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 .
Status of claims
This non-final office action on merits is in response to the Patent Application filed on 01/31/2025. Claims 1-20 are pending and considered below.
Information Disclosure Statement
The information disclosure statement (IDS) filed on 09/29/2025 has been acknowledged. The submission is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Objections
Claim 19 is objected to because of the following informalities: “obtaining a plurality of ECG data windows” should read “obtaining a plurality of electrocardiogram (ECG) data windows”. Appropriate correction is required.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1
Under step 1, the analysis is based on MPEP 2106.03, and claims 1-16 are drawn to a method, and claims 17-18 are drawn to a system, and claims 19-20 are drawn to one or more non-transitory computer-readable storage media. Thus, each claim, on its face, is directed to one of the statutory categories (i.e., useful process, machine, manufacture, or composition of matter) of 35 U.S.C. §101.
Step 2A Prong One
Claim 1 recites the limitations of determining a collective cardiovascular score based on the first cardiovascular score and the second cardiovascular score; identifying, based on the one or more demographics, an operating point threshold from a plurality of operating point thresholds; comparing the collective cardiovascular score to the operating point threshold; determining an indication of a cardiovascular condition based on the comparison. These limitations, as drafted, are processes that, under their broadest reasonable interpretations, cover performance of the limitations in the mind or by using a pen and paper. Even when considering the “via the first trained AI model” and “via the second trained AI model” language, the claim encompasses a user evaluating multiple cardiovascular scores, selecting an appropriate threshold based on demographic information, comparing the collective score to the selected threshold, and determining whether the comparison indicates a cardiovascular condition in their mind or by using a pen and paper. The nominal recitation of via the first trained AI model does not take the claim limitations out of the mental processes grouping. Thus, the claim recites a mental process which is an abstract idea.
Independent claims 17 and 19 recite identical or nearly identical steps with respect to claim 1 (and therefore also recite limitations that fall within this subject matter grouping of abstract ideas), and these claims are therefore determined to recite an abstract idea under the same analysis.
Under Step 2A Prong Two
The claimed limitations, as per method claim 1, includes:
receiving electrocardiogram (ECG) data of a subject, wherein the ECG data includes a time series of voltage measurements;
segmenting the received ECG data into a plurality of ECG data windows;
obtaining one or more demographics of the subject;
providing the plurality of ECG data windows and the one or more demographics as input to:
a first trained artificial intelligence (AI) model having a first framework; and
a second trained AI model having a second framework different from the first framework;
receiving, via the first trained AI model, a first cardiovascular score;
receiving, via the second trained AI model, a second cardiovascular score;
determining a collective cardiovascular score based on the first cardiovascular score and the second cardiovascular score;
identifying, based on the one or more demographics, an operating point threshold from a plurality of operating point thresholds;
comparing the collective cardiovascular score to the operating point threshold;
determining an indication of a cardiovascular condition based on the comparison; and
providing a notification based on the indication of the cardiovascular condition.
Examiner Note: underlined elements indicate additional elements of the claimed invention identified as performing the steps of the claimed invention.
The judicial exception expressed in claim 1 is not integrated into a practical application. The claim as a whole merely describes how to generally “apply” the concept of evaluating cardiovascular information to determine an indication of a cardiovascular condition in a computer environment. The claimed computer components (i.e., a first trained artificial intelligence (AI) model having a first framework; a second trained AI model having a second framework different from the first framework; via the first trained AI model; and via the second trained AI model) are recited at a high level of generality and are merely invoked as tools to perform an existing process of evaluating cardiovascular scores, selecting an operating point threshold based on demographic information, comparing the collective cardiovascular score to the selected threshold, and determining an indication of a cardiovascular condition. Simply implementing the abstract idea on a generic computer is not a practical application of the abstract idea. Accordingly, alone and in combination, these additional elements do not integrate the abstract idea into a practical application.
The judicial exception expressed in claim 1 is not integrated into a practical application. The claim recites the additional elements of receiving electrocardiogram (ECG) data of a subject, wherein the ECG data includes a time series of voltage measurements; segmenting the received ECG data into a plurality of ECG data windows; obtaining one or more demographics of the subject; providing the plurality of ECG data windows and the one or more demographics as input to; receiving a first cardiovascular score; receiving, a second cardiovascular score; and providing a notification based on the indication of the cardiovascular condition. These limitations are recited at a high level of generality (i.e., as a general means of collecting, preparing, and presenting information for use in the recited analysis), and amounts to merely data gathering, data preprocessing, and insignificant application, which are forms of insignificant extra-solution activities. Accordingly, even in combination, these additional elements do not integrate the abstract idea into a practical application. The claim is directed to an abstract idea.
Therefore, under step 2A, the claims are directed to the abstract idea, and require further analysis under Step 2B.
Under step 2B
Claim 1 does not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed with respect to Step 2A, the claim as a whole merely describes how to generally “apply” the concept of evaluating cardiovascular information to determine an indication of a cardiovascular condition in a computer environment. Thus, even when viewed as a whole, nothing in the claim adds significantly more (i.e., an inventive concept) to the abstract idea.
For claim 1, under step 2B, the additional elements of receiving electrocardiogram (ECG) data of a subject, wherein the ECG data includes a time series of voltage measurements; segmenting the received ECG data into a plurality of ECG data windows; obtaining one or more demographics of the subject; providing the plurality of ECG data windows and the one or more demographics as input to; receiving a first cardiovascular score; receiving, a second cardiovascular score; and providing a notification based on the indication of the cardiovascular condition have been evaluated. As noted in Electric Power Group, LLC v. Alstom S.A., 830 F.3d 1350, 1354, 119 USPQ2d 1739, 1742 (Fed. Cir. 2016), merely collecting information for analysis without a technological improvement does not add significantly more to an abstract idea. The use of the method is no more than collecting, preprocessing, and providing information before and after the abstract analysis, and therefore does not integrate the abstract idea into a practical application. Additionally, as noted in In re Brown, 645 Fed. App'x 1014, 1016-1017 (Fed. Cir. 2016), merely providing notification based on the determination indication of the cardiovascular condition represents an insignificant application of the underlying mental process, as the notification merely communicates the results of the abstract analysis to a user and does not impose any meaningful limitation or add any technological improvement. Therefore, the claim does not recite an inventive concept and is not patent eligible.
Claims 3, 5-7, 9-11 recite no further additional elements, and only further narrow the abstract idea. The previously identified additional elements, individually and as a combination, do not integrate the narrowed abstract idea into a practical application for reasons similar to those explained above, and do not amount to significantly more than the narrowed abstract idea for reasons similar to those explained above.
Claims 2, 4, 8, 12-16, 18, and 20 recite the additional element of using the first trained AI model (claims 2 and 14-15) segmenting the received ECG data (claim 4), providing the ECG data as input to a third trained AI model (claim 8), using the third trained AI model (claim 8), using the first trained AI model and the second trained AI model (claim 12), providing the one or more demographics as input to the first trained AI model includes (claim 13), and providing the notification recommending that the subject be scheduled for a diagnostic test to confirm the indication of the cardiovascular condition (claim 16), the one or more processors (claim 18), provide the subject demographics as input to the plurality of trained AI models (claim 18), segmenting the ECG data (claim 20). However, these additional elements amount to implementing an abstract idea on a generic computing device or data gathering, data preprocessing, and insignificant application (i.e., insignificant extra-solution activities)). As such, these additional elements, when considered individually or in combination with the previously identified additional elements, do not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea.
Thus, as the dependent claims remain directed to a judicial exception, and as the additional elements of the claims do not amount to significantly more, the dependent claims are not patent eligible.
Therefore, the claims here fail to contain any additional element(s) or combination of additional elements that can be considered as significantly more and the claims are rejected under 35 U.S.C. 101 for lacking eligible subject matter.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1-20 are rejected under 35 U.S.C. 103 as being unpatentable over Zimmerman et al. (U.S. Patent Publication 2022/0378379 A1), referred to hereinafter as Zimmerman, in view of Khosousi et al. (U.S. Patent Publication 2023/0289595 A1), referred to hereinafter as Khosousi.
