data store 330 comprises… medications taken by the patient at recorded times over the current time period”), which shows that Hadley tracks patient medication information over time as part of the patient data used by the system. However, Hadley does not expressly teach that the interaction data object corresponds to a second defined time period preceding the first defined time period, nor does it expressly teach that the one or more activity encodings comprise a one-hot encoding of such historical medication data.
Jimenez teaches using historical patient information from a defined look-back period preceding an index time for predictive modeling, including medication-related information represented as vectorized markers. Jimenez teaches that “The look-back period for these covariates is one year” for medication use (Jimenez, ¶[0025]-¶[0038]) and further teaches that “all comorbidities, procedures, and prescriptions… present in 1 year prior to the patients’ index date can be included in the predictive modeling” and that “Each marker is associated with a vector containing the set of patients for which the marker is true or false” (Jimenez, ¶[0044]; ¶[0047]), which shows that Jimenez uses medication-related information from a preceding historical period and represents that information in binary vector form for use in prediction. Jimenez therefore teaches an interaction data object corresponding to a defined time period preceding the target prediction period and teaches encoding historical medication usage during that preceding period in vectorized binary form, which corresponds to the recited one-hot style encoding of historical medication usage during the second defined time period.
Apostolova teaches generating vector representations from historical medical-context data and using those vector representations with current physiological or structured clinical data for prediction. Apostolova teaches that “Similar approach can be taken to additional multi-dimensional EMR structured data, such as CPT codes and medication lists. Once CPT code embeddings and medication embeddings are generated, a deep learning network can be trained” (Apostolova, ¶[0019]), which shows that coded historical medical-context data, including medication information, may be represented in vector form. Apostolova further teaches that “The Patient Context Vectors obtained from available EMR ICD codes, and from free-text notes are then used in conjunction with vital signs, and lab results to predict the patient’s outcome” (Apostolova, ¶[0021]), which shows that historical context vectors may be combined with current physiological data in a predictive model. Apostolova is further relied on to show that historical coded medical-context data, including medication information and other coded historical data, may be represented in vector form, including embedded representations, and used in combination with current physiological data in predictive models.
It would have been prima facie obvious before the effective filing date of the claimed invention to have modified Hadley in view of Jimenez and Apostolova so that Hadley’s medication-related contextual data would be taken from a defined historical time period preceding the initial physiological data period, encoded in vectorized form, including at least a one-hot style encoding of historical medication usage, and used together with the physiological feature set for predictive modeling. One of ordinary skill in the art would have found it obvious to do so because Jimenez teaches that historical medication and other clinical markers from a defined look-back period provide predictive value and may be represented as binary vectors, while Apostolova teaches that historical coded medical-context data, including medication information and other coded historical data, may be represented in vector form and used together with current physiological data in machine learning prediction systems. Hadley already incorporates patient-recorded medication and lifestyle data alongside physiological sensor data, such that one of ordinary skill in the art seeking to improve Hadley’s predictive accuracy would have been directed to formalize that historical medication context into a distinct, encodable preceding time period for use with the physiological data window. Thus, the combined teachings at least render obvious the first recited alternative, which is sufficient because the claim recites that the one or more activity encodings comprise “one or more of” the listed alternatives. Applying these teachings to Hadley would have represented a predictable use of recognized feature-engineering techniques to improve the predictive performance of Hadley’s machine learning system when the initial physiological data set is limited.
Also regarding claim 1, Hadley does not fully teach that the method comprises augmenting, by the one or more processors, the limited set of physiological features with the one or more activity encodings to generate a combined input feature vector for improving a predictive performance of a machine learning model with respect to the limited set of physiological features, wherein the machine learning model is trained on a labeled training dataset comprising a historical combined input feature vector and a ground truth physiological feature recorded at a target time duration relative to the historical combined input feature vector. Rather, the modified Hadley teaches that physiological sensor data and patient-recorded data, including medication and lifestyle information, are used together as inputs to a machine-learned model (Hadley, ¶[0082]: “Inputs measured by wearable sensors and lab tests or recorded manually by a patient may be encoded into a vector representation, for example a feature vector, that a machine learned model is configured to receive. A feature vector comprises an array of feature values each of which represents a measured or recorded value of an input biosignal”; ¶[0089]: “a representation of a patient’s metabolic state is generated by inputting wearable sensor data, lab test, and recorded patient data as input values to the model's function and parameters…”), but does not expressly describe forming a combined input feature vector through an augmentation operation that incorporates encoded historical medication data from a defined preceding time period, nor does it expressly describe training a machine learning model using a labeled dataset comprising historical combined input feature vectors aligned with corresponding physiological outcomes recorded at a target time duration relative to those vectors.
Jimenez teaches extracting historical patient information from a defined look-back period preceding an index time and representing such historical information, including medication usage, as vectorized features for predictive modeling (Jimenez, ¶[0044]: "all comorbidities, procedures, and prescriptions… present in 1 year prior to the patients' index date can be included in the predictive modeling"; ¶[0047]: "Each marker is associated with a vector containing the set of patients for which the marker is true or false", explaining that Jimenez represents historical medication and clinical markers from a preceding time period as binary vector features for use in prediction).
Apostolova teaches generating vector representations of historical medical context and combining those representations with current physiological or structured clinical data as inputs to a predictive model (Apostolova, ¶[0021]: "The Patient Context Vectors obtained from available EMR ICD codes… are then used in conjunction with vital signs, and lab results to predict the patient's outcome"; ¶[0019]: "PCVs (vectors of real numbers) can be simply added to the list of existing structured data variables (vital signs and lab results) and used in a variety of… machine learning models", explaining that historical context vectors are combined with current physiological data to form a unified input representation for prediction).
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Jimenez and Apostolova so that the physiological feature set derived from sensor data is augmented with encoded historical medication information and/or activity code embeddings from a preceding time period to form a combined input feature vector, and to train the machine learning model on labeled historical instances of such combined input feature vectors aligned with corresponding physiological outcomes. It would have been obvious to train the machine learning model on such combined input feature vectors aligned with ground truth physiological features recorded at a defined target time duration relative to each historical combined vector because (1) Hadley already employs a supervised machine-learned predictive framework that generates predictions of future physiological states, which inherently requires training data in which input features are associated with physiological outcome labels; (2) Jimenez teaches that historical medication markers from a preceding period are predictive covariates, establishing that such historical information should be incorporated into the training input representation used by the predictive model; (3) Apostolova teaches combining historical vectorized medical context with current physiological data as inputs to machine learning prediction models, thereby establishing that the same combined feature representation used at inference should also be used to train the model; and (4) one of ordinary skill in the art would have understood that a supervised machine learning model must be trained on training examples that have the same feature structure as the inputs it receives at inference time, and therefore would have recognized that training the model on historical combined input feature vectors (each paired with a ground truth physiological outcome recorded at a consistent target time duration after the combined vector's physiological data window) is the standard and necessary methodology for implementing the modified Hadley system. Applying these teachings to Hadley would have been a predictable implementation in which the model is trained to map the combined historical-and-current input representation to future physiological states, thereby improving predictive performance when only a limited set of physiological data is available.
Also regarding claim 1, Hadley does not explicitly teach that the method comprises initiating, by one or more processors, the performance of a prediction-based action based on the physiological prediction, wherein the prediction-based action comprises outputting the physiological prediction to a computing interface. Rather, the modified Hadley teaches determining a predicted metabolic state and, based on that predicted metabolic state, generating a patient-specific recommendation, communicating inconsistencies to a patient device, and generating notifications or reminders to a patient, doctor, or coach (Hadley, ¶[0061]-[0063]), which shows that it initiates actions based on the physiological prediction and presents resulting information through computing interfaces. The modified Hadley further teaches patient and provider devices configured to present medically relevant data and dashboards for tracking a patient’s metabolic health (Hadley, ¶[0022]-[0024]), which shows that Hadley uses computing interfaces to present information derived from its predictive system. However, it does not expressly state that the physiological prediction itself, as distinct from downstream recommendations or notifications derived from it, is output to the computing interface.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley so that, in addition to outputting recommendations, notifications, and other information derived from the predicted metabolic state, the system also outputs the physiological prediction itself to the patient device or provider device computing interface. One of ordinary skill in the art would have found it obvious to do so because the modified Hadley already determines the predicted metabolic state, already uses that predicted metabolic state to drive downstream actions, and already provides patient and provider computing interfaces for presenting medically relevant information, such that presenting the underlying prediction to the same interface would involve only using Hadley's existing processors, existing prediction result, and existing output interfaces to display an additional piece of information already generated by the same system. The benefit of doing so would have been to improve transparency of the predictive system, allow the patient, doctor, or coach to directly review the predicted physiological state underlying the recommendation or notification, and provide more complete medically relevant information for monitoring, assessment, and follow-up decision making.
Regarding claim 10, The modified Hadley teaches that the machine learning model is previously trained on a labeled training dataset comprising a plurality of historical combined input feature vectors and a plurality of corresponding historical recorded sensor values (Hadley, ¶[0082]: "Each metabolic model is trained using a training dataset made up of large volumes of historical patient data and biological data recorded for a significant volume of patients, respectively. The training set includes daily metabolic inputs and corresponding daily metabolic outputs. Inputs, for example, include patient data 320 recorded for a current time period (i.e., different foods, medication, sleep, exercise, etc.) and a patient's initial metabolic state before the patient data 320 was recorded (e.g., based on biosignals derived from sensor data and lab test data). Inputs measured by wearable sensors and lab tests or recorded manually by a patient may be encoded into a vector representation, for example a feature vector, that a machine learned model is configured to receive. A feature vector comprises an array of feature values each of which represents a measured or recorded value of an input biosignal", this teaches that Hadley trains its machine-learned metabolic models using a training dataset comprised of historical sensor-based biosignals and patient data inputs that are encoded into feature vectors, each feature vector representing a combination of multiple physiological and activity-related inputs, i.e., a plurality of historical combined input feature vectors used to train the model; ¶[0083]: "Outputs, for example, include the actual biological data 310, which represents biosignals characterizing a patient's metabolic health (i.e., blood glucose level, blood pressure, and cholesterol). These act as baseline models trained on historical data that can then be applied to new patients with metabolic issues needing treatment to make predictions about those new patients based on what the models have learned from historical patients. Once trained, the machine learned model may be applied to predict new metabolic states for the new patients based on new combinations of biosignals to predict how a novel set of input biosignals would result in different output signals, for example lowering blood glucose to improve diabetes or lowering blood pressure to improve hypertension", this shows that the labels used in Hadley’s training dataset are actual biological data 310, such as blood glucose and blood pressure, which correspond to historical recorded sensor and lab test values paired with the input feature vectors; ¶[0085]: "FIG. 6 is an illustration of the process for training a machine-learned model to output an aspect of a patient's metabolic health, according to one embodiment. The digital twin module 350 retrieves 610 a training dataset comprised of historical biosignals (e.g., historical sensor data and lab test data) and patient measured and/or recorded for an entire a population of patients. Each historical measurement of biological data and record of patient data is assigned a time stamp representing when the patient experienced the measurement/recording and a label identifying its impact on a patient's metabolic health, the patient's metabolic response to the measurement, or both. Using the training dataset of population-level data, the digital twin module 350 trains 620 a baseline model", this explains that Hadley explicitly describes retrieving a training dataset built from historical sensor data and patient data in which each data point is assigned a label reflecting its metabolic impact or response, i.e., a labeled training dataset comprising historical recorded sensor values and associated labels used to train the machine learning model in advance).
Regarding claim 11, The modified Hadley teaches that the labeled training dataset is updated with the combined input feature vector and a plurality of corresponding future recorded sensor values for the user (Hadley, ¶[0082]: "Inputs measured by wearable sensors and lab tests or recorded manually by a patient may be encoded into a vector representation, for example a feature vector, that a machine-learned model is configured to receive. A feature vector comprises an array of feature values each of which represents a measured or recorded value of an input biosignal", this explains that Hadley forms feature vectors by encoding measured and recorded biosignals into arrays of feature values , corresponding to combined input feature vectors built from recorded sensor values and other inputs for a patient; ¶[0084]: "For example, after a metabolic state model determines an aspect of a patient's true metabolic state for a time period, the digital twin module 350 may update a training dataset with the determined true metabolic state and a plurality of biosignals recorded during the time period that contributed to the true metabolic state. The metabolic state model(s) are periodically re-trained based on the updated training dataset. This continuously improves the model and allows it to accurately predict future metabolic states for each patient based on their biosignal inputs", taken together these passages show that Hadley’s training dataset consists of labeled examples pairing biosignal feature vectors with corresponding metabolic states and that, after the model has determined a patient’s metabolic state for a given time period, the system updates the training dataset with new examples formed from the biosignals recorded during that period and their associated metabolic states and periodically retrains the model on this updated dataset , which corresponds to updating a labeled training dataset with combined input feature vectors and a plurality of corresponding future physiological measurements derived from recorded sensor values for the user and further reflects that the determined metabolic state in ¶[0084] is a future physiological outcome relative to the biosignals recorded during the preceding time period).
