DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Status of the Claims
The status of the claims as of the response filed 10/08/2025, is as follows:
Claims 1-17 are pending.
Claims None are canceled.
The applicant has amended Claims 1, 2, 5, 7-10, 12, and 14-17 are amended and have been considered below.
Information Disclosure Statement
The information disclosure statements (IDS) submitted on 2025-07-22 are in
accordance with the provisions of 37 CFR 1.97 and are considered by the Examiner.
Response to Amendments
Double Patenting
Applicant’s arguments, see page 1 filed date 10/8/25 with respect to Claims 1-17 have been fully considered and are persuasive The Double Patenting
rejection is withdrawn.
The applicant requests the withdrawal of the double patenting rejection by submitting a Terminal Disclaimer. They argue that this filing cures the rejection without admitting that the invention was obvious (citing Quad Environmental Technologies Corp. v. Union Sanitary District).
Examiner respectfully agrees because the applicant has filed a complaint Terminal Disclaimer in accordance with 37 CFR 1.321(c).
Response to Arguments
35 U.S.C. § 101 - Subject Matter Eligibility
Applicants’ arguments, see page 11-13 filed date 10/8/25 with respect to Claims 1-17 have been fully considered and are not persuasive. The 35 U.S.C. § 101 rejection is sustained.
The Applicant asserts that amended Claim 1 recites a "specific technical pipeline" (generating identity-labeled aggregated data, training a binary determination model, and executing time-reversed priority resetting) that integrates the abstract idea into a practical application by controlling subsequent data acquisition and improving efficiency, and further argues that this specific logic constitutes an inventive concept that is not well-understood, routine, or conventional.
The Examiner is respectfully disagreeing and maintains the rejection because the claimed "technical pipeline" consists entirely of Mental Processes (evaluating data for identity, training models, calculating required numbers) which remain directed to the abstract idea under MPEP 2106.04(a)(2). The alleged practical application of "controlling data acquisition" constitutes Organization of Human Activity (managing a workflow), which fail to integrate the exception into a practical application as they do not improve the specific technical functioning of the computer itself (MPEP 2106.05(a)). Furthermore, the claims lack an inventive concept under Step 2B because the hardware components are generic (MPEP 2106.05(f)), and the specific algorithmic steps cannot supply the inventive concept as they are the abstract idea itself. The claim language recitations it focus on improve the abstract process “a feature selection optimization process to mathematically prove which patient data points are redundant and automatically reconfiguring the data collection requirements to exclude them to obtain same result that with the full data” efficiency, not computer efficiency. MPEP 2106.05(a) Consequently, the 35 U.S.C. § 101 rejections is sustained.
35 U.S.C. § 103
Applicant’s arguments, see page 14-16, filed date 10/8/25, with respect to Claims 1-17 have been fully considered and are not persuasive. The 35 U.S.C. § 103 rejection is sustained.
The Applicant asserts that the cited references (Yang and Chang) do not disclose or suggest the specific limitations of "generating aggregated data (X', y')," "training a binary determination model," and the "resetting... by inserting... from a later elapsed period" logic. Specifically, the applicant argues that Yang relates to predictive scoring and Chang to care delivery, and neither teaches the specific "technical pipeline" of training a secondary model to control data acquisition based on future requirements.
The Examiner respectfully disagrees. The combination of Yang (temporal framework), Chang '562 (feature importance calculation), and Chang et al. 2019 (dynamic scheduling logic) teaches every step of the claimed pipeline.
1. Regarding "Aggregated Data (X', y')" and "Binary Determination Model"
The Examiner maintains that Chang '562 explicitly discloses this data generation and thresholding process, as detailed in the previous Office Action:
Aggregated Data (X', y'): Chang '562 generates "modified feature values" (X') and compares the resulting ranking to the original ranking to generate a "rank-biased overlap" (RBO) score (y') (Chang '562, [0044]-[0048]). These scores are aggregated into an "importance score" (Chang '562, [0050]).
Binary Determination Model: Chang '562 applies a "feature removal threshold" to these importance scores to classify features into two categories: "high-importance" (keep) or "low-importance" (remove) (Chang '562, [0051]). This thresholding logic functions as the claimed binary determination model.
2. Regarding "Resetting... Earlier... Based on Later" (The "Technical Pipeline")
The Applicant argues that the "time-reversed" logic—updating an earlier model's requirements based on a later model's needs—is missing. To address this, the Examiner cites Chang et al. (2019), which explicitly discloses this exact "technical pipeline" of Dynamic Measurement Scheduling.