Regarding claim 1, Zimmerman teaches a method, comprising: receiving electrocardiogram (ECG) data of a subject, wherein the ECG data includes a time series of voltage measurements (Zimmerman [0021] "In the method, the plurality of leads may include a lead I, a lead V2, a lead V4, a lead V3, a lead V6, a lead II, a lead VI, and a lead V5. The electrocardiogram data may include first voltage data associated with the lead I and a first portion of the time interval, second voltage data associated with the lead V2 and a second portion of the time interval, third voltage data associated with the lead V4 and a third portion of the time interval, fourth voltage data associated with the lead V3 and the second portion of the time interval, fifth voltage data associated with the lead V6 and the third portion of the time interval, sixth voltage data associated with the lead II and the first portion of the time interval, seventh voltage data associated with the lead II and the second portion of the time interval, eighth voltage data associated with the lead II and the third portion of the time interval, ninth voltage data associated with the lead VI and the first portion of the time interval, tenth voltage data associated with the lead VI and the second portion of the time interval, eleventh voltage data associated with the lead VI and the third portion of the time interval, twelfth voltage data associated with the lead V5 and the first portion of the time interval, thirteenth voltage data associated with the lead V5 and the second portion of the time interval, and fourteenth voltage data associated with the lead V5 and the third portion of the time interval.”);
segmenting the received ECG data into a plurality of ECG data windows (Zimmerman [0297] “In one example, ECG data may be segmented into multiple portions. A model may be trained using all portions of the segmented ECG or with a varying number of portions. For example, the first or last portion of the ECG may be removed to avoid artifacts that may be present.”);
obtaining one or more demographics of the subject (Zimmerman [0036] “In one aspect, a method includes: receiving electrocardiogram data associated with a patient and an electrocardiogram configuration including a plurality of leads and a time interval, the electrocardiogram data comprising, for each lead included in the plurality of leads, voltage data associated with at least a portion of the time interval; receiving an age value associated with the patient; receiving a sex value associated with the patient; providing the age value, the sex value, and at least a portion of the electrocardiogram data to a trained model, the trained model being trained to generate a risk score based on input electrocardiogram data associated with the electrocardiogram configuration and supplementary information associated with the patient; receiving a risk score indicative of a likelihood the patient will suffer from aortic stenosis within a predetermined period of time from when the electrocardiogram data was generated; and outputting the risk score to at least one of a memory or a display for viewing by a medical practitioner or healthcare administrator.”);
providing the plurality of ECG data windows and the one or more demographics as input to: a first trained artificial intelligence (AI) model having a first framework (Zimmerman [0265] “In one instance, for the ECG trace models, a low-parameter convolutional neural network (CNN) was developed with 18,495 trainable parameters that consisted of six 1D CNN-Batch Normalization-ReLU (CBR) layer blocks followed by a two-layer multilayer perceptron and a final logistic output layer (Table 7). Specifically, Table 7 details a single output low-parameter CNN design for training on 8 non-derived ECG leads. The network contains a total of 18,945 trainable and 384 non-trainable parameters. Both Dropout layers were set at 25% drop rate. CBR is a brief notation for a sequence of 1D CNN, batch normalization, and ReLU layers.”, and
Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”);
cardiovascular score (Zimmerman [0125] “The training data database 124 can include a number of ECGs and clinical data. In some embodiments, the clinical data can include outcome data, such as whether or not a patient developed AF in a time period following the day that the ECG was taken. Exemplary time periods may include 1 month, 2 months, 3 months, 4 months, 5 months, 6 months, 7 months, 8 months, 9 months, 10 months, 11 months 12 months, 1 year, 2 years, 3 years, 4 years, 5 years, 6 years, 7 years, 8 years, 9 years, or 10 years. The ECGs and clinical data can be used for training a model to generate AF risk scores. In some embodiments, the training data database 124 can include multi-lead ECGs taken over a period of time (such as ten seconds) and corresponding clinical data. In some embodiments, the trained models database 128 can include a number of trained models that can receive raw ECGs and output AF risk scores. In other embodiments, a digital image of a lead for an ECG may be used. In some embodiments, trained models 136 can be stored in the computing device 104”);
identifying, based on the one or more demographics, an operating point threshold from a plurality of operating point thresholds (Zimmerman [0162] “To account for potential variability in the clinical implementation of such a model (matching the performance to the scope of available resources and desired screening characteristics), performance was evaluated across a range of operating points. An operating point can be the threshold of the model risk that was used to classify high or low risk for developing incident AF. For example, an operating point of 0.7 would indicate that model risk scores equal to and above 0.7 are considered high risk, and risk scores below 0.7 are low risk. Thus, overall model performance can be measured using AUROC and AUPRC scores that aggregate multiple operating point performances into a single metric. These points were defined based on maxima of the Fb score (for b=0.15, 0.5, 1, and 2) within the internal validation set. Fb scores are functions of precision and recall. A b value of 1 is the harmonic mean of precision and recall (e.g. sensitivity), a value of 2 emphasizes recall, and values of 0.15 and 0.5 attenuate the influence of recall correspondingly. Given the substantial variation in incidence of AF with age, the operating point was varied by age. The ECG with the highest risk for each patient acquired between the five-year period mentioned above was selected as the test set.”);
comparing the collective cardiovascular score to the operating point threshold; determining an indication of a cardiovascular condition based on the comparison; (Zimmerman [0194] “At 1316, the process can output the risk score to at least one of a memory (e.g., the memory 220 and/or the memory 240) or a display (e.g., the display 116, the display 208, and/or the display 228). In some embodiments, the display can be in view of a medical practitioner or healthcare administrator. In some embodiments, the process 1300 can generate and output a report based on the risk score. In some embodiments, the report can include the raw risk score and/or graphics related to the risk score. In some embodiments, the process 1300 can determine that the risk score is above a predetermined threshold associated with the condition (e.g., risk scores above the threshold can be indicative that the patient will suffer from the conditions within the predetermined time period). The process 1300 can then generate the report based on the determination that the risk score is above a predetermined threshold. In some embodiments, in response to determining that the risk score is above the predetermined threshold, the process 1300 can generate the report to include information (e.g., text) and/or links to sources (e.g., one or more hyperlinks) about treatments for the condition, causes of the condition, and/or other clinical information about the condition.”); and
providing a notification based on the indication of the cardiovascular condition (Zimmerman [0016] “The method may further include determining that the risk score is above a predetermined threshold associated with the condition, in response to determining that the risk score is above the predetermined threshold, generating a report including information and/or links to sources associated with at least one of treatments for the condition or causes of the condition, and outputting the report to at least one of a memory or a display for viewing by a medical practitioner or healthcare administrator.”).
Zimmerman fails to explicitly teach a second trained AI model having a second framework different from the first framework; receiving, via the first trained AI model, a first score; receiving, via the second trained AI model, a second score; and determining a collective cardiovascular score based on the first cardiovascular score and the second cardiovascular score.
Khosousi teaches a second trained AI model having a second framework different from the first framework (Khosousi [0076] “As used herein, the term “neural network” (and artificial neural networks (ANN)) refers to a family or framework of machine learning algorithms inspired by biological neural networks that can be instantiated in computing hardware and trained to perform tasks, including to learn on a set of features generated from a training set of data, e.g., to optimize one or more predictive models, which can be applied to data sources with unknown outcomes. Neural networks with fully-connected layers can define a family of functions that are parameterized by the weights of the network elements. Deep neural networks are examples of such multi-layer interconnected neural networks configured to recognize patterns directly from data sets with minimal preprocessing. Examples of classes of deep neural networks includes, for example, but not limited to, feed-forward neural networks, recurrent neural network, multi-layer perceptrons (MLP), convolutional neural networks, recursive neural networks, deep belief networks, convolutional deep belief networks, self-organizing maps, deep Boltzmann machines, stacked de-noising auto-encoders, etc. Convolutional neural networks (“CNNs”), and the likes, are particularly optimized to recognize patterns directly from a multi-dimensional data set (e.g., images). Examples of popular convolutional neural networks include GoogLeNets, ResNets, ResNeXts, DenseNets, DualPathNets, etc., each of which can be applied to the prediction or estimation of presence or absence of a disease state. The neural network, in some embodiments, uses deep learning methods such as CNNs to classify multi-dimensional data sets into one or more positive classes and/or one more negative classes based on machine-extractable features. As used herein, reference(s) to one or more neural network(s) can include one or more instance(s) of neural network architecture of the same type as well as instances of one or more instance(s) or combination(s) of neural network architectures of different types.”,
Khosousi [0113] “In some embodiments, assessment system 110 is configured with more than one neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof. Each of the neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof, may receive the normalized beat-to-beat data sets and generate a set of predictors that are combined (e.g., via an aggregation operator 202 as shown in FIG. 2A; or via operators 202 a, 202 b as shown in FIG. 2B).”, and
Khosousi [0172] “As noted above, two CNN models (referred to as “Model 85” and “Model 129”) were observed to satisfy the selection criteria on the validation set. Table 7 shows hyperparameters of the two CNN models.”).