Regarding claim 12, Hadley teaches that a system comprising: one or more processors; and one or more memories storing processor-executable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising: (Hadley, ¶[0032]: “As is known in the art, the computer 200 is adapted to execute computer program modules for providing functionality described herein. A module can be implemented in hardware, firmware, and/or software. In one embodiment, program modules are stored on the storage device 230, loaded into the memory 215, and executed by the processor 205”, this shows that Hadley discloses a computer system executing software modules using a processor and memory, thereby teaching the claimed system architecture); receiving a limited set of physiological features for a user that are based on a plurality of recorded sensor values for the user, recorded during a first defined time period (Hadley, ¶[0029]: "The memory 215 holds instructions and data used by the processor 205", this shows that sensor-based biological data stored in memory is directly accessed and used by processor 205, thereby supporting that the processor receives and uses physiological features computed from recorded sensor values; ¶[0046]: "The patient health management platform 130 receives biological data 310 recorded by a variety of technical sources. Biological data 310 includes sensor data comprising biosignals recorded by one or more sensors worn or implemented by a patient", this shows that Hadley receives biosignals measured by sensors, which constitute physiological features based on a plurality of recorded sensor values; ¶[0056]: "the digital twin module 350 retrieves all biological data 310 recorded within that time period (e.g., heart rate, exercise, continuous blood glucose, ketones, blood pressure, weight) to determine the patient's actual metabolic state", this explains that Hadley derives physiological features from a defined collection time period, which corresponds to the recited first defined time period; ¶[0051]: "The digital twin module 350 implements one or more machine-learned, metabolic models to analyze the patient data 320 recorded over a given time period to generate a prediction of the patient's metabolic state for that time period", explaining that Hadley uses a defined initial collection period of physiological data and then predicts future physiological values based on the collected data; ¶[0034]: "the platform 130 records measurements of various factors... include blood sugar, triglycerides, good cholesterol (high-density lipoprotein), blood pressure, and waist circumference", this shows that a plurality of sensor types are used to obtain physiological measurement values that are collected and used as physiological features for the user (see ¶[0040] for sensors); Collectively, this shows that a plurality of sensor types are used to obtain physiological measurement values that are collected and used as physiological features for the user; to the extent the recitation of a "limited" set of physiological features is argued to impose a constraint beyond what Hadley expressly teaches, it would have been prima facie obvious to one of ordinary skill in the art to apply Hadley's predictive modeling technique to circumstances in which only a bounded initial window of physiological sensor data is available for a given user, because such circumstances are routine in clinical monitoring contexts where a patient is newly enrolled in a monitoring program, has experienced gaps in sensor wear, or where predictions are required early in a monitoring period before a full longitudinal data history has accumulated; accommodating a limited initial physiological dataset within a predictive framework is a recognized design goal in the field, and one of ordinary skill in the art would have been motivated to configure Hadley's system to operate on whatever bounded set of sensor features was available within the defined collection period); generating one or more activity encodings for the user based on an interaction data object for the user (Hadley, ¶[0047]: “platform 130 also receives patient data 320 that is recorded manually by a patient via an application interface on a patient device 110. Patient data 320 includes nutrition data, medication data, symptom data, and lifestyle data...”, this shows that Hadley processes multiple structured patient-recorded interaction entries (nutrition, medication, symptoms, lifestyle) which serve as interaction data objects used to generate activity encodings; ¶[0080]: “The digital twin module 350 may include a combination of machine-learned models to generate various representations of a metabolic state, for example metabolic models trained to predictively model a patient’s metabolic state based on recorded nutrition data, medication data, symptom data and lifestyle data, and to model a patient’s true metabolic state based on sensor data and lab test data”, this shows that Hadley converts interaction data objects (nutrition, medication, symptoms, lifestyle) into inputs for machine-learned models, which corresponds to generating activity encodings); inputting the combined input feature vector to the machine learning model to produce a physiological prediction for the user based on the combined input feature vector (Hadley, ¶[0035]: “The platform determines a current metabolic state of a human body by analyzing a unique combination of continuous biosignals ... including, but not limited to, near-real-time data from wearable sensors ... periodic lab tests ... nutrition data, medicine data, and symptom data ...”; ¶[0082]: “Inputs measured by wearable sensors and lab tests or recorded manually by a patient may be encoded into a vector representation, for example a feature vector, that a machine learned model is configured to receive. A feature vector comprises an array of feature values each of which represents a measured or recorded value of an input biosignal”; ¶[0089]: “Briefly, a representation of a patient's metabolic state is generated by inputting wearable sensor data, lab test, and recorded patient data as input values to the model's function and parameters ...”; Abstract: “The platform implements a short-term prediction model to generate a daily prediction of the patient's glucose level based on nutrition data reported by the patient and sensor data and lab test data collected for the patient”, Hadley teaches inputting a feature vector comprising physiological and patient-recorded data into a machine learned model to generate a physiological prediction for the patient).
Also regarding claim 12, Hadley does not explicitly teach that the interaction data object corresponds to a second defined time period preceding the first defined time period, and the one or more activity encodings comprises one or more of: (i) a one-hot encoding of a historical medication usage of the user during the second defined time period, or (ii) a feature embedding encoding a presence of one or more activity codes within the interaction data object. Rather, Hadley teaches that patient data, including medication data, is stored as part of an ongoing timeline of entries for a current time period, such that the system retains medications taken by the patient at recorded times over that current time period (Hadley, ¶[0048]: “the patient data store 330 stores biological data 310 and patient data 330 as an ongoing recorded timeline of entries for a current time period… Accordingly, the timeline of entries stored in the patient data store 330 comprises… medications taken by the patient at recorded times over the current time period”), which shows that Hadley tracks patient medication information over time as part of the patient data used by the system. However, Hadley does not expressly teach that the interaction data object corresponds to a second defined time period preceding the first defined time period, nor does it expressly teach that the one or more activity encodings comprise a one-hot encoding of such historical medication data.
Jimenez teaches using historical patient information from a defined look-back period preceding an index time for predictive modeling, including medication-related information represented as vectorized markers. Jimenez teaches that “The look-back period for these covariates is one year” for medication use (Jimenez, ¶[0025]-¶[0038]) and further teaches that “all comorbidities, procedures, and prescriptions… present in 1 year prior to the patients’ index date can be included in the predictive modeling” and that “Each marker is associated with a vector containing the set of patients for which the marker is true or false” (Jimenez, ¶[0044]; ¶[0047]), which shows that Jimenez uses medication-related information from a preceding historical period and represents that information in binary vector form for use in prediction. Jimenez therefore teaches an interaction data object corresponding to a defined time period preceding the target prediction period and teaches encoding historical medication usage during that preceding period in vectorized binary form, which corresponds to the recited one-hot style encoding of historical medication usage during the second defined time period.
Apostolova teaches generating vector representations from historical medical-context data and using those vector representations with current physiological or structured clinical data for prediction. Apostolova teaches that “Similar approach can be taken to additional multi-dimensional EMR structured data, such as CPT codes and medication lists. Once CPT code embeddings and medication embeddings are generated, a deep learning network can be trained” (Apostolova, ¶[0019]), which shows that coded historical medical-context data, including medication information, may be represented in vector form. Apostolova further teaches that “The Patient Context Vectors obtained from available EMR ICD codes, and from free-text notes are then used in conjunction with vital signs, and lab results to predict the patient’s outcome” (Apostolova, ¶[0021]), which shows that historical context vectors may be combined with current physiological data in a predictive model. Apostolova is further relied on to show that historical coded medical-context data, including medication information and other coded historical data, may be represented in vector form, including embedded representations, and used in combination with current physiological data in predictive models.
It would have been prima facie obvious before the effective filing date of the claimed invention to have modified Hadley in view of Jimenez and Apostolova so that Hadley’s medication-related contextual data would be taken from a defined historical time period preceding the initial physiological data period, encoded in vectorized form, including at least a one-hot style encoding of historical medication usage, and used together with the physiological feature set for predictive modeling. One of ordinary skill in the art would have found it obvious to do so because Jimenez teaches that historical medication and other clinical markers from a defined look-back period provide predictive value and may be represented as binary vectors, while Apostolova teaches that historical coded medical-context data, including medication information and other coded historical data, may be represented in vector form and used together with current physiological data in machine learning prediction systems. Hadley already incorporates patient-recorded medication and lifestyle data alongside physiological sensor data, such that one of ordinary skill in the art seeking to improve Hadley’s predictive accuracy would have been directed to formalize that historical medication context into a distinct, encodable preceding time period for use with the physiological data window. Thus, the combined teachings at least render obvious the first recited alternative, which is sufficient because the claim recites that the one or more activity encodings comprise “one or more of” the listed alternatives. Applying these teachings to Hadley would have represented a predictable use of recognized feature-engineering techniques to improve the predictive performance of Hadley’s machine learning system when the initial physiological data set is limited.
Also regarding claim 12, Hadley does not fully teach augmenting the limited set of physiological features with the one or more activity encodings to generate a combined input feature vector for improving a predictive performance of a machine learning model with respect to the limited set of physiological features, wherein the machine learning model is trained on a labeled training dataset comprising a historical combined input feature vector and a ground truth physiological feature recorded at a target time period relative to the historical combined input feature vector. Rather, the modified Hadley teaches that physiological sensor data and patient-recorded data, including medication and lifestyle information, are used together as inputs to a machine-learned model (Hadley, ¶[0082]: “Inputs measured by wearable sensors and lab tests or recorded manually by a patient may be encoded into a vector representation, for example a feature vector, that a machine learned model is configured to receive. A feature vector comprises an array of feature values each of which represents a measured or recorded value of an input biosignal”; ¶[0089]: “a representation of a patient’s metabolic state is generated by inputting wearable sensor data, lab test, and recorded patient data as input values to the model's function and parameters…”), but does not expressly describe forming a combined input feature vector through an augmentation operation that incorporates encoded historical medication data from a defined preceding time period, nor does it expressly describe training a machine learning model using a labeled dataset comprising historical combined input feature vectors aligned with corresponding physiological outcomes recorded at a target time period relative to those vectors.
Jimenez teaches extracting historical patient information from a defined look-back period preceding an index time and representing such historical information, including medication usage, as vectorized features for predictive modeling (Jimenez, ¶[0044]: "all comorbidities, procedures, and prescriptions… present in 1 year prior to the patients' index date can be included in the predictive modeling"; ¶[0047]: "Each marker is associated with a vector containing the set of patients for which the marker is true or false", explaining that Jimenez represents historical medication and clinical markers from a preceding time period as binary vector features for use in prediction).
Apostolova teaches generating vector representations of historical medical context and combining those representations with current physiological or structured clinical data as inputs to a predictive model (Apostolova, ¶[0021]: "The Patient Context Vectors obtained from available EMR ICD codes… are then used in conjunction with vital signs, and lab results to predict the patient's outcome"; ¶[0019]: "PCVs (vectors of real numbers) can be simply added to the list of existing structured data variables (vital signs and lab results) and used in a variety of… machine learning models", explaining that historical context vectors are combined with current physiological data to form a unified input representation for prediction).
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Jimenez and Apostolova so that the physiological feature set derived from sensor data is augmented with encoded historical medication information and/or activity code embeddings from a preceding time period to form a combined input feature vector, and to train the machine learning model on labeled historical instances of such combined input feature vectors aligned with corresponding physiological outcomes. It would have been obvious to train the machine learning model on such combined input feature vectors aligned with ground truth physiological features recorded at a defined target time period relative to each historical combined vector because (1) Hadley already employs a supervised machine-learned predictive framework that generates predictions of future physiological states, which inherently requires training data in which input features are associated with physiological outcome labels; (2) Jimenez teaches that historical medication markers from a preceding period are predictive covariates, establishing that such historical information should be incorporated into the training input representation used by the predictive model; (3) Apostolova teaches combining historical vectorized medical context with current physiological data as inputs to machine learning prediction models, thereby establishing that the same combined feature representation used at inference should also be used to train the model; and (4) one of ordinary skill in the art would have understood that a supervised machine learning model must be trained on training examples that have the same feature structure as the inputs it receives at inference time, and therefore would have recognized that training the model on historical combined input feature vectors (each paired with a ground truth physiological outcome recorded at a consistent target time period after the combined vector's physiological data window) is the standard and necessary methodology for implementing the modified Hadley system. Applying these teachings to Hadley would have been a predictable implementation in which the model is trained to map the combined historical-and-current input representation to future physiological states, thereby improving predictive performance when only a limited set of physiological data is available.
Also regarding claim 12, Hadley does not explicitly teach that the system comprises initiating the performance of a prediction-based action based on the physiological prediction, wherein the prediction-based action comprises outputting the physiological prediction to a computing interface. Rather, the modified Hadley teaches determining a predicted metabolic state and, based on that predicted metabolic state, generating a patient-specific recommendation, communicating inconsistencies to a patient device, and generating notifications or reminders to a patient, doctor, or coach (Hadley, ¶[0061]-[0063]), which shows that it initiates actions based on the physiological prediction and presents resulting information through computing interfaces. The modified Hadley further teaches patient and provider devices configured to present medically relevant data and dashboards for tracking a patient’s metabolic health (Hadley, ¶[0022]-[0024]), which shows that Hadley uses computing interfaces to present information derived from its predictive system. However, it does not expressly state that the physiological prediction itself, as distinct from downstream recommendations or notifications derived from it, is output to the computing interface.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley so that, in addition to outputting recommendations, notifications, and other information derived from the predicted metabolic state, the system also outputs the physiological prediction itself to the patient device or provider device computing interface. One of ordinary skill in the art would have found it obvious to do so because the modified Hadley already determines the predicted metabolic state, already uses that predicted metabolic state to drive downstream actions, and already provides patient and provider computing interfaces for presenting medically relevant information, such that presenting the underlying prediction to the same interface would involve only using Hadley's existing processors, existing prediction result, and existing output interfaces to display an additional piece of information already generated by the same system. The benefit of doing so would have been to improve transparency of the predictive system, allow the patient, doctor, or coach to directly review the predicted physiological state underlying the recommendation or notification, and provide more complete medically relevant information for monitoring, assessment, and follow-up decision making.