Claim Limitation: "resetting the priority... of an earlier elapsed period... based on... a model of a later elapsed period."
Chang et al. (2019) Disclosure: Chang et al. teaches a system that answers the question "What and when should be measured... to scale [measurements]... by scheduling strategically-timed measurements" (Abstract).
The "Earlier" Policy: Chang et al. trains an agent (Deep Q-Network) to determine the measurement policy at the current time step (t). This corresponds to the "priority feature-value types associated with a model of an earlier elapsed period."
The "Later" Requirement: The policy is learned by "minimizing the Bellman-equation square error" (Section 3.2), which updates the current action value based on the "Q-target" (future reward/predictive gain) of the next time step (t+1).
The "Resetting": Chang et al. explicitly teaches "resetting" the policy for the earlier period during the training phase. The system looks at the predictive gain achieved in the "Later Elapsed Period" (the future reward) and backpropagates this to update the policy for the "Earlier Elapsed Period" (the current state).
The Applicant argues that the prior art cannot "look ahead." However, Chang et al. (2019) clarifies that this "look ahead" capability is the fundamental definition of Reinforcement Learning training as applied to measurement scheduling. By training on historical longitudinal data (as taught by Yang), the system uses the known outcomes of Later Elapsed Periods to optimize ("reset") the data collection instructions for Earlier Elapsed Periods.
It would have been obvious to a Person Having Ordinary Skill in the Art (PHOSITA) to combine these references. Yang provides the longitudinal windows. Chang '562 provides the mathematical method for scoring feature importance (RBO). Chang et al. (2019) provides the motivation and mechanism to use those scores to dynamically update the scheduling policy. As Chang et al. (2019) states, this approach reduces the total number of measurements by 31%... or improves predictive gain, providing a strong motivation for the combination (Chang et al. 2019, Abstract). Refer to below 35 U.S.C 103 rational for the resetting limitation for further details.
Consequently, the 35 U.S.C. § 103 rejection is maintained.
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.
Non-Statutory Rejection
The claim 17 does/do not fall within at least one of the four categories of patent eligible subject matter. Claim 17 recites a "computer readable storage medium storing thereon a program." Under MPEP §2106 the claim's scope covers both tangible, non-transitory media (e.g., disk, flash drive) and transitory signals (e.g., electromagnetic or carrier waves), it fails to meet a statutory category and must be rejected under §101. The specification (see [0040]) expressly defines both non-transitory media (e.g., magnetic disk, CD-ROM, semiconductor memory) and transitory media (e.g., electromagnetic signals, wired or wireless channels). Thus, Claim 17 is not limited to a statutory category and is rejected under §101 unless amended to "non-transitory computer readable storage medium."
Subject Matter eligibility Rejection 35 U.S.C 101
Claims 1-17 are rejected under 35 U.S.C. § 101 because the claimed subject matter is directed to a judicial exception (an abstract idea) without reciting elements that integrate the exception into a practical application or provide an inventive concept amounting to significantly more than the exception itself.
Step 1: Statutory Categories Analysis
The claims are directed to statutory subject matter, encompassing the following statutory categories:
Machine (Claims 1–9): Directed to a device with memory and processor executing instructions, qualifying as a machine under MPEP §2106.
Process (Claims 10–16): These claims describe a method involving data acquisition and output processing, aligning with the process definition.
Non-Statutory
Claim 17: This claim includes transitory embodiments under BRI and thus fails to meet a statutory category per MPEP §2106. The specification (0040) confirms this.
Proceeding to Step 2A, Prong One: evaluation of abstract idea eligibility.
Step 2A, Prong One: Judicial Exception Analysis
Step 2A, Prong One it is to verify if a claim recites a specific judicial exception before determining if that exception is actually integrated into a practical application under prong two.
The whole invention is related to A device that identifies the minimum subset of patient data required to yield the same treatment recommendation as a full dataset and dynamically sets those data types as the priority for collection. Refer to par. 0058 for further details.
More specifically, the claims 1-17 are directed to a judicial exception because they recite the abstract idea of Mental Process. Because under broadest reasonable interpretation (BRI) the claims recite a feature selection optimization process to mathematically prove which patient data points are redundant and automatically reconfiguring the data collection requirements to exclude them to obtain same result that with the full data. Cognitive process that a doctor could do by years of experience attending different patient and continuously updated his procedure to reduces inputs data and obtain same results.