receiving, via the first trained AI model, a first score; receiving, via the second trained AI model, a second score (Khosousi [0112] “Referring still to FIG. 3 , assessment system 110 is configured to then input the normalized beat-to-beat data set to a neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.), or an ensemble(s) thereof). The system may apply the normalized beat-to-beat data set to train, at step 312, the neural network 132 (e.g., deep neural network, convolutional neural network, etc.), or ensemble(s) thereof. The system may alternatively apply the normalized beat-to-beat data set to be classified, at step 314, by the trained neural network (e.g., trained deep neural network, trained convolutional neural network, etc.), or trained ensemble(s), to predict the presence or non-presence of a disease state or condition (e.g., presence or non-presence of coronary artery disease or other condition) in a patient and/or presence/location of disease or condition in a coronary artery.”, and
Khosousi [0113] “In some embodiments, assessment system 110 is configured with more than one neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof. Each of the neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof, may receive the normalized beat-to-beat data sets and generate a set of predictors that are combined (e.g., via an aggregation operator 202 as shown in FIG. 2A; or via operators 202 a, 202 b as shown in FIG. 2B).”); and
determining a collective cardiovascular score based on the first cardiovascular score and the second cardiovascular score (Khosousi [0113] “In some embodiments, assessment system 110 is configured with more than one neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof. Each of the neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof, may receive the normalized beat-to-beat data sets and generate a set of predictors that are combined (e.g., via an aggregation operator 202 as shown in FIG. 2A; or via operators 202 a, 202 b as shown in FIG. 2B).”, and
Khosousi [0120] “Referring still to FIG. 5 , for each data preparation and learning step (referred to as an epoch), assessment system 110 executes a pass through all of the training data set and calculates an AUC score from a validation data set. Because a single phase-gradient biophysical data set (e.g., wide-band phase gradient biophysical signal data set) can be segmented into a plurality of windowed data sets, the system, in some embodiments, is configured to combine all, or a substantial portion of, predictions on the plurality of patient's heart beats via a mean operator to provide the combined AUC score. Assessment system 110 may perform the learning and evaluating steps in a loop that can be run across multiple machines simultaneously without synchronization.”).
Khosousi teaches analyzing segmented cardiac signal windows using multiple trained neural-network models, generating a respective cardiovascular predictor from each model, and combining the predictors through an aggregation operator. Khosousi further teaches that the multiple models may employ different neural network architectures, including models having different numbers of convolutional layers, filter sizes, activation functions, input frequencies, and pooling parameters. It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Zimmerman’s ECG-based cardiovascular risk prediction method to employ Khosousi’s multiple model aggregation architecture, such that the ECG data windows and patient demographic information are processed using first and second trained AI models having different frameworks and the respective cardiovascular scores are combined to generate a collective cardiovascular score. A person of ordinary skill would have been motivated to make this modification to leverage the complementary predictive characteristics of differently configured models and thereby improve the robustness and diagnostic accuracy of Zimmerman’s cardiovascular risk prediction.
The proposed modification would have amounted to applying Khosousi’s known ensemble technique according to its established function in Zimmerman’s analogous ECG analysis system, yielding the predictable result of a combined cardiovascular risk score derived from multiple trained models. In the resulting combination, the collective cardiovascular score generated through Khosousi’s aggregation operator would serve as the final model risk score in Zimmerman’s system and would be compared with Zimmerman’s operating point threshold selected based on the subject’s demographic information, such as age, to determine whether the subject is at elevated risk for the cardiovascular condition and to provide the corresponding notification. One of ordinary skill would have had a reasonable expectation of success because both Zimmerman and Khosousi use trained machine learning models to analyze cardiac signal data and generate cardiovascular disease predictions, and Khosousi teaches combining outputs from multiple models for that purpose.
Regarding claim 2, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein determining the first cardiovascular score includes: generating, using the first trained AI model, a plurality of cardiovascular window scores, each corresponding to a different one of the plurality of ECG data windows (Khosousi [0108] “FIG. 4 is a diagram showing a beat-to-beat isolation of FIG. 3 , in accordance with an illustrative embodiment. As shown in FIG. 4 , assessment system 110 detects the maximum peaks (shown as 402 a, 402 b, 402 c) in the acquired biophysical-signal data set 108 (or an intermediary data set derived from the biophysical-signal data set such as the down-sampled signal data set) then isolates each beat as a data set defined in a fixed window (shown as 404 a, 404 b, 404 c) placed around the maximum peak (108 a, 108 b, 108 c) with the peak at the center of the window. In some embodiments, assessment system 110 a employs the Pan-Tompkins algorithm as described in Pan and Tompkins, “A Real-Time QRS Detection Algorithm,” IEEE Trans. Biomed. Eng., Vol. 32, No. 3, (March 1985), the entirety of which is hereby incorporated by reference herein, to detect peaks (402 a, 402 b, 402 c) in the down-sampled signal data set (108 a, 108 b, 108 c). In some embodiments, assessment system 110 generates a fixed-window of about 0.75 second, which corresponds to a heart rate of 80 beats per second. Other window sizes and centering techniques maybe used.”, and
Khosousi [0109] “In some embodiments, to preserve the phase-gradient information among the acquired biophysical data set (or the intermediary dataset being processed in the assessment analysis), assessment system 110 applies the same time window (e.g., 404 a, 404 b, 404 c) as obtained in the peak detection of the first channel (e.g., ORTH1) to extract the beats from one or more of the other channels (e.g., ORTH2 and ORTH3). As shown, assessment system 110 generates a first beat-to-beat segment from channel “1” (406 a) (also referred to as channel “ORTH1”) that are phase aligned with both a beat-to-beat segment from channel “2” (406 b) (also referred to as channel “ORTH2”) and a beat-to-beat segment from channel “3” (406 c) (also referred to as channel “ORTH3”), which can be used collectively as one input to the convolutional neural network. A second set of inputs are shown as a beat-to-beat segment from channel “1” (408 a), a beat-to-beat segment from channel “2” (408 b), and a beat-to-beat segment from channel “3” (408 c). A third set of inputs are shown as a beat-to-beat segment from channel “1” (410 a), a beat-to-beat segment from channel “2” (410 b), and a beat-to-beat segment from channel “3” (410 c). Indeed, output(s) of the Pre-Processing module 116 to be provided as the input(s) to the one or more neural networks (e.g., deep neural network such as convolutional neural networks, etc.), or ensemble(s) thereof, is a set of data segments (comprising a single complete cardiac cycle) from a phased-aligned time window (e.g., a 0.75-second window) from each, or a portion, of the acquisition channels. In some embodiments, all data segments extracted from the Pre-Processing module 116 are provided as input to the neural network(s) 132 (e.g., to the deep neural network(s), to the convolutional neural network(s), etc.), or ensemble(s) thereof (e.g., for training or analysis). In other embodiment, data segments extracted from some, but not all, of the acquisition channels are provided to the neural network(s) 132 (e.g., deep neural network(s), convolutional neural network(s), etc.), or ensemble(s) thereof (e.g., for training or analysis). In yet other embodiments, data segments extracted from some, but not all, of a given acquisition channel are provided to the neural network(s) 132 (e.g., deep neural network(s), convolutional neural network(s), etc.), or ensemble(s) thereof.”); and
determining the first cardiovascular score based on a combination of the plurality of cardiovascular window scores (Khosousi [0112] “Referring still to FIG. 3 , assessment system 110 is configured to then input the normalized beat-to-beat data set to a neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.), or an ensemble(s) thereof). The system may apply the normalized beat-to-beat data set to train, at step 312, the neural network 132 (e.g., deep neural network, convolutional neural network, etc.), or ensemble(s) thereof. The system may alternatively apply the normalized beat-to-beat data set to be classified, at step 314, by the trained neural network (e.g., trained deep neural network, trained convolutional neural network, etc.), or trained ensemble(s), to predict the presence or non-presence of a disease state or condition (e.g., presence or non-presence of coronary artery disease or other condition) in a patient and/or presence/location of disease or condition in a coronary artery.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to configure the modified Zimmerman system to generate a cardiovascular prediction for each segmented ECG data window and to combine the resulting window level predictions into a single cardiovascular score, as taught by Khosousi. Khosousi teaches segmenting cardiac signals into a plurality of beat to beat windows, inputting each window to a trained neural network to generate a respective disease prediction, and combining the predictions from the plurality of windows using a mean operator to produce a final prediction for the subject. A person of ordinary skill in the art would have been motivated to incorporate this known prediction aggregation technique into Zimmerman's ECG analysis because combining predictions from multiple ECG windows reduces the influence of noise or transient signal variations present in individual cardiac cycles, improves the robustness of the overall cardiovascular assessment, and represents the predictable use of a known aggregation technique to improve the accuracy and reliability of machine learning cardiovascular disease prediction, with a reasonable expectation of success.