Regarding claim 18, Hadley teaches that one or more non-transitory computer-readable media storing processor executable instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising: (Hadley, FIG. 2; ¶[0028]-¶[0030]: “FIG. 2 is a high-level block diagram illustrating physical components of an example computer 200 that may be used as part of a client device (e.g., devices 110, 120, 150), application server 130, and/or database server 140 from FIG. 1, according to one embodiment ... The storage device 230 is any non-transitory computer-readable storage medium, such as a hard drive, compact disk read-only memory (CD-ROM), DVD, or a solid state memory device”; ¶[0032]: “As is known in the art, the computer 200 is adapted to execute computer program modules for providing functionality described herein. A module can be implemented in hardware, firmware, and/or software. In one embodiment, program modules are stored on the storage device 230, loaded into the memory 215, and executed by the processor 205”; ¶[0135]: “Any of the steps, operations, or processes described herein may be performed or implemented with one or more hardware or software modules ... In one embodiment, a software module is implemented with a compter program product including a computer-readable non-transitory medium containing computer program code, which can be executed by a computer processor for performing any or all of the steps, operations, or processes described”; ¶[0136]: “Embodiments of the invention may also relate to a product that is produced by a computing process described herein. Such a product may include information resulting from a computing process, where the information is stored on a non-transitory, tangible computer readable storage medium and may include any embodiment of a computer program product or other data combination described herein”, Hadley teaches non-transitory computer-readable media storing instructions that are executed by a processor to perform the disclosed operations); receiving a limited set of physiological features for a user that are based on a plurality of recorded sensor values for the user, recorded during a first defined time period (Hadley, ¶[0029]: "The memory 215 holds instructions and data used by the processor 205", this shows that sensor-based biological data stored in memory is directly accessed and used by processor 205, thereby supporting that the processor receives and uses physiological features computed from recorded sensor values; ¶[0046]: "The patient health management platform 130 receives biological data 310 recorded by a variety of technical sources. Biological data 310 includes sensor data comprising biosignals recorded by one or more sensors worn or implemented by a patient", this shows that Hadley receives biosignals measured by sensors, which constitute physiological features based on a plurality of recorded sensor values; ¶[0056]: "the digital twin module 350 retrieves all biological data 310 recorded within that time period (e.g., heart rate, exercise, continuous blood glucose, ketones, blood pressure, weight) to determine the patient's actual metabolic state", this explains that Hadley derives physiological features from a defined collection time period, which corresponds to the recited first defined time period; ¶[0051]: "The digital twin module 350 implements one or more machine-learned, metabolic models to analyze the patient data 320 recorded over a given time period to generate a prediction of the patient's metabolic state for that time period", explaining that Hadley uses a defined initial collection period of physiological data and then predicts future physiological values based on the collected data; ¶[0034]: "the platform 130 records measurements of various factors... include blood sugar, triglycerides, good cholesterol (high-density lipoprotein), blood pressure, and waist circumference", this shows that a plurality of sensor types are used to obtain physiological measurement values that are collected and used as physiological features for the user (see ¶[0040] for sensors); Collectively, this shows that a plurality of sensor types are used to obtain physiological measurement values that are collected and used as physiological features for the user; to the extent the recitation of a "limited" set of physiological features is argued to impose a constraint beyond what Hadley expressly teaches, it would have been prima facie obvious to one of ordinary skill in the art to apply Hadley's predictive modeling technique to circumstances in which only a bounded initial window of physiological sensor data is available for a given user, because such circumstances are routine in clinical monitoring contexts where a patient is newly enrolled in a monitoring program, has experienced gaps in sensor wear, or where predictions are required early in a monitoring period before a full longitudinal data history has accumulated; accommodating a limited initial physiological dataset within a predictive framework is a recognized design goal in the field, and one of ordinary skill in the art would have been motivated to configure Hadley's system to operate on whatever bounded set of sensor features was available within the defined collection period); generating one or more activity encodings for the user based on an interaction data object for the user (Hadley, ¶[0047]: “platform 130 also receives patient data 320 that is recorded manually by a patient via an application interface on a patient device 110. Patient data 320 includes nutrition data, medication data, symptom data, and lifestyle data...”, this shows that Hadley processes multiple structured patient-recorded interaction entries (nutrition, medication, symptoms, lifestyle) which serve as interaction data objects used to generate activity encodings; ¶[0080]: “The digital twin module 350 may include a combination of machine-learned models to generate various representations of a metabolic state, for example metabolic models trained to predictively model a patient’s metabolic state based on recorded nutrition data, medication data, symptom data and lifestyle data, and to model a patient’s true metabolic state based on sensor data and lab test data”, this shows that Hadley converts interaction data objects (nutrition, medication, symptoms, lifestyle) into inputs for machine-learned models, which corresponds to generating activity encodings); inputting the combined input feature vector to the machine learning model to produce a physiological prediction for the user based on the combined input feature vector (Hadley, ¶[0035]: “The platform determines a current metabolic state of a human body by analyzing a unique combination of continuous biosignals ... including, but not limited to, near-real-time data from wearable sensors ... periodic lab tests ... nutrition data, medicine data, and symptom data ...”; ¶[0082]: “Inputs measured by wearable sensors and lab tests or recorded manually by a patient may be encoded into a vector representation, for example a feature vector, that a machine learned model is configured to receive. A feature vector comprises an array of feature values each of which represents a measured or recorded value of an input biosignal”; ¶[0089]: “Briefly, a representation of a patient’s metabolic state is generated by inputting wearable sensor data, lab test, and recorded patient data as input values to the model’s function and parameters ...”; Abstract: “The platform implements a short-term prediction model to generate a daily prediction of the patient’s glucose level based on nutrition data reported by the patient and sensor data and lab test data collected for the patient”, Hadley teaches inputting a feature vector comprising physiological and patient-recorded data into a machine learned model to generate a physiological prediction for the patient).
Also regarding claim 18, Hadley does not explicitly teach that the interaction data object corresponds to a second defined time period preceding the first defined time period, and the one or more activity encodings comprises one or more of: (i) a one-hot encoding of a historical medication usage of the user during the second defined time period, or (ii) a feature embedding encoding a presence of one or more activity codes within the interaction data object. Rather, Hadley teaches that patient data, including medication data, is stored as part of an ongoing timeline of entries for a current time period, such that the system retains medications taken by the patient at recorded times over that current time period (Hadley, ¶[0048]: “the patient data store 330 stores biological data 310 and patient data 330 as an ongoing recorded timeline of entries for a current time period… Accordingly, the timeline of entries stored in the patient data store 330 comprises… medications taken by the patient at recorded times over the current time period”), which shows that Hadley tracks patient medication information over time as part of the patient data used by the system. However, Hadley does not expressly teach that the interaction data object corresponds to a second defined time period preceding the first defined time period, nor does it expressly teach that the one or more activity encodings comprise a one-hot encoding of such historical medication data.
Jimenez teaches using historical patient information from a defined look-back period preceding an index time for predictive modeling, including medication-related information represented as vectorized markers. Jimenez teaches that “The look-back period for these covariates is one year” for medication use (Jimenez, ¶[0025]-¶[0038]) and further teaches that “all comorbidities, procedures, and prescriptions… present in 1 year prior to the patients’ index date can be included in the predictive modeling” and that “Each marker is associated with a vector containing the set of patients for which the marker is true or false” (Jimenez, ¶[0044]; ¶[0047]), which shows that Jimenez uses medication-related information from a preceding historical period and represents that information in binary vector form for use in prediction. Jimenez therefore teaches an interaction data object corresponding to a defined time period preceding the target prediction period and teaches encoding historical medication usage during that preceding period in vectorized binary form, which corresponds to the recited one-hot style encoding of historical medication usage during the second defined time period.
Apostolova teaches generating vector representations from historical medical-context data and using those vector representations with current physiological or structured clinical data for prediction. Apostolova teaches that “Similar approach can be taken to additional multi-dimensional EMR structured data, such as CPT codes and medication lists. Once CPT code embeddings and medication embeddings are generated, a deep learning network can be trained” (Apostolova, ¶[0019]), which shows that coded historical medical-context data, including medication information, may be represented in vector form. Apostolova further teaches that “The Patient Context Vectors obtained from available EMR ICD codes, and from free-text notes are then used in conjunction with vital signs, and lab results to predict the patient’s outcome” (Apostolova, ¶[0021]), which shows that historical context vectors may be combined with current physiological data in a predictive model. Apostolova is further relied on to show that historical coded medical-context data, including medication information and other coded historical data, may be represented in vector form, including embedded representations, and used in combination with current physiological data in predictive models.
It would have been prima facie obvious before the effective filing date of the claimed invention to have modified Hadley in view of Jimenez and Apostolova so that Hadley’s medication-related contextual data would be taken from a defined historical time period preceding the initial physiological data period, encoded in vectorized form, including at least a one-hot style encoding of historical medication usage, and used together with the physiological feature set for predictive modeling. One of ordinary skill in the art would have found it obvious to do so because Jimenez teaches that historical medication and other clinical markers from a defined look-back period provide predictive value and may be represented as binary vectors, while Apostolova teaches that historical coded medical-context data, including medication information and other coded historical data, may be represented in vector form and used together with current physiological data in machine learning prediction systems. Hadley already incorporates patient-recorded medication and lifestyle data alongside physiological sensor data, such that one of ordinary skill in the art seeking to improve Hadley’s predictive accuracy would have been directed to formalize that historical medication context into a distinct, encodable preceding time period for use with the physiological data window. Thus, the combined teachings at least render obvious the first recited alternative, which is sufficient because the claim recites that the one or more activity encodings comprise “one or more of” the listed alternatives. Applying these teachings to Hadley would have represented a predictable use of recognized feature-engineering techniques to improve the predictive performance of Hadley’s machine learning system when the initial physiological data set is limited.
Also regarding claim 18, Hadley does not fully teach that augmenting the limited set of physiological features with the one or more activity encodings to generate a combined input feature vector for improving a predictive performance of a machine learning model with respect to the limited set of physiological features, wherein the machine learning model is trained on a labeled training dataset comprising a historical combined input feature vector and a ground truth physiological feature recorded at a target time period relative to the historical combined input feature vector. Rather, the modified Hadley teaches that physiological sensor data and patient-recorded data, including medication and lifestyle information, are used together as inputs to a machine-learned model (Hadley, ¶[0082]: “Inputs measured by wearable sensors and lab tests or recorded manually by a patient may be encoded into a vector representation, for example a feature vector, that a machine learned model is configured to receive. A feature vector comprises an array of feature values each of which represents a measured or recorded value of an input biosignal”; ¶[0089]: “a representation of a patient’s metabolic state is generated by inputting wearable sensor data, lab test, and recorded patient data as input values to the model's function and parameters…”), but does not expressly describe forming a combined input feature vector through an augmentation operation that incorporates encoded historical medication data from a defined preceding time period, nor does it expressly describe training a machine learning model using a labeled dataset comprising historical combined input feature vectors aligned with corresponding physiological outcomes recorded at a target time period relative to those vectors.
Jimenez teaches extracting historical patient information from a defined look-back period preceding an index time and representing such historical information, including medication usage, as vectorized features for predictive modeling (Jimenez, ¶[0044]: “all comorbidities, procedures, and prescriptions… present in 1 year prior to the patients' index date can be included in the predictive modeling”; ¶[0047]: “Each marker is associated with a vector containing the set of patients for which the marker is true or false”, explaining that Jimenez represents historical medication and clinical markers from a preceding time period as binary vector features for use in prediction).
Apostolova teaches generating vector representations of historical medical context and combining those representations with current physiological or structured clinical data as inputs to a predictive model (Apostolova, ¶[0021]: “The Patient Context Vectors obtained from available EMR ICD codes… are then used in conjunction with vital signs, and lab results to predict the patient's outcome”; ¶[0019]: “PCVs (vectors of real numbers) can be simply added to the list of existing structured data variables (vital signs and lab results) and used in a variety of… machine learning models”, explaining that historical context vectors are combined with current physiological data to form a unified input representation for prediction).
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Jimenez and Apostolova so that the physiological feature set derived from sensor data is augmented with encoded historical medication information and/or activity code embeddings from a preceding time period to form a combined input feature vector, and to train the machine learning model on labeled historical instances of such combined input feature vectors aligned with corresponding physiological outcomes. It would have been obvious to train the machine learning model on such combined input feature vectors aligned with ground truth physiological features recorded at a defined target time period relative to each historical combined vector because (1) Hadley already employs a supervised machine-learned predictive framework that generates predictions of future physiological states, which inherently requires training data in which input features are associated with physiological outcome labels; (2) Jimenez teaches that historical medication markers from a preceding period are predictive covariates, establishing that such historical information should be incorporated into the training input representation used by the predictive model; (3) Apostolova teaches combining historical vectorized medical context with current physiological data as inputs to machine learning prediction models, thereby establishing that the same combined feature representation used at inference should also be used to train the model; and (4) one of ordinary skill in the art would have understood that a supervised machine learning model must be trained on training examples that have the same feature structure as the inputs it receives at inference time, and therefore would have recognized that training the model on historical combined input feature vectors (each paired with a ground truth physiological outcome recorded at a consistent target time period after the combined vector’s physiological data window) is the standard and necessary methodology for implementing the modified Hadley system. Applying these teachings to Hadley would have been a predictable implementation in which the model is trained to map the combined historical-and-current input representation to future physiological states, thereby improving predictive performance when only a limited set of physiological data is available.
Also regarding claim 18, Hadley does not explicitly teach that initiating the performance of a prediction-based action based on the physiological prediction, wherein the prediction-based action comprises outputting the physiological prediction to a computing interface. Rather, the modified Hadley teaches determining a predicted metabolic state and, based on that predicted metabolic state, generating a patient-specific recommendation, communicating inconsistencies to a patient device, and generating notifications or reminders to a patient, doctor, or coach (Hadley, ¶[0061]-[0063]), which shows that it initiates actions based on the physiological prediction and presents resulting information through computing interfaces. The modified Hadley further teaches patient and provider devices configured to present medically relevant data and dashboards for tracking a patient’s metabolic health (Hadley, ¶[0022]-[0024]), which shows that Hadley uses computing interfaces to present information derived from its predictive system. However, it does not expressly state that the physiological prediction itself, as distinct from downstream recommendations or notifications derived from it, is output to the computing interface.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley so that, in addition to outputting recommendations, notifications, and other information derived from the predicted metabolic state, the system also outputs the physiological prediction itself to the patient device or provider device computing interface. One of ordinary skill in the art would have found it obvious to do so because the modified Hadley already determines the predicted metabolic state, already uses that predicted metabolic state to drive downstream actions, and already provides patient and provider computing interfaces for presenting medically relevant information, such that presenting the underlying prediction to the same interface would involve only using Hadley’s existing processors, existing prediction result, and existing output interfaces to display an additional piece of information already generated by the same system. The benefit of doing so would have been to improve transparency of the predictive system, allow the patient, doctor, or coach to directly review the predicted physiological state underlying the recommendation or notification, and provide more complete medically relevant information for monitoring, assessment, and follow-up decision making.