Independent Claims Recites the following non-bold parts:
Claim 10:
An information processing method comprising:
acquiring a model that is generated for each elapsed period, and has learned by machine learning to output a treatment for a human by receiving input of a plurality of types of feature value representing a condition of the patient;
collecting first output that is obtained when a predetermined number of the types of feature value are input to the model of each elapsed period, and second output that is obtained when some types of feature value in the predetermined number of the types of feature value are input to the model of each elapsed period;
and on a basis of the first output and the second output, setting types to be associated with the model of each elapsed period by:
generating, for each elapsed period, aggregated data including a pair (X', y') where X' identifies a varied subset of the types of feature value, and y' indicates whether a second output obtained with a subset is identical to the first output; training, for each elapsed period, a binary determination model using the aggregated data;
determining, across multiple patients, a required number of priority feature-value types based on outputs of the binary determination model;
resetting the priority feature-value types associated with a model of an earlier elapsed period by inserting, based on the required number and places in an order of priority associated with a model of a later elapsed period, one or more of the types of feature value from the later elapsed period at a position corresponding to the required number;
and outputting acquisition-instruction data to a user terminal to cause acquisition of the priority feature-value types for a subsequent elapsed period.
Note: The bolded portions represent additional elements evaluated in Prong Two and Step 2B. The non-bolded portions represent the abstract idea. Referenced applicant language comes from public application number.
Claim Abstract Classification Rational
Under their Broadest Reasonable Interpretation (MPEP § 2111), the independent claims 1, 10, and 17 abstract idea recite a feature selection optimization process to mathematically prove which patient data points are redundant and automatically reconfiguring the data collection requirements to exclude them to obtain same result that with the full data. This process aligns with the following abstract idea categories:
Mental Process (MPEP § 2106.04(a)(2)(III)): Under their Broadest Reasonable Interpretation (MPEP § 2111), the independent claims 1, 10, and 17 abstract idea recite a feature selection optimization process to mathematically prove which patient data points are redundant and automatically reconfiguring the data collection requirements to exclude them to obtain the same result as with the full data. This process aligns with the Mental Process category. The independent claims 1, 10, and 17 recite "acquiring a model... collecting first output... setting types... determining... a required number... [and] resetting the priority feature-value types." These limitations describe observations (collecting output), evaluations (comparing first and second outputs), and judgments (determining required numbers and resetting priorities) which are concepts performed in the human mind. The specification supports this by describing the invention as a tool for decision support: “Thereby, the information processing device 100 can be used for assistance of decision-making by a user, or the like” (Spec., Abstract).
Certain Method of Organizing Human Activity (MPEP § 2106.04(a)(2)(II)): Alternatively, the independent claims 1, 10, and 17 recite "setting types to be associated with the model... resetting the priority feature-value types... and outputting acquisition-instruction data." This describes a managed workflow of interaction regarding what data to collect, which falls under the sub-category of Managing Personal Behavior or Relationships or Interactions Between People. The specification supports this, stating: “Since collection of information about patients requires costs and time... it is necessary to enhance the efficiency of collection of information about subjects for measure proposals” (Spec., para. [0005]). The claims organize the activity of data collection to maximize efficiency.
Manual Replication Scenario (Human Equivalence)
The abstract nature of the claims is reinforced because the entire process is analogous to fundamental human activities:
A Chief Physician retrieves specific clinical protocols for Day 1 and Day 3. She reviews old charts to test if diagnoses made with partial data match those made with full data, creating a 'Match/No-Match' scorecard for each combination. Analyzing the scorecard patterns, she calculates that exactly two specific vitals are sufficient. She updates the Day 1 protocol to include a critical vital identified from Day 3 and hands the nurse a new checklist requiring only those two vitals.
The dependent claims 2-9 and 11-16 are also directed to an abstract idea.
Claims 2, 4, 11-16: These claims recite under BRI specific data content limitations and detailed algorithmic sub-steps, including "condition... includes bio-information," "collecting the second output," "setting... a required number," and "generating... a binary determination model" (Claims 2, 11, 13, 16). This subject matter falls within the Mental Process (MPEP § 2106.04(a)(2)(III)) categories. The claims merely elaborate on the mathematical calculations and logical determinations required to perform the abstract feature selection process. For instance, Claim 16 describes "generating... the binary determination model" and "determining the required number on a basis of output," which describes a mathematical relationship and calculation. As these claims describe mathematical calculations and mental evaluations of data without adding non-abstract hardware limitations, they remain directed to the abstract idea.