Regarding claim 3, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein the collective cardiovascular score includes a score of pulmonary hypertension, and the cardiovascular condition is pulmonary hypertension (Khosousi [0090] “The assessment system 110 determines a value (e.g., risk/likelihood, binary indication) indicative of presence of cardiac disease or condition (e.g., coronary arterial disease, pulmonary hypertension, pulmonary arterial hypertension, left heart failure, right-sided heart failure, abnormal left-ventricular end diastolic pressure (LVEDP)) by directly inputting the pre-processed data set to a set of one or more neural networks 232 a, 232 b (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or ensemble(s) thereof, trained with one or more biophysical signal data sets acquired from a plurality of patients labeled with a diagnosis of presence of coronary artery disease (e.g., significant coronary artery disease). In some embodiments, the label for the presence of coronary artery disease comprises a Gensini-based score determined as a combination of a severity-weighted scoring and location-weighted scoring for a coronary lesion diagnosed in the patient's myocardium.”, and
Khosousi [0113] “In some embodiments, assessment system 110 is configured with more than one neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof. Each of the neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof, may receive the normalized beat-to-beat data sets and generate a set of predictors that are combined (e.g., via an aggregation operator 202 as shown in FIG. 2A; or via operators 202 a, 202 b as shown in FIG. 2B).”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to configure the modified Zimmerman system to generate a collective cardiovascular score indicative of pulmonary hypertension, as taught by Khosousi. Khosousi teaches that one or more trained neural networks, or ensembles thereof, may determine a value, such as a risk, likelihood, or binary indication, for cardiovascular conditions including pulmonary hypertension and pulmonary arterial hypertension, and further teaches combining outputs from multiple neural networks through an aggregation operator. A person of ordinary skill in the art would have been motivated to incorporate Khosousi's disease specific prediction capability into Zimmerman's ECG cardiovascular assessment system to extend the system to identify additional clinically significant cardiovascular conditions using the same AI analysis framework. Doing this would have represented the predictable use of known machine learning techniques to evaluate another recognized cardiovascular disease from cardiac signal data, which improves the clinical utility of the system while yielding predictable results with a reasonable expectation of success.
Regarding claim 4, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein segmenting the received ECG data is based on one or more of a step size or a window size (Khosousi [0108] “FIG. 4 is a diagram showing a beat-to-beat isolation of FIG. 3 , in accordance with an illustrative embodiment. As shown in FIG. 4 , assessment system 110 detects the maximum peaks (shown as 402 a, 402 b, 402 c) in the acquired biophysical-signal data set 108 (or an intermediary data set derived from the biophysical-signal data set such as the down-sampled signal data set) then isolates each beat as a data set defined in a fixed window (shown as 404 a, 404 b, 404 c) placed around the maximum peak (108 a, 108 b, 108 c) with the peak at the center of the window. In some embodiments, assessment system 110 a employs the Pan-Tompkins algorithm as described in Pan and Tompkins, “A Real-Time QRS Detection Algorithm,” IEEE Trans. Biomed. Eng., Vol. 32, No. 3, (March 1985), the entirety of which is hereby incorporated by reference herein, to detect peaks (402 a, 402 b, 402 c) in the down-sampled signal data set (108 a, 108 b, 108 c). In some embodiments, assessment system 110 generates a fixed-window of about 0.75 second, which corresponds to a heart rate of 80 beats per second. Other window sizes and centering techniques maybe used.”, and
Khosousi [ [0109] “In some embodiments, to preserve the phase-gradient information among the acquired biophysical data set (or the intermediary dataset being processed in the assessment analysis), assessment system 110 applies the same time window (e.g., 404 a, 404 b, 404 c) as obtained in the peak detection of the first channel (e.g., ORTH1) to extract the beats from one or more of the other channels (e.g., ORTH2 and ORTH3). As shown, assessment system 110 generates a first beat-to-beat segment from channel “1” (406 a) (also referred to as channel “ORTH1”) that are phase aligned with both a beat-to-beat segment from channel “2” (406 b) (also referred to as channel “ORTH2”) and a beat-to-beat segment from channel “3” (406 c) (also referred to as channel “ORTH3”), which can be used collectively as one input to the convolutional neural network. A second set of inputs are shown as a beat-to-beat segment from channel “1” (408 a), a beat-to-beat segment from channel “2” (408 b), and a beat-to-beat segment from channel “3” (408 c). A third set of inputs are shown as a beat-to-beat segment from channel “1” (410 a), a beat-to-beat segment from channel “2” (410 b), and a beat-to-beat segment from channel “3” (410 c). Indeed, output(s) of the Pre-Processing module 116 to be provided as the input(s) to the one or more neural networks (e.g., deep neural network such as convolutional neural networks, etc.), or ensemble(s) thereof, is a set of data segments (comprising a single complete cardiac cycle) from a phased-aligned time window (e.g., a 0.75-second window) from each, or a portion, of the acquisition channels. In some embodiments, all data segments extracted from the Pre-Processing module 116 are provided as input to the neural network(s) 132 (e.g., to the deep neural network(s), to the convolutional neural network(s), etc.), or ensemble(s) thereof (e.g., for training or analysis). In other embodiment, data segments extracted from some, but not all, of the acquisition channels are provided to the neural network(s) 132 (e.g., deep neural network(s), convolutional neural network(s), etc.), or ensemble(s) thereof (e.g., for training or analysis). In yet other embodiments, data segments extracted from some, but not all, of a given acquisition channel are provided to the neural network(s) 132 (e.g., deep neural network(s), convolutional neural network(s), etc.), or ensemble(s) thereof.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to configure Zimmerman's ECG segmentation to utilize a selected window size as taught by Khosousi. Khosousi expressly teaches segmenting cardiac signals into fixed windows centered on detected cardiac peaks, such as a 0.75 second window, and further teaches that other window sizes may be selected depending on the implementation. A person of ordinary skill in the art would have been motivated to incorporate Khosousi's window segmentation technique into Zimmerman's ECG analysis because selecting an appropriate window size ensures that complete cardiac cycles are consistently captured for analysis, reduces signal variability, and improves the quality and consistency of the data supplied to the trained AI models, thereby predictably improving the robustness and accuracy of the cardiovascular assessment with a reasonable expectation of success.
Regarding claim 5, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein determining the collective cardiovascular score includes: computing an average of the first cardiovascular score and the second cardiovascular score (Khosousi [0113] “In some embodiments, assessment system 110 is configured with more than one neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof. Each of the neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof, may receive the normalized beat-to-beat data sets and generate a set of predictors that are combined (e.g., via an aggregation operator 202 as shown in FIG. 2A; or via operators 202 a, 202 b as shown in FIG. 2B).”, and
Khosousi [0120] “Referring still to FIG. 5 , for each data preparation and learning step (referred to as an epoch), assessment system 110 executes a pass through all of the training data set and calculates an AUC score from a validation data set. Because a single phase-gradient biophysical data set (e.g., wide-band phase gradient biophysical signal data set) can be segmented into a plurality of windowed data sets, the system, in some embodiments, is configured to combine all, or a substantial portion of, predictions on the plurality of patient's heart beats via a mean operator to provide the combined AUC score. Assessment system 110 may perform the learning and evaluating steps in a loop that can be run across multiple machines simultaneously without synchronization.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to determine the collective cardiovascular score by computing an average of the first and second cardiovascular scores, as taught by Khosousi. Khosousi teaches combining multiple cardiovascular predictions using a mean operator, which is a well known averaging technique, to generate a combined prediction. A person of ordinary skill in the art would have been motivated to employ Khosousi's averaging technique as the aggregation operation in the modified Zimmerman system because averaging multiple model predictions reduces the influence of variability associated with any individual model, improves the stability and robustness of the resulting cardiovascular assessment, and represents the predictable use of a known statistical aggregation technique to improve machine learning prediction performance, with a reasonable expectation of success.
Regarding claim 6, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein determining the collective cardiovascular score includes: providing the first cardiovascular score and the second cardiovascular score as input to a classifier (Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”, and
Khosousi [0113] “In some embodiments, assessment system 110 is configured with more than one neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof. Each of the neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof, may receive the normalized beat-to-beat data sets and generate a set of predictors that are combined (e.g., via an aggregation operator 202 as shown in FIG. 2A; or via operators 202 a, 202 b as shown in FIG. 2B).”); and
determining the collective cardiovascular score based on output of the classifier (Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to determine the collective cardiovascular score by providing the first and second cardiovascular scores as inputs to a classifier and determining the collective cardiovascular score based on the classifier's output. Zimmerman teaches generating outputs from multiple independently trained neural networks, concatenating those outputs into a feature vector, and supplying the feature vector to a classification pipeline, such as an XGBoost classifier, to generate a recommendation score. In view of Khosousi's teaching of generating multiple cardiovascular predictors from separate neural networks, a person of ordinary skill in the art would have been motivated to use those predictors as inputs to Zimmerman's classifier to produce a single collective cardiovascular score. Doing this would have predictably improve diagnostic performance by allowing the classifier to integrate complementary information from multiple AI models into a unified cardiovascular assessment, with a reasonable expectation of success.