Claims 2, 13, and 21 are rejected under 35 U.S.C. 103 as obvious over Hadley et al. (US 20220061710 A1), hereto referred as Hadley, and further in view of Jimenez et al. (US 20210216894 A1), hereto referred as Jimenez, and further in view of Apostolova et al. (US 20200381090 A1) , hereto referred as Apostolova, and further in view of Bergenstal et al. (Bergenstal, Richard M et al. “Recommendations for Standardizing Glucose Reporting and Analysis to Optimize Clinical Decision Making in Diabetes: The Ambulatory Glucose Profile (AGP).” Diabetes technology & therapeutics 15.3 (2013)) , hereto referred as Bergenstal, and further in view of Frank et al. (US 20210401330 A1) , hereto referred as Frank, as evidence.
The modified Hadley teaches claim 1 as described above. The modified Hadley teaches claim 12 as described above. The modified Hadley teaches claim 18 as described above.
Regarding claim 2, the modified Hadley teaches that the limited set of physiological features comprises an aggregated glucose value (Hadley, ¶[0092]: “The glucose twin module 515 includes a short-term prediction module 810 and a long-term prediction module 825”; ¶[0093]: “both the short-term prediction module 810 and the long-term prediction module 825 apply a training dataset of historical blood glucose data from a population of patients”; ¶[0126]: “The glucose twin module 651 determines 866 an aggregate estimate of the patient's A1c for a rolling 10-day blood glucose average by concatenating the prediction output by the first model with the prediction output by the second model”, Hadley teaches use of glucose-related features derived from historical blood glucose data and expressly teaches an aggregate estimate based on a rolling 10-day blood glucose average, which corresponds to an aggregated glucose value).
Also regarding claim 2, the modified Hadley does not explicitly teach that the limited set of physiological features comprises an aggregated glucose variability value. Rather, Hadley teaches using historical blood glucose data, sequential features, and static features for glucose prediction and aggregate glucose estimation (Hadley, ¶[0092]-¶[0093]; ¶[0126]), but does not expressly disclose that one of the physiological features in the limited set is an aggregated glucose variability value.
Dave teaches that glucose-prediction feature sets may include variability-based glucose features derived from CGM measurements. Dave teaches that “Cichosz et al used variability-based features extracted from CGM readings in a 30-minute window” and that “The authors also included glycemic variability at specific intervals during the night as well as static contextual information” (Dave, p. 843, 'Introduction'). Dave further teaches that its extracted features include “sd_2hr Standard deviation of CGM observations observed in the past two hours”, “sd_4hr Standard deviation of CGM observations observed in the past four hours”, and “SD Standard deviation of all CGM observations across the patient” (Dave, Table 2, p. 844), which shows that glucose variability values aggregated over multiple CGM observations and time periods were used as predictive physiological features.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Dave so that the limited set of physiological features used by Hadley’s glucose prediction framework further comprises an aggregated glucose variability value in addition to the aggregated glucose value. One of ordinary skill in the art would have found it obvious to do so because Dave teaches that variability-based glucose features, including standard deviation values aggregated over sets of CGM observations and over defined time windows, are used as predictive features for glucose modeling, thereby demonstrating that glucose variability provides additional predictive information beyond absolute glucose levels alone. Hadley already relies on historical blood glucose data and aggregated glucose information as inputs to a machine-learning prediction model. One of ordinary skill in the art would have recognized that incorporating variability-based glucose features into Hadley’s existing feature set would have been a straightforward and compatible modification, as both references operate on CGM-derived glucose data and use machine-learning frameworks for prediction. Further, it would have been understood that aggregated glucose values capture central tendency (e.g., average glucose levels), whereas variability features capture fluctuations and dispersion in glucose measurements, and that combining both types of features improves model robustness and predictive accuracy. Therefore, incorporating Dave’s aggregated glucose variability features into Hadley would have been a predictable use of known feature-engineering techniques to enhance predictive performance by enabling the model to account for both baseline glucose levels and temporal variability in glucose behavior.
Regarding claim 13, the modified Hadley teaches that the limited set of physiological features comprises an aggregated glucose value (Hadley, ¶[0092]: “The glucose twin module 515 includes a short-term prediction module 810 and a long-term prediction module 825”; ¶[0093]: “both the short-term prediction module 810 and the long-term prediction module 825 apply a training dataset of historical blood glucose data from a population of patients”; ¶[0126]: “The glucose twin module 651 determines 866 an aggregate estimate of the patient's A1c for a rolling 10-day blood glucose average by concatenating the prediction output by the first model with the prediction output by the second model”, Hadley teaches use of glucose-related features derived from historical blood glucose data and expressly teaches an aggregate estimate based on a rolling 10-day blood glucose average, which corresponds to an aggregated glucose value).
Also regarding claim 13, the modified Hadley does not explicitly teach that the limited set of physiological features comprises an aggregated glucose variability value. Rather, Hadley teaches using historical blood glucose data, sequential features, and static features for glucose prediction and aggregate glucose estimation (Hadley, ¶[0092]-¶[0093]; ¶[0126]), but does not expressly disclose that one of the physiological features in the limited set is an aggregated glucose variability value.
Dave teaches that glucose-prediction feature sets may include variability-based glucose features derived from CGM measurements. Dave teaches that “Cichosz et al used variability-based features extracted from CGM readings in a 30-minute window” and that “The authors also included glycemic variability at specific intervals during the night as well as static contextual information” (Dave, p. 843, 'Introduction'). Dave further teaches that its extracted features include “sd_2hr Standard deviation of CGM observations observed in the past two hours”, “sd_4hr Standard deviation of CGM observations observed in the past four hours”, and “SD Standard deviation of all CGM observations across the patient” (Dave, Table 2, p. 844), which shows that glucose variability values aggregated over multiple CGM observations and time periods were used as predictive physiological features.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Dave so that the limited set of physiological features used by Hadley’s glucose prediction framework further comprises an aggregated glucose variability value in addition to the aggregated glucose value. One of ordinary skill in the art would have found it obvious to do so because Dave teaches that variability-based glucose features, including standard deviation values aggregated over sets of CGM observations and over defined time windows, are used as predictive features for glucose modeling, thereby demonstrating that glucose variability provides additional predictive information beyond absolute glucose levels alone. Hadley already relies on historical blood glucose data and aggregated glucose information as inputs to a machine-learning prediction model. One of ordinary skill in the art would have recognized that incorporating variability-based glucose features into Hadley’s existing feature set would have been a straightforward and compatible modification, as both references operate on CGM-derived glucose data and use machine-learning frameworks for prediction. Further, it would have been understood that aggregated glucose values capture central tendency (e.g., average glucose levels), whereas variability features capture fluctuations and dispersion in glucose measurements, and that combining both types of features improves model robustness and predictive accuracy. Therefore, incorporating Dave’s aggregated glucose variability features into Hadley would have been a predictable use of known feature-engineering techniques to enhance predictive performance by enabling the model to account for both baseline glucose levels and temporal variability in glucose behavior.
Regarding claim 21, the modified Hadley teaches that the limited set of physiological features comprises an aggregated glucose value (Hadley, ¶[0092]: “The glucose twin module 515 includes a short-term prediction module 810 and a long-term prediction module 825”; ¶[0093]: “both the short-term prediction module 810 and the long-term prediction module 825 apply a training dataset of historical blood glucose data from a population of patients”; ¶[0126]: “The glucose twin module 651 determines 866 an aggregate estimate of the patient's A1c for a rolling 10-day blood glucose average by concatenating the prediction output by the first model with the prediction output by the second model”, Hadley teaches use of glucose-related features derived from historical blood glucose data and expressly teaches an aggregate estimate based on a rolling 10-day blood glucose average, which corresponds to an aggregated glucose value).
Also regarding claim 21, the modified Hadley does not explicitly teach that the limited set of physiological features comprises an aggregated glucose variability value. Rather, Hadley teaches using historical blood glucose data, sequential features, and static features for glucose prediction and aggregate glucose estimation (Hadley, ¶[0092]-¶[0093]; ¶[0126]), but does not expressly disclose that one of the physiological features in the limited set is an aggregated glucose variability value.
Dave teaches that glucose-prediction feature sets may include variability-based glucose features derived from CGM measurements. Dave teaches that “Cichosz et al used variability-based features extracted from CGM readings in a 30-minute window” and that “The authors also included glycemic variability at specific intervals during the night as well as static contextual information” (Dave, p. 843, 'Introduction'). Dave further teaches that its extracted features include “sd_2hr Standard deviation of CGM observations observed in the past two hours”, “sd_4hr Standard deviation of CGM observations observed in the past four hours”, and “SD Standard deviation of all CGM observations across the patient” (Dave, Table 2, p. 844), which shows that glucose variability values aggregated over multiple CGM observations and time periods were used as predictive physiological features.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Dave so that the limited set of physiological features used by Hadley’s glucose prediction framework further comprises an aggregated glucose variability value in addition to the aggregated glucose value. One of ordinary skill in the art would have found it obvious to do so because Dave teaches that variability-based glucose features, including standard deviation values aggregated over sets of CGM observations and over defined time windows, are used as predictive features for glucose modeling, thereby demonstrating that glucose variability provides additional predictive information beyond absolute glucose levels alone. Hadley already relies on historical blood glucose data and aggregated glucose information as inputs to a machine-learning prediction model. One of ordinary skill in the art would have recognized that incorporating variability-based glucose features into Hadley’s existing feature set would have been a straightforward and compatible modification, as both references operate on CGM-derived glucose data and use machine-learning frameworks for prediction. Further, it would have been understood that aggregated glucose values capture central tendency (e.g., average glucose levels), whereas variability features capture fluctuations and dispersion in glucose measurements, and that combining both types of features improves model robustness and predictive accuracy. Therefore, incorporating Dave’s aggregated glucose variability features into Hadley would have been a predictable use of known feature-engineering techniques to enhance predictive performance by enabling the model to account for both baseline glucose levels and temporal variability in glucose behavior.
Claims 3-4, 14-15, and 22-23 are rejected under 35 U.S.C. 103 as obvious over Hadley et al. (US 20220061710 A1), hereto referred as Hadley, and further in view of Jimenez et al. (US 20210216894 A1), hereto referred as Jimenez, and further in view of Apostolova et al. (US 20200381090 A1) , hereto referred as Apostolova, and further in view of Bergenstal et al. (Bergenstal, Richard M et al. “Recommendations for Standardizing Glucose Reporting and Analysis to Optimize Clinical Decision Making in Diabetes: The Ambulatory Glucose Profile (AGP).” Diabetes technology & therapeutics 15.3 (2013)) , hereto referred as Bergenstal, and further in view of Frank et al. (US 20210401330 A1) , hereto referred as Frank, as evidence, and further in view of Dave et al. (Dave, Darpit et al. “Feature-Based Machine Learning Model for Real-Time Hypoglycemia Prediction.” Journal of diabetes science and technology 15.4 (2021)) , hereto referred as Dave.
The modified Hadley teaches claim 1 as described above. The modified Hadley teaches claim 12 as described above. The modified Hadley teaches claim 18 as described above.
Regarding claim 3, the modified Hadley teaches that the aggregated glucose value comprises glucose values aggregated from sensor measurements over a defined period of time (Hadley, ¶[0126]: “The glucose twin module 651 determines 866 an aggregate estimate of the patient's A1c for a rolling 10-day blood glucose average by concatenating the prediction output by the first model with the prediction output by the second model”, Hadley teaches an aggregated glucose value based on blood glucose information over a multi-day period; see also Hadley, ¶[0059]: “The illustrated interface displays biological data recorded by wearable devices over a period of time including signal curves of 5-day average blood glucose measurements (5DG CGM) 391, 1-day average blood glucose measurements (1DG-CGM) 392 ... Each point on the signal curve represents an average value of the signal measured on that day. For example, each point along the signal curve of 1DG-CGM 392 measurements represents a patient's 1-day average glucose for a given day”, Hadley further teaches generating daily aggregated glucose values from multiple recorded sensor measurements for a user over time).
Also regarding claim 3, the modified Hadley does not explicitly teach that the aggregated glucose value comprises an arithmetic average of a plurality of daily median recorded sensor measurements for the user over the first defined time period. Rather, Hadley teaches aggregated glucose values such as rolling blood glucose averages and 1-day average glucose values derived from multiple recorded sensor measurements, but does not expressly disclose using daily median recorded sensor measurements as the intermediate daily values that are then arithmetically averaged over the first defined time period.