Because the claims are directed to an abstract idea, the analysis proceeds to Step 2A, Prong Two.
Step 2A, Prong Two: Integration into a Practical Application
Step 2A, Prong Two evaluates whether the claim as a whole integrates the recited judicial exception into a practical application. The claims' additional elements do not overcome Prong Two because they merely recite generic computer components and data gathering steps that do not meaningfully limit the abstract idea.
Evaluation of independent Claims 1, 10, and 17 Additional Elements
Generic Computer Components and Machine Learning: The recitation of "at least one memory," "at least one processor," "machine learning," and "user terminal" fails to integrate the abstract idea because it:
The additional elements of "at least one processor," "memory," "machine learning," and "user terminal" fail to integrate the abstract idea. As explained in MPEP § 2106.05(f), mere instructions to implement an abstract idea on a computer do not render a claim eligible. Here, the processor and memory are invoked merely as tools to perform the "acquiring," "collecting," and "setting" steps of the abstract idea. Furthermore, per MPEP § 2106.05(a), utilizing "machine learning" to "output a measure" is simply using a mathematical tool to achieve the abstract result, not an improvement to the technical functioning of the computer. Finally, the "user terminal" serves only as a data gathering/outputting endpoint (MPEP § 2106.05(h)). Consequently, these elements do not impose meaningful limits on the judicial exception.
When viewed as a whole, the combination of these elements does not integrate the abstract idea. The claim describes a generic arrangement of standard computing hardware performing the abstract analysis (feature value optimization) via standard software tools (machine learning), which does not transform the abstract idea into an eligible application.
Dependent claims does not add any additional elements merely narrowing the abstract idea identified in prong one.
Because the claims are directed to an abstract idea without integrating it into a practical application, the analysis proceeds to Step 2B.
Step 2B: Inventive Concept Analysis
Step 2B determines whether the claim recites additional elements that amount to significantly more than the judicial exception. The additional elements in Claims 1, 10, and 17, representing generic computer components and standard machine learning techniques, do not provide an inventive concept because they are well-understood, routine, conventional activities used in a generic manner.
Generic Computer Components and Machine Learning:
The additional elements fail to provide an inventive concept. Under MPEP § 2106.05(f), the "processor" and "memory" are recited at a high level of generality, which the specification confirms are "typical information processing device[s]" (Spec., para. [0060]). The use of "machine learning" is a mathematical tool used to "learn the relationship of the experience data" (Spec., para. [0048]), which does not improve the computer's functionality (MPEP § 2106.05(a)) but rather uses the computer's processing power for abstract calculations. The "user terminal" merely links the result to a user environment (MPEP § 2106.05(h)). These elements represent well-understood, routine, and conventional activities in the field of information processing, as evidenced by the specification's admission of using "typical" hardware and standard processing units like GPUs and TPUs (Spec., paras. 0060, 0070).
Dependent claims does not add any additional elements merely narrowing the abstract idea identified in prong one.
When viewed as a whole, the combination of additional elements does not amount to significantly more than the abstract idea. It simply amounts to "applying" the abstract idea of feature value optimization using "typical" computer hardware and standard machine learning techniques, which does not constitute an inventive concept.
Claim Rejections - 35 USC § 103
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 1-17 are rejected under 35 U.S.C. § 103 as being unpatentable over Yang (WO2018/228852) in view of Chang (US 2021/0374562 A1) referred to as Chang and in view of Chang, C.-H., Mai, M., & Goldenberg, A. (2019). Dynamic measurement scheduling for event forecasting using deep RL. Referred as Chang et al. (2019) Refer to PTO-892
Yang, teaches Claim 10:
An information processing method comprising: (Yang, [0010], [0032], [0063])
Yang explicitly discloses a "healthcare predictive analysis system" and a "healthcare analytics prediction system 101", which are types of "information processing device" as broadly interpreted. Yang further states that this system "can be implemented in one or more computing device, mobile device or other processing systems", directly aligning with the definition of an "information processing device" being an electronic device capable of processing data.