Regarding claim 7, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein determining the indication of the cardiovascular condition includes: determining a binary prediction of whether the subject has the cardiovascular condition based on the comparison (Zimmerman [0162] “To account for potential variability in the clinical implementation of such a model (matching the performance to the scope of available resources and desired screening characteristics), performance was evaluated across a range of operating points. An operating point can be the threshold of the model risk that was used to classify high or low risk for developing incident AF. For example, an operating point of 0.7 would indicate that model risk scores equal to and above 0.7 are considered high risk, and risk scores below 0.7 are low risk. Thus, overall model performance can be measured using AUROC and AUPRC scores that aggregate multiple operating point performances into a single metric. These points were defined based on maxima of the Fb score (for b=0.15, 0.5, 1, and 2) within the internal validation set. Fb scores are functions of precision and recall. A b value of 1 is the harmonic mean of precision and recall (e.g. sensitivity), a value of 2 emphasizes recall, and values of 0.15 and 0.5 attenuate the influence of recall correspondingly. Given the substantial variation in incidence of AF with age, the operating point was varied by age. The ECG with the highest risk for each patient acquired between the five-year period mentioned above was selected as the test set.”; and
Zimmerman [0194] “At 1316, the process can output the risk score to at least one of a memory (e.g., the memory 220 and/or the memory 240) or a display (e.g., the display 116, the display 208, and/or the display 228). In some embodiments, the display can be in view of a medical practitioner or healthcare administrator. In some embodiments, the process 1300 can generate and output a report based on the risk score. In some embodiments, the report can include the raw risk score and/or graphics related to the risk score. In some embodiments, the process 1300 can determine that the risk score is above a predetermined threshold associated with the condition (e.g., risk scores above the threshold can be indicative that the patient will suffer from the conditions within the predetermined time period). The process 1300 can then generate the report based on the determination that the risk score is above a predetermined threshold. In some embodiments, in response to determining that the risk score is above the predetermined threshold, the process 1300 can generate the report to include information (e.g., text) and/or links to sources (e.g., one or more hyperlinks) about treatments for the condition, causes of the condition, and/or other clinical information about the condition.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to determine a binary prediction of whether the subject has the cardiovascular condition based on the comparison of the collective cardiovascular score to the selected operating point threshold. Zimmerman =teaches comparing a cardiovascular risk score to an operating point threshold to classify a patient according to whether the patient is predicted to have a cardiovascular condition. After modifying Zimmerman to determine a collective cardiovascular score using the techniques taught by Khosousi, a person of ordinary skill in the art would have been motivated to apply Zimmerman's threshold comparison to the collective score to produce a binary prediction because thresholding continuous prediction scores into positive or negative classifications is a well known and predictable technique for facilitating clinical diagnosis and decision making, with a reasonable expectation of success.
Regarding claim 8, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach further comprising: providing the ECG data as input to a third trained AI model (Zimmerman [0265] “In one instance, for the ECG trace models, a low-parameter convolutional neural network (CNN) was developed with 18,495 trainable parameters that consisted of six 1D CNN-Batch Normalization-ReLU (CBR) layer blocks followed by a two-layer multilayer perceptron and a final logistic output layer (Table 7). Specifically, Table 7 details a single output low-parameter CNN design for training on 8 non-derived ECG leads. The network contains a total of 18,945 trainable and 384 non-trainable parameters. Both Dropout layers were set at 25% drop rate. CBR is a brief notation for a sequence of 1D CNN, batch normalization, and ReLU layers.”, and
Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”);
determining, using the third trained AI model, a third cardiovascular score (Zimmerman [0125] “The training data database 124 can include a number of ECGs and clinical data. In some embodiments, the clinical data can include outcome data, such as whether or not a patient developed AF in a time period following the day that the ECG was taken. Exemplary time periods may include 1 month, 2 months, 3 months, 4 months, 5 months, 6 months, 7 months, 8 months, 9 months, 10 months, 11 months 12 months, 1 year, 2 years, 3 years, 4 years, 5 years, 6 years, 7 years, 8 years, 9 years, or 10 years. The ECGs and clinical data can be used for training a model to generate AF risk scores. In some embodiments, the training data database 124 can include multi-lead ECGs taken over a period of time (such as ten seconds) and corresponding clinical data. In some embodiments, the trained models database 128 can include a number of trained models that can receive raw ECGs and output AF risk scores. In other embodiments, a digital image of a lead for an ECG may be used. In some embodiments, trained models 136 can be stored in the computing device 104”,
Khosousi [0112] “Referring still to FIG. 3 , assessment system 110 is configured to then input the normalized beat-to-beat data set to a neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.), or an ensemble(s) thereof). The system may apply the normalized beat-to-beat data set to train, at step 312, the neural network 132 (e.g., deep neural network, convolutional neural network, etc.), or ensemble(s) thereof. The system may alternatively apply the normalized beat-to-beat data set to be classified, at step 314, by the trained neural network (e.g., trained deep neural network, trained convolutional neural network, etc.), or trained ensemble(s), to predict the presence or non-presence of a disease state or condition (e.g., presence or non-presence of coronary artery disease or other condition) in a patient and/or presence/location of disease or condition in a coronary artery.”, and
Khosousi [0113] “In some embodiments, assessment system 110 is configured with more than one neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof. Each of the neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof, may receive the normalized beat-to-beat data sets and generate a set of predictors that are combined (e.g., via an aggregation operator 202 as shown in FIG. 2A; or via operators 202 a, 202 b as shown in FIG. 2B).”); and
determining the collective cardiovascular score based on the third cardiovascular score (Khosousi [0113] “In some embodiments, assessment system 110 is configured with more than one neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof. Each of the neural networks 132 a, 132 b (e.g., deep neural networks, convolutional neural networks, etc.), or ensemble(s) thereof, may receive the normalized beat-to-beat data sets and generate a set of predictors that are combined (e.g., via an aggregation operator 202 as shown in FIG. 2A; or via operators 202 a, 202 b as shown in FIG. 2B).”, and
Khosousi [0120] “Referring still to FIG. 5 , for each data preparation and learning step (referred to as an epoch), assessment system 110 executes a pass through all of the training data set and calculates an AUC score from a validation data set. Because a single phase-gradient biophysical data set (e.g., wide-band phase gradient biophysical signal data set) can be segmented into a plurality of windowed data sets, the system, in some embodiments, is configured to combine all, or a substantial portion of, predictions on the plurality of patient's heart beats via a mean operator to provide the combined AUC score. Assessment system 110 may perform the learning and evaluating steps in a loop that can be run across multiple machines simultaneously without synchronization.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify the combined Zimmerman and Khosousi system to provide the ECG data as input to a third trained AI model, determine a third cardiovascular score, and determine the collective cardiovascular score based on the third cardiovascular score. Zimmerman teaches employing multiple independently trained CNN models that receive ECG data as input to generate cardiovascular risk scores, while Khosousi teaches configuring an assessment system with more than one neural network, each generating a respective predictor that is combined using an aggregation operator to produce a collective prediction. A person of ordinary skill in the art would have been motivated to incorporate an additional trained AI model into the modified system because aggregating predictions from multiple independently trained models is a well known ensemble learning technique that improves robustness, reduces prediction variance, and increases diagnostic accuracy compared to relying on a single model. Extending the aggregation from two model outputs to include a third model output would have been a routine and predictable application of Khosousi's aggregation framework, yielding no more than the expected improvement in machine-learning performance with a reasonable expectation of success.