Bergenstal teaches CGM data summarization using median-based statistical representations of glucose measurements derived from repeated daily CGM recordings across a multi-day period. Specifically, Bergenstal teaches that the Ambulatory Glucose Profile (AGP) presents “smoothed curves representing the median (50th) … frequency percentiles” computed from glucose values collected across multiple days (Bergenstal, p. 206-207, Part 2), which shows that within the CGM data analysis field, the median is an established central-tendency statistic used to summarize glucose measurements derived from sensor readings. Bergenstal further teaches that “modal day displays of glucose medians, peaks, troughs, and variability were highly reflective of what the display would look like after 30 days of CGM use” (Bergenstal, p. 203-204, Proposed AGP), which shows that median-based glucose summaries derived from multi-day CGM data are recognized as stable and representative summaries of a patient’s glucose behavior across time. Bergenstal thereby establishes that the median is an art-recognized statistic for summarizing CGM sensor measurements into a representative glucose value, motivated by its robustness to transient excursions and intraday noise.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Bergenstal so that the aggregated glucose value used in Hadley’s feature set comprises an arithmetic average of a plurality of daily median recorded sensor measurements for the user over the first defined time period. One of ordinary skill in the art would have found it obvious to do so because Bergenstal establishes that the median is a preferred statistic for representing CGM-derived glucose measurements when robustness to transient excursions and intraday sensor noise is desired, while Hadley teaches forming aggregated glucose features using a daily representative glucose value arithmetically averaged across a multi-day period. One of ordinary skill in the art implementing Hadley’s per-day glucose aggregation step would have had a clear, articulable reason in view of Bergenstal to select the intraday median of each day’s recorded sensor measurements as the per-day representative value, rather than another summary statistic, because Bergenstal teaches that the median is more robust than mean-based measures for characterizing CGM glucose behavior in the presence of intraday excursions and noise. Having selected the daily median as the per-day representative value in view of Bergenstal, the subsequent step of arithmetically averaging those daily median values across the plurality of days in the first defined time period follows directly from Hadley’s existing multi-day averaging methodology. The result, an arithmetic average of daily median sensor readings, represents a straightforward and predictable combination in which the median is applied at the within-day level to suppress intraday noise and transient outliers, while the arithmetic average is applied across days to produce a stable multi-day summary of central glucose tendency. Therefore, incorporating Bergenstal’s median-based CGM summarization into the per-day aggregation step of Hadley’s multi-day averaging framework would have been a predictable use of known CGM statistical analysis techniques, motivated by improving robustness of the daily representative glucose value, to arrive at the claimed arithmetic average of daily median recorded sensor measurements. Further, Frank confirms that median glucose values and daily aggregated glucose features derived from CGM measurements were well-established inputs in glucose-prediction machine learning systems (Frank, ¶[0096]: "may also or alternately include one or more of mean glucose (e.g., over the duration of the observation period or daily), median glucose, inter quartile range of the glucose measurements 110, variance of the glucose measurements 110 ..."), showing that median glucose statistics and daily aggregated glucose measures were recognized and routinely used as input features in predictive glucose modeling.
Regarding claim 4, the modified Hadley teaches that the limited set of physiological features comprises an aggregated glucose variability value, as established in the rejection of claim 2 above.
Also regarding claim 4, the modified Hadley does not explicitly teach that the aggregated glucose variability value comprises an arithmetic average of a plurality of daily median glycemic variability measurements for the user over the first defined time period. Rather, the claim 2 art teaches aggregated glucose variability features derived from CGM data, but does not expressly disclose first computing a daily median of the glycemic variability measurements within each day and then arithmetically averaging those daily median values across the first defined time period.
Bergenstal teaches that median-based CGM summarization is a recognized technique for generating representative glucose-related statistics from repeated daily CGM data across a multi-day period. Specifically, Bergenstal teaches that the Ambulatory Glucose Profile (AGP) presents “smoothed curves representing the median (50th) … frequency percentiles” computed from glucose values collected across multiple days (Bergenstal, p. 206-207, Part 2), which shows that median-based summarization is an established statistical technique in CGM data analysis for generating representative values from repeated measurements. Bergenstal further teaches that “modal day displays of glucose medians, peaks, troughs, and variability were highly reflective of what the display would look like after 30 days of CGM use” (Bergenstal, p. 203-204, Proposed AGP), which confirms that median-based summarization is recognized as a stable and representative technique for characterizing both glucose levels and variability across a multi-day CGM period.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Dave and Bergenstal so that the aggregated glucose variability value used in Hadley’s feature set comprises an arithmetic average of a plurality of daily median glycemic variability measurements for the user over the first defined time period. One of ordinary skill in the art would have found it obvious to do so for the following reasons. First, Dave teaches generating glycemic variability measurements, specifically rolling-window standard deviations such as sd_2hr and sd_4hr, computed at regular CGM reading intervals throughout each day, which produces an intraday time-series of glycemic variability values within each day of the first defined time period. Second, one of ordinary skill in the art implementing Hadley’s aggregated glucose variability feature would have recognized that those intraday variability values are subject to the same transient excursions and within-day noise as intraday glucose readings, and would therefore have been motivated, in view of Bergenstal’s teaching that the median is a preferred robust summary statistic for CGM-derived data, to select the daily median of each day’s intraday variability measurements as the per-day representative variability value. Third, a person of ordinary skill in the art would have applied the same median-then-average methodology to the glucose variability feature as a direct and obvious parallel, because both the glucose feature and the glucose variability feature are derived from the same intraday CGM time-series, are susceptible to the same sources of intraday noise, and are aggregated across the same multi-day time period, such that using a consistent per-day summarization methodology for both features would have been an obvious design choice. The subsequent step of arithmetically averaging those daily median variability values across the plurality of days in the first defined time period follows directly from Hadley’s use of multi-day aggregated feature values. The result, an arithmetic average of daily median glycemic variability measurements, represents a predictable combination in which the median suppresses transient within-day fluctuations at the intraday level, while the arithmetic average produces a stable multi-day summary of glycemic variability. Therefore, incorporating Dave’s intraday glycemic variability measurements into Hadley’s feature set and applying Bergenstal’s median-based CGM summarization methodology to derive daily representative variability values would have been a predictable use of known CGM statistical analysis techniques, motivated by the same robustness considerations identified for claim 22 and by the obvious design symmetry of applying consistent aggregation methodology to parallel feature types. Further, Frank confirms that median glucose values and daily aggregated CGM-derived statistics, including variance of glucose measurements, were well-established inputs in glucose-prediction machine learning systems as of the filing date (Frank, ¶[0096]: “may also or alternately include one or more of mean glucose (e.g., over the duration of the observation period or daily), median glucose, inter quartile range of the glucose measurements 110, variance of the glucose measurements 110 ...”), showing that both median-based and variance-based CGM statistics were recognized and routinely used as input features in predictive glucose modeling frameworks.
Regarding claim 14, the modified Hadley teaches that the aggregated glucose value comprises glucose values aggregated from sensor measurements over a defined period of time (Hadley, ¶[0126]: “The glucose twin module 651 determines 866 an aggregate estimate of the patient's A1c for a rolling 10-day blood glucose average by concatenating the prediction output by the first model with the prediction output by the second model”, Hadley teaches an aggregated glucose value based on blood glucose information over a multi-day period; see also Hadley, ¶[0059]: “The illustrated interface displays biological data recorded by wearable devices over a period of time including signal curves of 5-day average blood glucose measurements (5DG CGM) 391, 1-day average blood glucose measurements (1DG-CGM) 392 ... Each point on the signal curve represents an average value of the signal measured on that day. For example, each point along the signal curve of 1DG-CGM 392 measurements represents a patient's 1-day average glucose for a given day”, Hadley further teaches generating daily aggregated glucose values from multiple recorded sensor measurements for a user over time).
Also regarding claim 14, the modified Hadley does not explicitly teach that the aggregated glucose value comprises an arithmetic average of a plurality of daily median recorded sensor measurements for the user over the first defined time period. Rather, Hadley teaches aggregated glucose values such as rolling blood glucose averages and 1-day average glucose values derived from multiple recorded sensor measurements, but does not expressly disclose using daily median recorded sensor measurements as the intermediate daily values that are then arithmetically averaged over the first defined time period.
Bergenstal teaches CGM data summarization using median-based statistical representations of glucose measurements derived from repeated daily CGM recordings across a multi-day period. Specifically, Bergenstal teaches that the Ambulatory Glucose Profile (AGP) presents “smoothed curves representing the median (50th) … frequency percentiles” computed from glucose values collected across multiple days (Bergenstal, p. 206-207, Part 2), which shows that within the CGM data analysis field, the median is an established central-tendency statistic used to summarize glucose measurements derived from sensor readings. Bergenstal further teaches that “modal day displays of glucose medians, peaks, troughs, and variability were highly reflective of what the display would look like after 30 days of CGM use” (Bergenstal, p. 203-204, Proposed AGP), which shows that median-based glucose summaries derived from multi-day CGM data are recognized as stable and representative summaries of a patient’s glucose behavior across time. Bergenstal thereby establishes that the median is an art-recognized statistic for summarizing CGM sensor measurements into a representative glucose value, motivated by its robustness to transient excursions and intraday noise.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Bergenstal so that the aggregated glucose value used in Hadley’s feature set comprises an arithmetic average of a plurality of daily median recorded sensor measurements for the user over the first defined time period. One of ordinary skill in the art would have found it obvious to do so because Bergenstal establishes that the median is a preferred statistic for representing CGM-derived glucose measurements when robustness to transient excursions and intraday sensor noise is desired, while Hadley teaches forming aggregated glucose features using a daily representative glucose value arithmetically averaged across a multi-day period. One of ordinary skill in the art implementing Hadley’s per-day glucose aggregation step would have had a clear, articulable reason in view of Bergenstal to select the intraday median of each day’s recorded sensor measurements as the per-day representative value, rather than another summary statistic, because Bergenstal teaches that the median is more robust than mean-based measures for characterizing CGM glucose behavior in the presence of intraday excursions and noise. Having selected the daily median as the per-day representative value in view of Bergenstal, the subsequent step of arithmetically averaging those daily median values across the plurality of days in the first defined time period follows directly from Hadley’s existing multi-day averaging methodology. The result, an arithmetic average of daily median sensor readings, represents a straightforward and predictable combination in which the median is applied at the within-day level to suppress intraday noise and transient outliers, while the arithmetic average is applied across days to produce a stable multi-day summary of central glucose tendency. Therefore, incorporating Bergenstal’s median-based CGM summarization into the per-day aggregation step of Hadley’s multi-day averaging framework would have been a predictable use of known CGM statistical analysis techniques, motivated by improving robustness of the daily representative glucose value, to arrive at the claimed arithmetic average of daily median recorded sensor measurements. Further, Frank confirms that median glucose values and daily aggregated glucose features derived from CGM measurements were well-established inputs in glucose-prediction machine learning systems (Frank, ¶[0096]: "may also or alternately include one or more of mean glucose (e.g., over the duration of the observation period or daily), median glucose, inter quartile range of the glucose measurements 110, variance of the glucose measurements 110 ..."), showing that median glucose statistics and daily aggregated glucose measures were recognized and routinely used as input features in predictive glucose modeling.
Regarding claim 15, the modified Hadley teaches that the limited set of physiological features comprises an aggregated glucose variability value, as established in the rejection of claim 13 above.
Also regarding claim 15, the modified Hadley does not explicitly teach that the aggregated glucose variability value comprises an arithmetic average of a plurality of daily median glycemic variability measurements for the user over the first defined time period. Rather, the claim 13 art teaches aggregated glucose variability features derived from CGM data, but does not expressly disclose first computing a daily median of the glycemic variability measurements within each day and then arithmetically averaging those daily median values across the first defined time period.
Bergenstal teaches that median-based CGM summarization is a recognized technique for generating representative glucose-related statistics from repeated daily CGM data across a multi-day period. Specifically, Bergenstal teaches that the Ambulatory Glucose Profile (AGP) presents “smoothed curves representing the median (50th) … frequency percentiles” computed from glucose values collected across multiple days (Bergenstal, p. 206-207, Part 2), which shows that median-based summarization is an established statistical technique in CGM data analysis for generating representative values from repeated measurements. Bergenstal further teaches that “modal day displays of glucose medians, peaks, troughs, and variability were highly reflective of what the display would look like after 30 days of CGM use” (Bergenstal, p. 203-204, Proposed AGP), which confirms that median-based summarization is recognized as a stable and representative technique for characterizing both glucose levels and variability across a multi-day CGM period.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Dave and Bergenstal so that the aggregated glucose variability value used in Hadley’s feature set comprises an arithmetic average of a plurality of daily median glycemic variability measurements for the user over the first defined time period. One of ordinary skill in the art would have found it obvious to do so for the following reasons. First, Dave teaches generating glycemic variability measurements, specifically rolling-window standard deviations such as sd_2hr and sd_4hr, computed at regular CGM reading intervals throughout each day, which produces an intraday time-series of glycemic variability values within each day of the first defined time period. Second, one of ordinary skill in the art implementing Hadley’s aggregated glucose variability feature would have recognized that those intraday variability values are subject to the same transient excursions and within-day noise as intraday glucose readings, and would therefore have been motivated, in view of Bergenstal’s teaching that the median is a preferred robust summary statistic for CGM-derived data, to select the daily median of each day’s intraday variability measurements as the per-day representative variability value. Third, a person of ordinary skill in the art would have applied the same median-then-average methodology to the glucose variability feature as a direct and obvious parallel, because both the glucose feature and the glucose variability feature are derived from the same intraday CGM time-series, are susceptible to the same sources of intraday noise, and are aggregated across the same multi-day time period, such that using a consistent per-day summarization methodology for both features would have been an obvious design choice. The subsequent step of arithmetically averaging those daily median variability values across the plurality of days in the first defined time period follows directly from Hadley’s use of multi-day aggregated feature values. The result, an arithmetic average of daily median glycemic variability measurements, represents a predictable combination in which the median suppresses transient within-day fluctuations at the intraday level, while the arithmetic average produces a stable multi-day summary of glycemic variability. Therefore, incorporating Dave’s intraday glycemic variability measurements into Hadley’s feature set and applying Bergenstal’s median-based CGM summarization methodology to derive daily representative variability values would have been a predictable use of known CGM statistical analysis techniques, motivated by the same robustness considerations identified for claim 22 and by the obvious design symmetry of applying consistent aggregation methodology to parallel feature types. Further, Frank confirms that median glucose values and daily aggregated CGM-derived statistics, including variance of glucose measurements, were well-established inputs in glucose-prediction machine learning systems as of the filing date (Frank, ¶[0096]: “may also or alternately include one or more of mean glucose (e.g., over the duration of the observation period or daily), median glucose, inter quartile range of the glucose measurements 110, variance of the glucose measurements 110 ...”), showing that both median-based and variance-based CGM statistics were recognized and routinely used as input features in predictive glucose modeling frameworks.