Regarding claim 9, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein the first framework includes a ResNet architecture or an Inception architecture (Khosousi [0076] “As used herein, the term “neural network” (and artificial neural networks (ANN)) refers to a family or framework of machine learning algorithms inspired by biological neural networks that can be instantiated in computing hardware and trained to perform tasks, including to learn on a set of features generated from a training set of data, e.g., to optimize one or more predictive models, which can be applied to data sources with unknown outcomes. Neural networks with fully-connected layers can define a family of functions that are parameterized by the weights of the network elements. Deep neural networks are examples of such multi-layer interconnected neural networks configured to recognize patterns directly from data sets with minimal preprocessing. Examples of classes of deep neural networks includes, for example, but not limited to, feed-forward neural networks, recurrent neural network, multi-layer perceptrons (MLP), convolutional neural networks, recursive neural networks, deep belief networks, convolutional deep belief networks, self-organizing maps, deep Boltzmann machines, stacked de-noising auto-encoders, etc. Convolutional neural networks (“CNNs”), and the likes, are particularly optimized to recognize patterns directly from a multi-dimensional data set (e.g., images). Examples of popular convolutional neural networks include GoogLeNets, ResNets, ResNeXts, DenseNets, DualPathNets, etc., each of which can be applied to the prediction or estimation of presence or absence of a disease state. The neural network, in some embodiments, uses deep learning methods such as CNNs to classify multi-dimensional data sets into one or more positive classes and/or one more negative classes based on machine-extractable features. As used herein, reference(s) to one or more neural network(s) can include one or more instance(s) of neural network architecture of the same type as well as instances of one or more instance(s) or combination(s) of neural network architectures of different types.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to implement the first trained AI model of the modified Zimmerman system using a ResNet architecture, as taught by Khosousi. Khosousi identifies ResNet as a known convolutional neural network architecture suitable for predicting the presence or absence of a disease state from multidimensional physiological data and further teaches that one or more neural networks may employ architectures of the same type or different types. A person of ordinary skill in the art would have been motivated to utilize a ResNet architecture in Zimmerman's ECG cardiovascular prediction system because ResNet was a well known deep-learning architecture designed to improve feature learning and training performance in deep neural networks through residual connections, which improves the robustness and predictive accuracy of cardiovascular disease detection. Substituting one known neural network architecture for another to perform the same predictive function would have been a predictable variation yielding no more than the expected improvement in machine learning performance, with a reasonable expectation of success.
Regarding claim 10, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein the one or more demographics include a representation of sex of the subject and a representation of age of the subject (Zimmerman [0036] “In one aspect, a method includes: receiving electrocardiogram data associated with a patient and an electrocardiogram configuration including a plurality of leads and a time interval, the electrocardiogram data comprising, for each lead included in the plurality of leads, voltage data associated with at least a portion of the time interval; receiving an age value associated with the patient; receiving a sex value associated with the patient; providing the age value, the sex value, and at least a portion of the electrocardiogram data to a trained model, the trained model being trained to generate a risk score based on input electrocardiogram data associated with the electrocardiogram configuration and supplementary information associated with the patient; receiving a risk score indicative of a likelihood the patient will suffer from aortic stenosis within a predetermined period of time from when the electrocardiogram data was generated; and outputting the risk score to at least one of a memory or a display for viewing by a medical practitioner or healthcare administrator.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to implement the demographics of the modified Zimmerman system as including representations of the subject's age and sex, as taught by Zimmerman. Zimmerman discloses receiving an age value and a sex value associated with a patient and providing those demographic values together with ECG data to a trained model for generating a cardiovascular risk score. A person of ordinary skill in the art would have recognized that age and sex are well known demographic factors that influence cardiovascular disease risk and are commonly incorporated into predictive models to improve diagnostic accuracy and individualized risk assessment. Accordingly, utilizing age and sex as the demographic representations in the modified system would have been a predictable implementation of known techniques with a reasonable expectation of success.
Regarding claim 11, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein identifying the operating point threshold includes: identifying the operating point threshold from at least four possible operating points based on the one or more demographics (Zimmerman [0162] “To account for potential variability in the clinical implementation of such a model (matching the performance to the scope of available resources and desired screening characteristics), performance was evaluated across a range of operating points. An operating point can be the threshold of the model risk that was used to classify high or low risk for developing incident AF. For example, an operating point of 0.7 would indicate that model risk scores equal to and above 0.7 are considered high risk, and risk scores below 0.7 are low risk. Thus, overall model performance can be measured using AUROC and AUPRC scores that aggregate multiple operating point performances into a single metric. These points were defined based on maxima of the Fb score (for b=0.15, 0.5, 1, and 2) within the internal validation set. Fb scores are functions of precision and recall. A b value of 1 is the harmonic mean of precision and recall (e.g. sensitivity), a value of 2 emphasizes recall, and values of 0.15 and 0.5 attenuate the influence of recall correspondingly. Given the substantial variation in incidence of AF with age, the operating point was varied by age. The ECG with the highest risk for each patient acquired between the five-year period mentioned above was selected as the test set.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to identify the operating point threshold from at least four possible operating points based on one or more demographics, as taught by Zimmerman. Zimmerman teaches selecting an operating point threshold from multiple candidate operating points based on patient demographic information, such as age, so that the classification threshold is tailored to the characteristics of the patient population. A person of ordinary skill in the art would have been motivated to employ multiple candidate operating points because selecting an appropriate threshold based on demographic information is a well known technique for balancing diagnostic sensitivity and specificity across different patient groups, thereby improving the accuracy and clinical usefulness of cardiovascular risk prediction with a reasonable expectation of success.
Regarding claim 12, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein identifying the operating point threshold includes: identifying the operating point threshold that minimizes a difference between a sensitivity and a specificity of indications of the cardiovascular condition determined using the first trained AI model and the second trained AI model for subjects having the one or more demographics (Zimmerman [0162] “To account for potential variability in the clinical implementation of such a model (matching the performance to the scope of available resources and desired screening characteristics), performance was evaluated across a range of operating points. An operating point can be the threshold of the model risk that was used to classify high or low risk for developing incident AF. For example, an operating point of 0.7 would indicate that model risk scores equal to and above 0.7 are considered high risk, and risk scores below 0.7 are low risk. Thus, overall model performance can be measured using AUROC and AUPRC scores that aggregate multiple operating point performances into a single metric. These points were defined based on maxima of the Fb score (for b=0.15, 0.5, 1, and 2) within the internal validation set. Fb scores are functions of precision and recall. A b value of 1 is the harmonic mean of precision and recall (e.g. sensitivity), a value of 2 emphasizes recall, and values of 0.15 and 0.5 attenuate the influence of recall correspondingly. Given the substantial variation in incidence of AF with age, the operating point was varied by age. The ECG with the highest risk for each patient acquired between the five-year period mentioned above was selected as the test set.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to identify the operating-point threshold that minimizes the difference between sensitivity and specificity for subjects having one or more demographics. Zimmerman teaches selecting an operating point threshold from multiple candidate operating points based on patient demographics so that the trained AI model provides an appropriate clinical classification. A person of ordinary skill in the art would have recognized that selecting an operating point that balances or minimizes the difference between sensitivity and specificity is a well known optimization technique for binary classifiers because it provides a clinically useful tradeoff between false positives and false negatives for a particular patient population. Accordingly, optimizing Zimmerman's demographic specific operating point selection to minimize the difference between sensitivity and specificity would have been a routine optimization of a known result-effective variable, yielding predictable improvements in classification performance with a reasonable expectation of success.
Regarding claim 13, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein providing the one or more demographics as input to the first trained AI model includes: providing the one or more demographics to a subject demographic block of the first trained AI model (Zimmerman [0036] “In one aspect, a method includes: receiving electrocardiogram data associated with a patient and an electrocardiogram configuration including a plurality of leads and a time interval, the electrocardiogram data comprising, for each lead included in the plurality of leads, voltage data associated with at least a portion of the time interval; receiving an age value associated with the patient; receiving a sex value associated with the patient; providing the age value, the sex value, and at least a portion of the electrocardiogram data to a trained model, the trained model being trained to generate a risk score based on input electrocardiogram data associated with the electrocardiogram configuration and supplementary information associated with the patient; receiving a risk score indicative of a likelihood the patient will suffer from aortic stenosis within a predetermined period of time from when the electrocardiogram data was generated; and outputting the risk score to at least one of a memory or a display for viewing by a medical practitioner or healthcare administrator.”, and
Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to provide the one or more demographics to a subject demographic block of the first trained AI model. Zimmerman teaches providing demographic information, including age and sex, as inputs to a trained AI model together with ECG data for cardiovascular risk prediction. A person of ordinary skill in the art would have understood that an AI model receiving dedicated demographic inputs necessarily includes a corresponding input component or portion configured to receive and process those demographic features. Implementing the demographic inputs through a dedicated subject demographic block represents a routine design choice that organizes known model inputs according to their function, yielding predictable processing of demographic information with a reasonable expectation of success.
Regarding claim 14, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein determining the first cardiovascular score using the first trained AI model includes: providing the plurality of ECG data windows as input to a convolutional block of the first trained AI model (Zimmerman [0265] “In one instance, for the ECG trace models, a low-parameter convolutional neural network (CNN) was developed with 18,495 trainable parameters that consisted of six 1D CNN-Batch Normalization-ReLU (CBR) layer blocks followed by a two-layer multilayer perceptron and a final logistic output layer (Table 7). Specifically, Table 7 details a single output low-parameter CNN design for training on 8 non-derived ECG leads. The network contains a total of 18,945 trainable and 384 non-trainable parameters. Both Dropout layers were set at 25% drop rate. CBR is a brief notation for a sequence of 1D CNN, batch normalization, and ReLU layers.”, and
Khosousi [0112] “Referring still to FIG. 3 , assessment system 110 is configured to then input the normalized beat-to-beat data set to a neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.), or an ensemble(s) thereof). The system may apply the normalized beat-to-beat data set to train, at step 312, the neural network 132 (e.g., deep neural network, convolutional neural network, etc.), or ensemble(s) thereof. The system may alternatively apply the normalized beat-to-beat data set to be classified, at step 314, by the trained neural network (e.g., trained deep neural network, trained convolutional neural network, etc.), or trained ensemble(s), to predict the presence or non-presence of a disease state or condition (e.g., presence or non-presence of coronary artery disease or other condition) in a patient and/or presence/location of disease or condition in a coronary artery.”);
providing the one or more demographics as input to a subject demographic block of the first trained AI model (Zimmerman [0036] “In one aspect, a method includes: receiving electrocardiogram data associated with a patient and an electrocardiogram configuration including a plurality of leads and a time interval, the electrocardiogram data comprising, for each lead included in the plurality of leads, voltage data associated with at least a portion of the time interval; receiving an age value associated with the patient; receiving a sex value associated with the patient; providing the age value, the sex value, and at least a portion of the electrocardiogram data to a trained model, the trained model being trained to generate a risk score based on input electrocardiogram data associated with the electrocardiogram configuration and supplementary information associated with the patient; receiving a risk score indicative of a likelihood the patient will suffer from aortic stenosis within a predetermined period of time from when the electrocardiogram data was generated; and outputting the risk score to at least one of a memory or a display for viewing by a medical practitioner or healthcare administrator.”); and
determining the first cardiovascular score based on output of the convolutional block and output of the subject demographic block (Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to determine the first cardiovascular score by providing the plurality of ECG data windows as input to a convolutional block of the first trained AI model, providing the one or more demographics as input to a subject demographic block of the first trained AI model, and determining the first cardiovascular score based on the outputs of the convolutional block and the subject demographic block. Zimmerman teaches a convolutional neural network comprising multiple convolutional layer blocks for processing ECG data and further teaches providing demographic information, including age and sex, together with ECG data to the trained AI model to generate a cardiovascular risk score. Khosousi teaches providing segmented beat to beat ECG windows as inputs to convolutional neural networks for cardiovascular disease prediction. A person of ordinary skill in the art would have been motivated to configure the modified system so that ECG windows are processed through the convolutional portion of the model while demographic information is provided through a dedicated demographic input component, with the resulting features combined to generate the cardiovascular score, because integrating learned ECG features with patient demographic information is a well known machine learning technique that improves predictive performance by allowing the model to utilize complementary sources of information. This modification would have represented the predictable use of known AI model architectures with a reasonable expectation of success.
Regarding claim 15, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein determining the first cardiovascular score using the first trained AI model includes: providing the plurality of ECG data windows as input to a convolutional block of the first trained AI model (Zimmerman [0265] “In one instance, for the ECG trace models, a low-parameter convolutional neural network (CNN) was developed with 18,495 trainable parameters that consisted of six 1D CNN-Batch Normalization-ReLU (CBR) layer blocks followed by a two-layer multilayer perceptron and a final logistic output layer (Table 7). Specifically, Table 7 details a single output low-parameter CNN design for training on 8 non-derived ECG leads. The network contains a total of 18,945 trainable and 384 non-trainable parameters. Both Dropout layers were set at 25% drop rate. CBR is a brief notation for a sequence of 1D CNN, batch normalization, and ReLU layers.”, and
Khosousi [0112] “Referring still to FIG. 3 , assessment system 110 is configured to then input the normalized beat-to-beat data set to a neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.), or an ensemble(s) thereof). The system may apply the normalized beat-to-beat data set to train, at step 312, the neural network 132 (e.g., deep neural network, convolutional neural network, etc.), or ensemble(s) thereof. The system may alternatively apply the normalized beat-to-beat data set to be classified, at step 314, by the trained neural network (e.g., trained deep neural network, trained convolutional neural network, etc.), or trained ensemble(s), to predict the presence or non-presence of a disease state or condition (e.g., presence or non-presence of coronary artery disease or other condition) in a patient and/or presence/location of disease or condition in a coronary artery.”);
providing the one or more demographics as input to a subject demographic block of the first trained AI model (Zimmerman [0036] “In one aspect, a method includes: receiving electrocardiogram data associated with a patient and an electrocardiogram configuration including a plurality of leads and a time interval, the electrocardiogram data comprising, for each lead included in the plurality of leads, voltage data associated with at least a portion of the time interval; receiving an age value associated with the patient; receiving a sex value associated with the patient; providing the age value, the sex value, and at least a portion of the electrocardiogram data to a trained model, the trained model being trained to generate a risk score based on input electrocardiogram data associated with the electrocardiogram configuration and supplementary information associated with the patient; receiving a risk score indicative of a likelihood the patient will suffer from aortic stenosis within a predetermined period of time from when the electrocardiogram data was generated; and outputting the risk score to at least one of a memory or a display for viewing by a medical practitioner or healthcare administrator.”);
providing output of the convolutional block and output of the subject demographic block as input to one or more fully connected layers (Zimmerman [0265] “In one instance, for the ECG trace models, a low-parameter convolutional neural network (CNN) was developed with 18,495 trainable parameters that consisted of six 1D CNN-Batch Normalization-ReLU (CBR) layer blocks followed by a two-layer multilayer perceptron and a final logistic output layer (Table 7). Specifically, Table 7 details a single output low-parameter CNN design for training on 8 non-derived ECG leads. The network contains a total of 18,945 trainable and 384 non-trainable parameters. Both Dropout layers were set at 25% drop rate. CBR is a brief notation for a sequence of 1D CNN, batch normalization, and ReLU layers.”, and
Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”); and
determining the first cardiovascular score based on output of the one or more fully connected layers (Zimmerman [0265] “In one instance, for the ECG trace models, a low-parameter convolutional neural network (CNN) was developed with 18,495 trainable parameters that consisted of six 1D CNN-Batch Normalization-ReLU (CBR) layer blocks followed by a two-layer multilayer perceptron and a final logistic output layer (Table 7). Specifically, Table 7 details a single output low-parameter CNN design for training on 8 non-derived ECG leads. The network contains a total of 18,945 trainable and 384 non-trainable parameters. Both Dropout layers were set at 25% drop rate. CBR is a brief notation for a sequence of 1D CNN, batch normalization, and ReLU layers.”, and
Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to determine the first cardiovascular score by providing the plurality of ECG data windows as input to a convolutional block of the first trained AI model, providing the one or more demographics as input to a subject demographic block, providing the outputs of the convolutional block and the subject demographic block as inputs to one or more fully connected layers, and determining the first cardiovascular score based on the output of the fully connected layers. Zimmerman teaches a convolutional neural network architecture comprising multiple convolutional layer blocks followed by a multilayer perceptron having fully connected layers that generate a cardiovascular risk score, while also teaching the use of demographic information, including age and sex, as inputs to the trained model. Khosousi teaches supplying segmented ECG windows to a convolutional neural network for cardiovascular disease prediction. A person of ordinary skill in the art would have been motivated to combine ECG-derived features and demographic information before processing them through fully connected layers because integrating complementary feature types within the fully connected portion of a neural network is a well known machine learning architecture that improves predictive performance by enabling the model to learn relationships between physiological signal features and patient-specific characteristics. This modification would have been a predictable application of known neural-network design techniques with a reasonable expectation of success.
Regarding claim 16, Zimmerman and Khosousi teach the invention in claim 1, as discussed above, and further teach wherein providing the notification based on the indication of the cardiovascular condition includes: providing the notification recommending that the subject be scheduled for a diagnostic test to confirm the indication of the cardiovascular condition (Zimmerman [0302] “A patient identified as high-risk may prompt a notification, either at the time of diagnostic testing using the ECG or during a patient evaluation period where the patient's records are scanned for high-risk events. Upon notification, a physician may consider increased, more aggressive monitoring for their patient or request the patient receive more diagnostic testing, such as imaging of their heart via an ultrasound/echocardiography.”, and
Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to provide a notification recommending that the subject be scheduled for a diagnostic test to confirm the indication of the cardiovascular condition, as taught by Zimmerman. Zimmerman teaches that when a patient is identified as high risk, a notification is generated that prompts a physician to consider increased monitoring or request that the patient undergo additional diagnostic testing, such as cardiac imaging by ultrasound or echocardiography. A person of ordinary skill in the art would have recognized that providing such a notification facilitates appropriate clinical follow up by recommending confirmatory diagnostic testing after an AI cardiovascular risk indication, which improves patient evaluation and enabling timely diagnosis using known clinical practices, with a reasonable expectation of success.