Regarding claim 22, the modified Hadley teaches that the aggregated glucose value comprises glucose values aggregated from sensor measurements over a defined period of time (Hadley, ¶[0126]: “The glucose twin module 651 determines 866 an aggregate estimate of the patient's A1c for a rolling 10-day blood glucose average by concatenating the prediction output by the first model with the prediction output by the second model”, Hadley teaches an aggregated glucose value based on blood glucose information over a multi-day period; see also Hadley, ¶[0059]: “The illustrated interface displays biological data recorded by wearable devices over a period of time including signal curves of 5-day average blood glucose measurements (5DG CGM) 391, 1-day average blood glucose measurements (1DG-CGM) 392 ... Each point on the signal curve represents an average value of the signal measured on that day. For example, each point along the signal curve of 1DG-CGM 392 measurements represents a patient's 1-day average glucose for a given day”, Hadley further teaches generating daily aggregated glucose values from multiple recorded sensor measurements for a user over time).
Also regarding claim 22, the modified Hadley does not explicitly teach that the aggregated glucose value comprises an arithmetic average of a plurality of daily median recorded sensor measurements for the user over the first defined time period. Rather, Hadley teaches aggregated glucose values such as rolling blood glucose averages and 1-day average glucose values derived from multiple recorded sensor measurements, but does not expressly disclose using daily median recorded sensor measurements as the intermediate daily values that are then arithmetically averaged over the first defined time period.
Bergenstal teaches CGM data summarization using median-based statistical representations of glucose measurements derived from repeated daily CGM recordings across a multi-day period. Specifically, Bergenstal teaches that the Ambulatory Glucose Profile (AGP) presents “smoothed curves representing the median (50th) … frequency percentiles” computed from glucose values collected across multiple days (Bergenstal, p. 206-207, Part 2), which shows that within the CGM data analysis field, the median is an established central-tendency statistic used to summarize glucose measurements derived from sensor readings. Bergenstal further teaches that “modal day displays of glucose medians, peaks, troughs, and variability were highly reflective of what the display would look like after 30 days of CGM use” (Bergenstal, p. 203-204, Proposed AGP), which shows that median-based glucose summaries derived from multi-day CGM data are recognized as stable and representative summaries of a patient’s glucose behavior across time. Bergenstal thereby establishes that the median is an art-recognized statistic for summarizing CGM sensor measurements into a representative glucose value, motivated by its robustness to transient excursions and intraday noise.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Bergenstal so that the aggregated glucose value used in Hadley’s feature set comprises an arithmetic average of a plurality of daily median recorded sensor measurements for the user over the first defined time period. One of ordinary skill in the art would have found it obvious to do so because Bergenstal establishes that the median is a preferred statistic for representing CGM-derived glucose measurements when robustness to transient excursions and intraday sensor noise is desired, while Hadley teaches forming aggregated glucose features using a daily representative glucose value arithmetically averaged across a multi-day period. One of ordinary skill in the art implementing Hadley’s per-day glucose aggregation step would have had a clear, articulable reason in view of Bergenstal to select the intraday median of each day’s recorded sensor measurements as the per-day representative value, rather than another summary statistic, because Bergenstal teaches that the median is more robust than mean-based measures for characterizing CGM glucose behavior in the presence of intraday excursions and noise. Having selected the daily median as the per-day representative value in view of Bergenstal, the subsequent step of arithmetically averaging those daily median values across the plurality of days in the first defined time period follows directly from Hadley’s existing multi-day averaging methodology. The result, an arithmetic average of daily median sensor readings, represents a straightforward and predictable combination in which the median is applied at the within-day level to suppress intraday noise and transient outliers, while the arithmetic average is applied across days to produce a stable multi-day summary of central glucose tendency. Therefore, incorporating Bergenstal’s median-based CGM summarization into the per-day aggregation step of Hadley’s multi-day averaging framework would have been a predictable use of known CGM statistical analysis techniques, motivated by improving robustness of the daily representative glucose value, to arrive at the claimed arithmetic average of daily median recorded sensor measurements. Further, Frank confirms that median glucose values and daily aggregated glucose features derived from CGM measurements were well-established inputs in glucose-prediction machine learning systems (Frank, ¶[0096]: "may also or alternately include one or more of mean glucose (e.g., over the duration of the observation period or daily), median glucose, inter quartile range of the glucose measurements 110, variance of the glucose measurements 110 ..."), showing that median glucose statistics and daily aggregated glucose measures were recognized and routinely used as input features in predictive glucose modeling.
Regarding claim 23, the modified Hadley teaches that the limited set of physiological features comprises an aggregated glucose variability value, as established in the rejection of claim 21 above.
Also regarding claim 23, the modified Hadley does not explicitly teach that the aggregated glucose variability value comprises an arithmetic average of a plurality of daily median glycemic variability measurements for the user over the first defined time period. Rather, the claim 21 art teaches aggregated glucose variability features derived from CGM data, but does not expressly disclose first computing a daily median of the glycemic variability measurements within each day and then arithmetically averaging those daily median values across the first defined time period.
Bergenstal teaches that median-based CGM summarization is a recognized technique for generating representative glucose-related statistics from repeated daily CGM data across a multi-day period. Specifically, Bergenstal teaches that the Ambulatory Glucose Profile (AGP) presents “smoothed curves representing the median (50th) … frequency percentiles” computed from glucose values collected across multiple days (Bergenstal, p. 206-207, Part 2), which shows that median-based summarization is an established statistical technique in CGM data analysis for generating representative values from repeated measurements. Bergenstal further teaches that “modal day displays of glucose medians, peaks, troughs, and variability were highly reflective of what the display would look like after 30 days of CGM use” (Bergenstal, p. 203-204, Proposed AGP), which confirms that median-based summarization is recognized as a stable and representative technique for characterizing both glucose levels and variability across a multi-day CGM period.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of Dave and Bergenstal so that the aggregated glucose variability value used in Hadley’s feature set comprises an arithmetic average of a plurality of daily median glycemic variability measurements for the user over the first defined time period. One of ordinary skill in the art would have found it obvious to do so for the following reasons. First, Dave teaches generating glycemic variability measurements, specifically rolling-window standard deviations such as sd_2hr and sd_4hr, computed at regular CGM reading intervals throughout each day, which produces an intraday time-series of glycemic variability values within each day of the first defined time period. Second, one of ordinary skill in the art implementing Hadley’s aggregated glucose variability feature would have recognized that those intraday variability values are subject to the same transient excursions and within-day noise as intraday glucose readings, and would therefore have been motivated, in view of Bergenstal’s teaching that the median is a preferred robust summary statistic for CGM-derived data, to select the daily median of each day’s intraday variability measurements as the per-day representative variability value. Third, a person of ordinary skill in the art would have applied the same median-then-average methodology to the glucose variability feature as a direct and obvious parallel, because both the glucose feature and the glucose variability feature are derived from the same intraday CGM time-series, are susceptible to the same sources of intraday noise, and are aggregated across the same multi-day time period, such that using a consistent per-day summarization methodology for both features would have been an obvious design choice. The subsequent step of arithmetically averaging those daily median variability values across the plurality of days in the first defined time period follows directly from Hadley’s use of multi-day aggregated feature values. The result, an arithmetic average of daily median glycemic variability measurements, represents a predictable combination in which the median suppresses transient within-day fluctuations at the intraday level, while the arithmetic average produces a stable multi-day summary of glycemic variability. Therefore, incorporating Dave’s intraday glycemic variability measurements into Hadley’s feature set and applying Bergenstal’s median-based CGM summarization methodology to derive daily representative variability values would have been a predictable use of known CGM statistical analysis techniques, motivated by the same robustness considerations identified for claim 22 and by the obvious design symmetry of applying consistent aggregation methodology to parallel feature types. Further, Frank confirms that median glucose values and daily aggregated CGM-derived statistics, including variance of glucose measurements, were well-established inputs in glucose-prediction machine learning systems as of the filing date (Frank, ¶[0096]: “may also or alternately include one or more of mean glucose (e.g., over the duration of the observation period or daily), median glucose, inter quartile range of the glucose measurements 110, variance of the glucose measurements 110 ...”), showing that both median-based and variance-based CGM statistics were recognized and routinely used as input features in predictive glucose modeling frameworks.
Claims 8-9, 19-20, and 24-25 are rejected under 35 U.S.C. 103 as obvious over Hadley et al. (US 20220061710 A1), hereto referred as Hadley, and further in view of Jimenez et al. (US 20210216894 A1), hereto referred as Jimenez, and further in view of Apostolova et al. (US 20200381090 A1), hereto referred as Apostolova, and further in view of McMahon et al. (US 20180272064 A1), hereto referred as McMahon.
Regarding claim 8, the modified Hadley teaches that the physiological prediction for the user comprises a prediction of a physiological parameter based on recorded sensor values over a defined time period (Hadley, Abstract: “A patient health management platform implements a machine learned metabolic model to generate a prediction of a patient’s glucose level. The platform implements a short-term prediction model to generate a daily prediction of the patient’s glucose level based on nutrition data reported by the patient and sensor data and lab test data collected for the patient. The platform implements a long-term prediction model generate a prediction of the patient’s glucose level during an extended time period based on sensor data and lab test data collected for the patient”, Hadley teaches generating future glucose predictions from recorded sensor data over both short-term and extended future periods);.
Also regarding claim 8, the modified Hadley does not explicitly teach that the physiological prediction for the user comprises a plurality of predicted average sensor values for the user during a future time period corresponding to the plurality of recorded sensor values. Rather, the modified Hadley teaches generating a daily prediction of the patient’s glucose level and a longer-term prediction of glucose level during an extended time period, but does not expressly disclose that the physiological prediction comprises a plurality of predicted average sensor values across multiple future intervals within the future time period.
McMahon teaches generating a plurality of predicted average glucose values for future time intervals and presenting those predicted values as a future glucose trajectory. McMahon teaches that “the LSTM cell 1004 calculates or otherwise determines an average glucose value 1009 associated with the 11 AM interval utilizing the forecasting model for the 11 AM hourly interval” and that the “forecasted glucose value 1009 for the 11 AM interval ... [is] input to the 12 PM hourly interval LSTM cell 1006 for determining a forecasted glucose value for the 12 PM interval ... and so on” (McMahon, ¶[0128]), which shows that McMahon generates a sequence of forecasted average glucose values for successive future intervals. McMahon further teaches that the system determines “ensemble predicted glucose values within those different prediction horizons as weighted averages of the outputs of the different patient-specific glucose prediction models” (McMahon, ¶[0138]), which confirms that McMahon’s predicted values are predicted average sensor values across multiple prediction intervals, constituting a plurality of predicted average sensor values over the future time period.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of McMahon so that the physiological prediction output by Hadley comprises a plurality of predicted average sensor values for the user during a future time period corresponding to the plurality of recorded sensor values. One of ordinary skill in the art would have found it obvious to do so because Hadley already teaches machine-learning prediction of a patient’s future glucose level based on recorded sensor data over a defined time period, while McMahon teaches that such a future physiological prediction is advantageously expressed as a sequence of predicted average glucose values over successive intervals within the future period, rather than as a single undifferentiated future value, because doing so provides a multi-point future trajectory that more granularly characterizes expected future glucose behavior. One of ordinary skill in the art would have recognized that structuring Hadley’s future glucose prediction as a plurality of predicted average values over successive intervals within the future prediction period represents a predictable and straightforward implementation choice, motivated by the recognized clinical benefit of enabling the patient, provider, or system to evaluate expected future glucose levels at multiple points across the prediction horizon rather than at a single future time point, thereby improving monitoring, treatment planning, and prediction-based decision making.
Regarding claim 9, the modified Hadley teaches that the physiological prediction for the user comprises a plurality of predicted average sensor values for the user during a future time period corresponding to the plurality of recorded sensor values, as established in the rejection of claim 8.
Also regarding claim 9, the modified Hadley does not explicitly teach that the future time period comprises a seventy five day time period and the plurality of predicted average sensor values comprises a predicted daily average sensor value for one or more days of the seventy five day time period. Rather, the modified Hadley teaches a long-term prediction model that generates a prediction of the patient’s glucose level during an extended time period and teaches daily average glucose values as a unit of aggregation, while McMahon teaches expressing future physiological prediction as a sequence of predicted average glucose values across successive future intervals. However, the modified Hadley in view of McMahon does not expressly disclose selecting a seventy five day future prediction horizon or expressly disclosing that the predicted average sensor values are daily average values for one or more days within that seventy five day future period.
Hadley teaches that the platform implements a long-term prediction model to generate a prediction of the patient’s glucose level during an extended time period based on sensor data and lab test data collected for the patient (Hadley, Abstract), which shows that Hadley contemplates long-horizon future glucose prediction. Hadley further teaches that the system generates A1c-related predictions based on rolling average glucose measures, specifically that “the glucose twin module 651 determines 866 an aggregate estimate of the patient’s A1c for a rolling 10-day blood glucose average by concatenating the prediction output by the first model with the prediction output by the second model” (Hadley, ¶[0126]), which shows that Hadley’s long-term prediction framework is directed to A1c-relevant glycemic assessment based on multi-day average glucose values. Hadley further teaches generating daily average glucose values as the unit of aggregation, specifically that “each point along the signal curve of 1DG-CGM 392 measurements represents a patient’s 1-day average glucose for a given day” (Hadley, ¶[0059]), which shows that Hadley uses the daily average glucose value as an established granularity for glucose prediction output.
McMahon teaches that future glucose prediction may be expressed as a sequence of predicted average glucose values over successive future intervals within a selected prediction horizon (McMahon, ¶[0128]; ¶[0138]), as established in the rejection of claim 24, confirming that structuring a future glucose prediction as a sequence of per-interval predicted average values is a known implementation approach.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of McMahon so that the future time period for the plurality of predicted average sensor values comprises a seventy five day time period, with the plurality of predicted average sensor values including a predicted daily average sensor value for one or more days within that seventy five day period. One of ordinary skill in the art would have found it obvious to do so for the following reasons. First, Hadley already teaches long-term glucose prediction directed toward A1c-relevant glycemic assessment, using multi-day rolling average glucose values and aggregate A1c estimation as the clinical objective of its long-term prediction model. One of ordinary skill in the art would have recognized that the seventy five day future prediction horizon is directly and specifically motivated by the clinical significance of A1c as a measure of longer-term glycemic control. Because Hadley’s long-term prediction framework is already directed toward A1c-relevant glucose prediction, one of ordinary skill in the art implementing that framework would have been motivated to select a 75-day future prediction horizon as a clinically determined and technically appropriate window that aligns the long-term glucose prediction output with the A1c-relevant glucose contribution period. Second, once the 75-day horizon is selected, the use of daily average glucose values as the per-interval prediction unit follows directly from Hadley’s own teaching that the 1-day average glucose value is the standard unit of daily glucose aggregation (Hadley, ¶[0059]), combined with McMahon’s teaching that future glucose prediction may be expressed as a sequence of per-interval predicted average values. Applying Hadley’s daily average granularity to McMahon’s multi-interval prediction sequence structure over the selected 75-day horizon produces the claimed combination in a straightforward and predictable manner. Third, the claim requires only that the plurality of predicted average sensor values comprises a predicted daily average for one or more days of the seventy five day period, a requirement that is satisfied even by a single daily predicted value within the future horizon. Hadley already teaches generating a daily prediction of the patient’s glucose level (Hadley, Abstract), such that the primary non-disclosed element is the 75-day horizon itself, while the daily granularity prediction within that horizon is already taught or at least obvious from Hadley and McMahon for the reasons stated above. Therefore, selecting a seventy five day prediction horizon for Hadley’s long-term daily average glucose prediction, and expressing the resulting prediction as a predicted daily average glucose value for one or more days within that period, would have been a predictable and clinically motivated implementation of Hadley’s A1c-directed long-term prediction framework using McMahon’s multi-interval average-value output structure.
Regarding claim 19, the modified Hadley teaches that the physiological prediction for the user comprises a prediction of a physiological parameter based on recorded sensor values over a defined time period (Hadley, Abstract: “A patient health management platform implements a machine learned metabolic model to generate a prediction of a patient’s glucose level. The platform implements a short-term prediction model to generate a daily prediction of the patient’s glucose level based on nutrition data reported by the patient and sensor data and lab test data collected for the patient. The platform implements a long-term prediction model generate a prediction of the patient’s glucose level during an extended time period based on sensor data and lab test data collected for the patient”, Hadley teaches generating future glucose predictions from recorded sensor data over both short-term and extended future periods);.
Also regarding claim 19, the modified Hadley does not explicitly teach that the physiological prediction for the user comprises a plurality of predicted average sensor values for the user during a future time period corresponding to the plurality of recorded sensor values. Rather, the modified Hadley teaches generating a daily prediction of the patient’s glucose level and a longer-term prediction of glucose level during an extended time period, but does not expressly disclose that the physiological prediction comprises a plurality of predicted average sensor values across multiple future intervals within the future time period.
McMahon teaches generating a plurality of predicted average glucose values for future time intervals and presenting those predicted values as a future glucose trajectory. McMahon teaches that “the LSTM cell 1004 calculates or otherwise determines an average glucose value 1009 associated with the 11 AM interval utilizing the forecasting model for the 11 AM hourly interval” and that the “forecasted glucose value 1009 for the 11 AM interval ... [is] input to the 12 PM hourly interval LSTM cell 1006 for determining a forecasted glucose value for the 12 PM interval ... and so on” (McMahon, ¶[0128]), which shows that McMahon generates a sequence of forecasted average glucose values for successive future intervals. McMahon further teaches that the system determines “ensemble predicted glucose values within those different prediction horizons as weighted averages of the outputs of the different patient-specific glucose prediction models” (McMahon, ¶[0138]), which confirms that McMahon’s predicted values are predicted average sensor values across multiple prediction intervals, constituting a plurality of predicted average sensor values over the future time period.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of McMahon so that the physiological prediction output by Hadley comprises a plurality of predicted average sensor values for the user during a future time period corresponding to the plurality of recorded sensor values. One of ordinary skill in the art would have found it obvious to do so because Hadley already teaches machine-learning prediction of a patient’s future glucose level based on recorded sensor data over a defined time period, while McMahon teaches that such a future physiological prediction is advantageously expressed as a sequence of predicted average glucose values over successive intervals within the future period, rather than as a single undifferentiated future value, because doing so provides a multi-point future trajectory that more granularly characterizes expected future glucose behavior. One of ordinary skill in the art would have recognized that structuring Hadley’s future glucose prediction as a plurality of predicted average values over successive intervals within the future prediction period represents a predictable and straightforward implementation choice, motivated by the recognized clinical benefit of enabling the patient, provider, or system to evaluate expected future glucose levels at multiple points across the prediction horizon rather than at a single future time point, thereby improving monitoring, treatment planning, and prediction-based decision making.
Regarding claim 20, the modified Hadley teaches that the physiological prediction for the user comprises a plurality of predicted average sensor values for the user during a future time period corresponding to the plurality of recorded sensor values, as established in the rejection of claim 19.
Also regarding claim 20, the modified Hadley does not explicitly teach that the future time period comprises a seventy five day time period and the plurality of predicted average sensor values comprises a predicted daily average sensor value for one or more days of the seventy five day time period. Rather, the modified Hadley teaches a long-term prediction model that generates a prediction of the patient’s glucose level during an extended time period and teaches daily average glucose values as a unit of aggregation, while McMahon teaches expressing future physiological prediction as a sequence of predicted average glucose values across successive future intervals. However, the modified Hadley in view of McMahon does not expressly disclose selecting a seventy five day future prediction horizon or expressly disclosing that the predicted average sensor values are daily average values for one or more days within that seventy five day future period.
Hadley teaches that the platform implements a long-term prediction model to generate a prediction of the patient’s glucose level during an extended time period based on sensor data and lab test data collected for the patient (Hadley, Abstract), which shows that Hadley contemplates long-horizon future glucose prediction. Hadley further teaches that the system generates A1c-related predictions based on rolling average glucose measures, specifically that “the glucose twin module 651 determines 866 an aggregate estimate of the patient’s A1c for a rolling 10-day blood glucose average by concatenating the prediction output by the first model with the prediction output by the second model” (Hadley, ¶[0126]), which shows that Hadley’s long-term prediction framework is directed to A1c-relevant glycemic assessment based on multi-day average glucose values. Hadley further teaches generating daily average glucose values as the unit of aggregation, specifically that “each point along the signal curve of 1DG-CGM 392 measurements represents a patient’s 1-day average glucose for a given day” (Hadley, ¶[0059]), which shows that Hadley uses the daily average glucose value as an established granularity for glucose prediction output.
McMahon teaches that future glucose prediction may be expressed as a sequence of predicted average glucose values over successive future intervals within a selected prediction horizon (McMahon, ¶[0128]; ¶[0138]), as established in the rejection of claim 24, confirming that structuring a future glucose prediction as a sequence of per-interval predicted average values is a known implementation approach.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of McMahon so that the future time period for the plurality of predicted average sensor values comprises a seventy five day time period, with the plurality of predicted average sensor values including a predicted daily average sensor value for one or more days within that seventy five day period. One of ordinary skill in the art would have found it obvious to do so for the following reasons. First, Hadley already teaches long-term glucose prediction directed toward A1c-relevant glycemic assessment, using multi-day rolling average glucose values and aggregate A1c estimation as the clinical objective of its long-term prediction model. One of ordinary skill in the art would have recognized that the seventy five day future prediction horizon is directly and specifically motivated by the clinical significance of A1c as a measure of longer-term glycemic control. Because Hadley’s long-term prediction framework is already directed toward A1c-relevant glucose prediction, one of ordinary skill in the art implementing that framework would have been motivated to select a 75-day future prediction horizon as a clinically determined and technically appropriate window that aligns the long-term glucose prediction output with the A1c-relevant glucose contribution period. Second, once the 75-day horizon is selected, the use of daily average glucose values as the per-interval prediction unit follows directly from Hadley’s own teaching that the 1-day average glucose value is the standard unit of daily glucose aggregation (Hadley, ¶[0059]), combined with McMahon’s teaching that future glucose prediction may be expressed as a sequence of per-interval predicted average values. Applying Hadley’s daily average granularity to McMahon’s multi-interval prediction sequence structure over the selected 75-day horizon produces the claimed combination in a straightforward and predictable manner. Third, the claim requires only that the plurality of predicted average sensor values comprises a predicted daily average for one or more days of the seventy five day period, a requirement that is satisfied even by a single daily predicted value within the future horizon. Hadley already teaches generating a daily prediction of the patient’s glucose level (Hadley, Abstract), such that the primary non-disclosed element is the 75-day horizon itself, while the daily granularity prediction within that horizon is already taught or at least obvious from Hadley and McMahon for the reasons stated above. Therefore, selecting a seventy five day prediction horizon for Hadley’s long-term daily average glucose prediction, and expressing the resulting prediction as a predicted daily average glucose value for one or more days within that period, would have been a predictable and clinically motivated implementation of Hadley’s A1c-directed long-term prediction framework using McMahon’s multi-interval average-value output structure.
Regarding claim 24, the modified Hadley teaches that the physiological prediction for the user comprises a prediction of a physiological parameter based on recorded sensor values over a defined time period (Hadley, Abstract: “A patient health management platform implements a machine learned metabolic model to generate a prediction of a patient’s glucose level. The platform implements a short-term prediction model to generate a daily prediction of the patient’s glucose level based on nutrition data reported by the patient and sensor data and lab test data collected for the patient. The platform implements a long-term prediction model generate a prediction of the patient’s glucose level during an extended time period based on sensor data and lab test data collected for the patient”, Hadley teaches generating future glucose predictions from recorded sensor data over both short-term and extended future periods);.
Also regarding claim 24, the modified Hadley does not explicitly teach that the physiological prediction for the user comprises a plurality of predicted average sensor values for the user during a future time period corresponding to the plurality of recorded sensor values. Rather, the modified Hadley teaches generating a daily prediction of the patient’s glucose level and a longer-term prediction of glucose level during an extended time period, but does not expressly disclose that the physiological prediction comprises a plurality of predicted average sensor values across multiple future intervals within the future time period.
McMahon teaches generating a plurality of predicted average glucose values for future time intervals and presenting those predicted values as a future glucose trajectory. McMahon teaches that “the LSTM cell 1004 calculates or otherwise determines an average glucose value 1009 associated with the 11 AM interval utilizing the forecasting model for the 11 AM hourly interval” and that the “forecasted glucose value 1009 for the 11 AM interval ... [is] input to the 12 PM hourly interval LSTM cell 1006 for determining a forecasted glucose value for the 12 PM interval ... and so on” (McMahon, ¶[0128]), which shows that McMahon generates a sequence of forecasted average glucose values for successive future intervals. McMahon further teaches that the system determines “ensemble predicted glucose values within those different prediction horizons as weighted averages of the outputs of the different patient-specific glucose prediction models” (McMahon, ¶[0138]), which confirms that McMahon’s predicted values are predicted average sensor values across multiple prediction intervals, constituting a plurality of predicted average sensor values over the future time period.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of McMahon so that the physiological prediction output by Hadley comprises a plurality of predicted average sensor values for the user during a future time period corresponding to the plurality of recorded sensor values. One of ordinary skill in the art would have found it obvious to do so because Hadley already teaches machine-learning prediction of a patient’s future glucose level based on recorded sensor data over a defined time period, while McMahon teaches that such a future physiological prediction is advantageously expressed as a sequence of predicted average glucose values over successive intervals within the future period, rather than as a single undifferentiated future value, because doing so provides a multi-point future trajectory that more granularly characterizes expected future glucose behavior. One of ordinary skill in the art would have recognized that structuring Hadley’s future glucose prediction as a plurality of predicted average values over successive intervals within the future prediction period represents a predictable and straightforward implementation choice, motivated by the recognized clinical benefit of enabling the patient, provider, or system to evaluate expected future glucose levels at multiple points across the prediction horizon rather than at a single future time point, thereby improving monitoring, treatment planning, and prediction-based decision making.
Regarding claim 25, the modified Hadley teaches that the physiological prediction for the user comprises a plurality of predicted average sensor values for the user during a future time period corresponding to the plurality of recorded sensor values, as established in the rejection of claim 24.
Also regarding claim 25, the modified Hadley does not explicitly teach that the future time period comprises a seventy five day time period and the plurality of predicted average sensor values comprises a predicted daily average sensor value for one or more days of the seventy five day time period. Rather, the modified Hadley teaches a long-term prediction model that generates a prediction of the patient’s glucose level during an extended time period and teaches daily average glucose values as a unit of aggregation, while McMahon teaches expressing future physiological prediction as a sequence of predicted average glucose values across successive future intervals. However, the modified Hadley in view of McMahon does not expressly disclose selecting a seventy five day future prediction horizon or expressly disclosing that the predicted average sensor values are daily average values for one or more days within that seventy five day future period.