Claim 17 is analogous to claim 1, thus claim 17 is similarly analyzed and rejected in a manner consistent with the rejection of claim 1.
Regarding claim 18, Zimmerman and Khosousi teach the invention in claim 17, as discussed above, and further teach wherein the one or more processors provide the subject demographics as input to the plurality of trained AI models by being further configured to: provide the subject demographics to a subject demographic block of each trained AI model of the ensemble of trained AI models (Zimmerman [0036] “In one aspect, a method includes: receiving electrocardiogram data associated with a patient and an electrocardiogram configuration including a plurality of leads and a time interval, the electrocardiogram data comprising, for each lead included in the plurality of leads, voltage data associated with at least a portion of the time interval; receiving an age value associated with the patient; receiving a sex value associated with the patient; providing the age value, the sex value, and at least a portion of the electrocardiogram data to a trained model, the trained model being trained to generate a risk score based on input electrocardiogram data associated with the electrocardiogram configuration and supplementary information associated with the patient; receiving a risk score indicative of a likelihood the patient will suffer from aortic stenosis within a predetermined period of time from when the electrocardiogram data was generated; and outputting the risk score to at least one of a memory or a display for viewing by a medical practitioner or healthcare administrator.”, and
Zimmerman [0266] “Each CNN layer consisted of 16 kernels of size 5. The same network configuration was used to train one model per clinical outcome, resulting in 7 independently trained CNN models (FIG. 25B). Specifically, FIG. 25B displays a block diagram for a composite model that shows the classification pipeline for ECG trace and other EHR data. The output of each neural network (the triangles in FIG. 25B) applied to ECG trace data is concatenated to labs, vitals, and demographics to form a feature vector. The vector is the input to a classification pipeline (min-max scaling, mean imputation, and XGBoost classifier), which outputs a recommendation score for the patient.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to configure the one or more processors to provide the subject demographics to a subject demographic block of each trained AI model of the ensemble. Zimmerman teaches providing demographic information, including age and sex, as inputs to a trained AI model for cardiovascular risk prediction, while Khosousi teaches employing an ensemble of multiple trained neural networks for cardiovascular assessment. A person of ordinary skill in the art would have been motivated to supply the same demographic inputs to each trained AI model within the ensemble because each model relies on the available patient information to generate its respective prediction, and providing identical demographic inputs to each model ensures consistent feature availability across the ensemble. Implementing a dedicated demographic input component for each trained AI model represents a routine architectural design choice and the predictable application of known machine learning techniques, with a reasonable expectation of success.
Claim 19 is analogous to claim 1, thus claim 19 is similarly analyzed and rejected in a manner consistent with the rejection of claim 1.
Regarding claim 20, Zimmerman and Khosousi teach the invention in claim 19, as discussed above, and further teach wherein obtaining the plurality of ECG data windows based on the ECG data includes: segmenting the ECG data into the plurality ECG data windows, wherein each ECG data window overlaps with at least one other ECG data window (Khosousi [0108] “FIG. 4 is a diagram showing a beat-to-beat isolation of FIG. 3 , in accordance with an illustrative embodiment. As shown in FIG. 4 , assessment system 110 detects the maximum peaks (shown as 402 a, 402 b, 402 c) in the acquired biophysical-signal data set 108 (or an intermediary data set derived from the biophysical-signal data set such as the down-sampled signal data set) then isolates each beat as a data set defined in a fixed window (shown as 404 a, 404 b, 404 c) placed around the maximum peak (108 a, 108 b, 108 c) with the peak at the center of the window. In some embodiments, assessment system 110 a employs the Pan-Tompkins algorithm as described in Pan and Tompkins, “A Real-Time QRS Detection Algorithm,” IEEE Trans. Biomed. Eng., Vol. 32, No. 3, (March 1985), the entirety of which is hereby incorporated by reference herein, to detect peaks (402 a, 402 b, 402 c) in the down-sampled signal data set (108 a, 108 b, 108 c). In some embodiments, assessment system 110 generates a fixed-window of about 0.75 second, which corresponds to a heart rate of 80 beats per second. Other window sizes and centering techniques maybe used.”,
Khosousi [0109] “In some embodiments, to preserve the phase-gradient information among the acquired biophysical data set (or the intermediary dataset being processed in the assessment analysis), assessment system 110 applies the same time window (e.g., 404 a, 404 b, 404 c) as obtained in the peak detection of the first channel (e.g., ORTH1) to extract the beats from one or more of the other channels (e.g., ORTH2 and ORTH3). As shown, assessment system 110 generates a first beat-to-beat segment from channel “1” (406 a) (also referred to as channel “ORTH1”) that are phase aligned with both a beat-to-beat segment from channel “2” (406 b) (also referred to as channel “ORTH2”) and a beat-to-beat segment from channel “3” (406 c) (also referred to as channel “ORTH3”), which can be used collectively as one input to the convolutional neural network. A second set of inputs are shown as a beat-to-beat segment from channel “1” (408 a), a beat-to-beat segment from channel “2” (408 b), and a beat-to-beat segment from channel “3” (408 c). A third set of inputs are shown as a beat-to-beat segment from channel “1” (410 a), a beat-to-beat segment from channel “2” (410 b), and a beat-to-beat segment from channel “3” (410 c). Indeed, output(s) of the Pre-Processing module 116 to be provided as the input(s) to the one or more neural networks (e.g., deep neural network such as convolutional neural networks, etc.), or ensemble(s) thereof, is a set of data segments (comprising a single complete cardiac cycle) from a phased-aligned time window (e.g., a 0.75-second window) from each, or a portion, of the acquisition channels. In some embodiments, all data segments extracted from the Pre-Processing module 116 are provided as input to the neural network(s) 132 (e.g., to the deep neural network(s), to the convolutional neural network(s), etc.), or ensemble(s) thereof (e.g., for training or analysis). In other embodiment, data segments extracted from some, but not all, of the acquisition channels are provided to the neural network(s) 132 (e.g., deep neural network(s), convolutional neural network(s), etc.), or ensemble(s) thereof (e.g., for training or analysis). In yet other embodiments, data segments extracted from some, but not all, of a given acquisition channel are provided to the neural network(s) 132 (e.g., deep neural network(s), convolutional neural network(s), etc.), or ensemble(s) thereof.”, and
Khosousi [0120] “Referring still to FIG. 5 , for each data preparation and learning step (referred to as an epoch), assessment system 110 executes a pass through all of the training data set and calculates an AUC score from a validation data set. Because a single phase-gradient biophysical data set (e.g., wide-band phase gradient biophysical signal data set) can be segmented into a plurality of windowed data sets, the system, in some embodiments, is configured to combine all, or a substantial portion of, predictions on the plurality of patient's heart beats via a mean operator to provide the combined AUC score. Assessment system 110 may perform the learning and evaluating steps in a loop that can be run across multiple machines simultaneously without synchronization.”).
It would have been obvious to one of ordinary skill in the art before the effective filing date to configure the ECG segmentation such that each ECG data window overlaps with at least one other ECG data window. Khosousi teaches segmenting ECG data into a plurality of fixed, beat centered windows that are provided as inputs to one or more neural networks for cardiovascular analysis and further teaches processing a plurality of windowed data sets generated from a single ECG recording. A person of ordinary skill in the art would have been motivated to employ overlapping adjacent ECG windows because the amount of overlap between successive windows is a well known result effective variable in digital signal processing and machine learning preprocessing. Using overlapping windows predictably preserves waveform features occurring near window boundaries, reduces the likelihood of omitting diagnostically relevant information between successive segments, and improves the robustness and continuity of feature extraction for AI cardiovascular analysis, with a reasonable expectation of success.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Nemani et al. (U.S. Patent Publication 2024/0221936 A1) teaches a computer implemented system for configuring a single architecture deep learning model by receiving model parameters, configuring model segments, and processing segment, combination, and fully connected outputs to generate a predictive modeling output.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to KYRA R LAGOY whose telephone number is (703)756-1773. The examiner can normally be reached Monday - Friday, 8:00 am - 5:00 pm EST.
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, Kambiz Abdi can be reached at (571)272-6702. 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.
/K.R.L./
Examiner, Art Unit 3685
/KAMBIZ ABDI/Supervisory Patent Examiner, Art Unit 3685
Prosecution Insights