Hadley teaches that the platform implements a long-term prediction model to generate a prediction of the patient’s glucose level during an extended time period based on sensor data and lab test data collected for the patient (Hadley, Abstract), which shows that Hadley contemplates long-horizon future glucose prediction. Hadley further teaches that the system generates A1c-related predictions based on rolling average glucose measures, specifically that “the glucose twin module 651 determines 866 an aggregate estimate of the patient’s A1c for a rolling 10-day blood glucose average by concatenating the prediction output by the first model with the prediction output by the second model” (Hadley, ¶[0126]), which shows that Hadley’s long-term prediction framework is directed to A1c-relevant glycemic assessment based on multi-day average glucose values. Hadley further teaches generating daily average glucose values as the unit of aggregation, specifically that “each point along the signal curve of 1DG-CGM 392 measurements represents a patient’s 1-day average glucose for a given day” (Hadley, ¶[0059]), which shows that Hadley uses the daily average glucose value as an established granularity for glucose prediction output.
McMahon teaches that future glucose prediction may be expressed as a sequence of predicted average glucose values over successive future intervals within a selected prediction horizon (McMahon, ¶[0128]; ¶[0138]), as established in the rejection of claim 24, confirming that structuring a future glucose prediction as a sequence of per-interval predicted average values is a known implementation approach.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Hadley in view of McMahon so that the future time period for the plurality of predicted average sensor values comprises a seventy five day time period, with the plurality of predicted average sensor values including a predicted daily average sensor value for one or more days within that seventy five day period. One of ordinary skill in the art would have found it obvious to do so for the following reasons. First, Hadley already teaches long-term glucose prediction directed toward A1c-relevant glycemic assessment, using multi-day rolling average glucose values and aggregate A1c estimation as the clinical objective of its long-term prediction model. One of ordinary skill in the art would have recognized that the seventy five day future prediction horizon is directly and specifically motivated by the clinical significance of A1c as a measure of longer-term glycemic control. Because Hadley’s long-term prediction framework is already directed toward A1c-relevant glucose prediction, one of ordinary skill in the art implementing that framework would have been motivated to select a 75-day future prediction horizon as a clinically determined and technically appropriate window that aligns the long-term glucose prediction output with the A1c-relevant glucose contribution period. Second, once the 75-day horizon is selected, the use of daily average glucose values as the per-interval prediction unit follows directly from Hadley’s own teaching that the 1-day average glucose value is the standard unit of daily glucose aggregation (Hadley, ¶[0059]), combined with McMahon’s teaching that future glucose prediction may be expressed as a sequence of per-interval predicted average values. Applying Hadley’s daily average granularity to McMahon’s multi-interval prediction sequence structure over the selected 75-day horizon produces the claimed combination in a straightforward and predictable manner. Third, the claim requires only that the plurality of predicted average sensor values comprises a predicted daily average for one or more days of the seventy five day period, a requirement that is satisfied even by a single daily predicted value within the future horizon. Hadley already teaches generating a daily prediction of the patient’s glucose level (Hadley, Abstract), such that the primary non-disclosed element is the 75-day horizon itself, while the daily granularity prediction within that horizon is already taught or at least obvious from Hadley and McMahon for the reasons stated above. Therefore, selecting a seventy five day prediction horizon for Hadley’s long-term daily average glucose prediction, and expressing the resulting prediction as a predicted daily average glucose value for one or more days within that period, would have been a predictable and clinically motivated implementation of Hadley’s A1c-directed long-term prediction framework using McMahon’s multi-interval average-value output structure.
Response to Arguments
Objections
Applicant's arguments filed 3/10/2026, page 9, regarding the previous Objections of claim 12 have been fully considered and are persuasive. The previous Objections have been withdrawn.
35 U.S.C. §112(b)
Applicant's arguments filed 3/10/2026, pages 9-10, regarding the previous 112(b) Rejections of claims 1-20 have been fully considered and are persuasive. The previous 112(b) rejections have been withdrawn.
35 U.S.C. §101
Applicant's arguments filed 3/10/2026, pages 10-18, regarding the previous 101 Rejections of claims 1-20 have been fully considered and are persuasive in part for the reasons set forth below. Accordingly, the rejection of claims 1-20 under 35 U.S.C. § 101 is withdrawn.
Step 2A, Prong One
Upon reconsideration, the Examiner agrees that amended claim 1 does not recite a mental process. The claimed steps, including generating activity encodings from an interaction data object, augmenting physiological features with those encodings to construct a combined input feature vector, and inputting that vector to a trained machine learning model, require the use of a specific computing system and cannot practically be performed in the human mind. The Examiner further agrees that claim 1 is not directed to fundamental economic practices, commercial or legal interactions, or managing personal behavior or relationships between people.
However, the Examiner maintains that amended claim 1 recites a judicial exception under Prong One, specifically a mathematical concept. The core computational operations of the claim, including one-hot encoding of historical medication usage, feature embedding of activity codes, construction of a feature vector by augmentation, and machine learning inference, are mathematical in character as defined under MPEP § 2106.04(a)(2)(I).
Step 2A, Prong Two
The Examiner finds Applicant's Prong Two arguments persuasive. The Examiner has reconsidered the claim in light of Applicant's amendments and remarks, including the arguments presented under Ex Parte Desjardines (Decision on Request for Rehearing, September 26, 2025) and the December 2025 memorandum addressing Desjardins.
Applicant's reliance on example (xiv) is noted but is not the most precise fit. Example (xiv) addresses adjustments to parameters of a machine learning model. Amended claim 1 does not recite adjusting or updating model parameters as an active step. Instead, the machine learning model is constrained by the "wherein" clause to one that has been trained on a labeled dataset comprising a historical combined input feature vector and a ground truth physiological feature. This clause is treated as a meaningful structural limitation on the model.
The Examiner nevertheless agrees that the claim integrates the recited mathematical concept into a practical application. Amended claim 1 recites a specific technique in which physiological features derived from recorded sensor values during a first defined time period are augmented with activity encodings generated from an interaction data object corresponding to a second defined time period that precedes the first defined time period. This results in a combined input feature vector that is provided to a trained machine learning model.
This two-window temporal structure addresses a technical problem described in the specification, namely improving physiological prediction performance when only a limited initial time window of physiological sensor data is available. The specification explains that combining a limited physiological dataset with encoded historical context improves predictive accuracy and range. See Specification ¶¶[0076]-[0081].
The "wherein" clause further ties the claimed inference operation to a model that is trained using the same combined feature vector structure, thereby linking the feature construction technique to the configuration of the model itself. The claim therefore reflects a specific machine learning implementation in which the structure of the input data and the training of the model are aligned.
In view of these limitations, the claim does more than apply mathematical operations in a general manner. The claim recites a particular technical implementation that uses a defined temporal relationship and specific encoding techniques to improve the performance of a physiological prediction system. The Examiner determines that the recited judicial exception is integrated into a practical application.
Step 2B
Because the claim is determined to be patent eligible under Step 2A, Prong Two, it is not necessary to reach Step 2B.
Conclusion
Accordingly, the rejection of claims 1-20 under 35 U.S.C. § 101 is withdrawn.
The Examiner notes for the record that this determination is based on the following claim features: (1) the temporal relationship in which activity encodings from a preceding time period are combined with physiological features from a later time period; (2) the specific encoding techniques recited, including one-hot encoding and feature embedding; and (3) the limitation that the machine learning model is trained on a dataset comprising combined input feature vectors of the claimed structure. This determination should not be interpreted as a general finding that all claims involving machine learning are patent eligible.
35 U.S.C. §102/103
Applicant's arguments filed 3/10/2026, pages 18-20, regarding the previous 102/103 Rejections of claims 1-20 have been fully considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument. That is, there are new grounds of rejection. Additionally, the arguments are not persuasive as detailed below.
Applicant’s Argument: Applicant argues that Hadley fails to teach or suggest the amended limitations of claim 1, including (i) receiving a limited set of physiological features recorded during a first defined time period, (ii) generating activity encodings based on an interaction data object corresponding to a second defined time period preceding the first, including a one-hot encoding of historical medication usage, and (iii) using such information within the claimed machine-learning framework.
Examiner’s Response: Applicant’s arguments have been fully considered but are not persuasive.
With respect to receiving physiological features based on recorded sensor values during a first defined time period, Hadley teaches receiving biological data including sensor data comprising biosignals recorded by one or more sensors worn or implemented by a patient and retrieving biological data recorded within a time period to determine the patient’s actual metabolic state (Hadley, ¶[0046]; ¶[0056]). Hadley further teaches analyzing patient data recorded over a given time period to generate a prediction of the patient’s metabolic state (Hadley, ¶[0051]).
With respect to the “limited set” recitation, the claim does not define a specific quantitative threshold for ‘limited’ on the number of physiological features. To the extent Hadley processes sensor values recorded within a defined time period, the resulting collection constitutes a bounded set of features received by the system. More importantly, it would have been prima facie obvious before the effective filing date of the claimed invention to apply Hadley’s predictive modeling framework to situations in which only a limited amount of physiological data is available. Such scenarios routinely arise in clinical monitoring contexts, including newly enrolled patients, gaps in sensor wear, or early-stage monitoring where only a short observation window exists. One of ordinary skill in the art would have been motivated to configure Hadley’s system to operate on whatever subset of physiological features is available within the defined collection period in order to enable prediction under these common conditions.
With respect to the recitation of the interaction data object corresponding to a second defined time period and the one-hot encoding of historical medication usage, Applicant’s argument is persuasive only as to Hadley alone, but is not persuasive as to the rejection over Hadley in view of Jimenez and further in view of Apostolova. Hadley teaches that patient data, including medication data, is stored as part of an ongoing timeline of entries for a current time period, including medications taken by the patient at recorded times over that time period (Hadley, ¶[0048]). Jimenez teaches that medication-use covariates are taken from a defined look-back period preceding an index date and that prescriptions and other markers present in that prior period are included in predictive modeling, with each marker associated with a vector indicating whether it is true or false (Jimenez, ¶[0025]-¶[0038]; ¶[0044]; ¶[0047]). Jimenez therefore teaches both the use of medication-related information from a defined preceding period and its representation in binary vector form, which corresponds to the recited one-hot encoding of historical medication usage. Apostolova further teaches that historical coded medical-context data, including medication information, may be represented in vector form and used together with current physiological or structured clinical data for prediction (Apostolova, ¶[0019]; ¶[0021]). The combined teachings render at least one of the recited alternatives obvious. Because the claim recites “one or more of” the listed alternatives, establishing obviousness as to one alternative is sufficient. Applicant’s arguments with respect to this limitation are therefore not persuasive.
Applicant’s Argument: Applicant argues that Hadley fails to teach the recited machine-learning implementation, including inputting a combined input feature vector, augmenting physiological features with activity encodings, training on a labeled dataset, and initiating a prediction-based action comprising outputting the physiological prediction.
Examiner’s Response: Applicant’s arguments have been fully considered but are not persuasive.
With respect to inputting a combined input feature vector to produce a physiological prediction, Hadley teaches that inputs measured by wearable sensors and lab tests or recorded manually by a patient may be encoded into a vector representation that a machine learned model is configured to receive (Hadley, ¶[0082]). Hadley further teaches generating a representation of a patient’s metabolic state by inputting wearable sensor data, lab test data, and recorded patient data into the model (Hadley, ¶[0089]; Abstract). Hadley therefore teaches inputting a feature vector comprising physiological and patient-recorded data into a machine learned model to generate a physiological prediction.
With respect to augmenting physiological features with activity encodings and training on a labeled dataset, Hadley does not expressly disclose forming the combined input feature vector through augmentation with encoded historical medication information from a preceding time period, nor the specific labeled training dataset structure. However, Jimenez teaches that historical medication markers from a defined preceding period are predictive covariates represented in vector form (Jimenez, ¶[0044]; ¶[0047]). Apostolova teaches combining historical vectorized medical context with current physiological data as inputs to machine learning prediction models (Apostolova, ¶[0019]; ¶[0021]). It would have been prima facie obvious before the effective filing date of the claimed invention to modify Hadley in view of Jimenez and Apostolova so that the physiological feature set is augmented with encoded historical medication information to form a combined input feature vector. It would further have been obvious to train the model on labeled historical instances of such combined input feature vectors paired with corresponding physiological outcomes, because a supervised machine learning model is trained using examples that share the same feature structure as those used during prediction. One of ordinary skill in the art would have understood this as standard practice for implementing such a predictive system.
With respect to initiating a prediction-based action comprising outputting the physiological prediction, Hadley teaches determining a predicted metabolic state and generating patient-specific recommendations, communications, and notifications based on that prediction (Hadley, ¶[0061]-¶[0063]). Hadley further teaches devices configured to present medically relevant data and dashboards (Hadley, ¶[0022]-¶[0024]). Although Hadley does not expressly state that the physiological prediction itself is output to the computing interface, it would have been prima facie obvious to modify Hadley to output the physiological prediction using the same interface infrastructure, since the prediction is already generated and related information is already presented. Doing so would improve transparency and provide more complete information for monitoring and decision making.
Accordingly, Applicant’s arguments have been fully considered but are not persuasive. Therefore, the rejection of claim 1 under 35 U.S.C. § 103 over Hadley in view of Jimenez and further in view of Apostolova is maintained.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the
examiner should be directed to AARON MERRIAM whose telephone number is (703) 756-
5938. The examiner can normally be reached M-F 8:00 am - 5:00 pm.
Examiner interviews are available via telephone, in-person, and video conferencing
using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is
encouraged to use the USPTO Automated Interview Request (AIR) at
http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s
supervisor, Jason Sims can be reached on (571)272-4867. The fax phone number for the
organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained
from Patent Center. Unpublished application information in Patent Center is available to
registered users. To file and manage patent submissions in Patent Center, visit:
https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for
more information about Patent Center and https://www.uspto.gov/patents/docx for
information about filing in DOCX format. For additional questions, contact the Electronic
Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO
Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/AARON MERRIAM/Examiner, Art Unit 3791
/MATTHEW KREMER/Primary Examiner, Art Unit 3791
Prosecution Insights