acquiring a model that is generated for each elapsed period, and has learned by machine learning to output a treatment for a human by receiving input of a plurality of types of feature value representing a condition of the patient; (Yang, [0031], [0010], [0049])
Yang teaches a model that is "trained using a sliding-window approach or algorithm" where "for each of the windows among the plurality of windows: a current generation predictive model is trained". This training utilizes "various machine learning or predictive analysis algorithms". The model "predicts the occurrence of adverse events" related to "patient conditions", taking as input "a current set of features" which "can include patient demographic information (e.g., age, gender, weight, height, ethnicity, residence, distance from hospital, etc.) and hospital information". The system's purpose is to "monitor patient health data and conditions, and predict the occurrence of adverse events", functionally equivalent to outputting a "treatment for a patient".
collecting first output that is obtained when a predetermined number of the types of feature value are input to the model of each elapsed period,
Yang's system retrieves a set of historical data and identifies a plurality of "windows," which are subsets of data corresponding to a "sub-period of time", analogous to an "elapsed period." For each "current window," the system extracts a "current set of features and outcomes" and uses this input to train a "current generation predictive model", which produces an output. This process is repeated sequentially for each window to dynamically update the model.
generating, for each elapsed period, aggregated data ; (Yang, Claim 1, ... extract a current set of features and outcomes corresponding to the current window... train a current generation predictive model based on the extracted current set of features and outcomes ...)
Yang describes the compilation of data that contains distinct components used for analysis, specifically input variables and results. Since Yang extracts "current features" (identifying types of feature values) and "current outcomes" (a pair structure of Input/Result) for the training of the predictive model, Yang describes data including identified feature values paired with associated data points.
training, for each elapsed period, a ; (Yang, par. 0010, 0021, 0051,
Yang describes utilizing the information extracted from the specific time windows as the input for the learning algorithm. Since Yang trains the model "based on the extracted current set of features and outcomes," Yang teaches using the data collected for that period to perform the training.
determining, across multiple patients, a
Yang describes a system that analyzes data derived from a large group of individuals associated with a healthcare entity. Since Yang processes "historical claim feed data" which is "made up of a large number of claims associated with... the healthcare provider entity's patients" to train the model, Yang teaches determining analytic values using a multi-patient dataset.
Yang describes identifying specific data elements that are deemed important for the analysis and assigning them a relative value of importance. Since Yang teaches identifying "predictive variables" (priority features) and "assigning weights" to them based on "impact on outcomes," Yang describes establishing priority among feature types.
; (Yang, paragraph 0050)
Yang describes associating specific variables with the model and assigning them importance based on their predictive power. Since Yang teaches identifying "predictive variables" (feature-value types) and "assigning weights" to them for the "current generation predictive model," Yang describes priority types associated with a model.
and outputting
Yang describes a network architecture where the analytics system communicates with devices operated by end-users. Since Yang discloses that the "healthcare analytics prediction system 101 is communicatively coupled to end-user systems" and that users can use these systems to "monitor patient conditions," Yang teaches transmitting data to a user device.
Obvious Rational:
Yang teaches a healthcare predictive analysis system that uses a "predictive model" ([0001]) trained on sequential "windows" (elapsed periods) of historical patient data to predict outcomes.
However, Yang does not disclose collecting a second output from a subset of features within the same elapsed period to compare against a first output, nor setting the feature types for the model based on that comparison to improve efficiency.
But Chang describes a "feature removal framework to streamline machine learning" (Title) designed to reduce "latency, processor usage, memory usage... and/or other types of resource overhead" ([0015, 0039]). The framework works by calculating "an importance score for each feature" which represents the "impact of a corresponding feature on the entity rankings" produced by the model (abstract, [0017], [0046]). It then identifies and removes the features with the "lowest importance scores" ([0019, 0046]) to create a "simplified version of the machine learning model using a second subset of the features that excludes the first subset" ([0096], [0047], Claim 1).
It would have been obvious to combine Yang with Chang because both address the challenge of processing large datasets with machine learning models to make predictions about specific populations ("healthcare predictive analysis," "produce entity rankings"). A person of ordinary skill in the art, facing the high cost and latency associated with processing the "historical claim feed data" (Yang, [0031]) in Yang's system, would be motivated to integrate Chang's framework. Chang explicitly addresses this motivation by providing a system to reduce "resource overhead" such as "latency, processor usage, memory usage" (Chang, [0016], [0039]). A person of ordinary skill in the art would be motivated to integrate Chang's feature removal framework because it provides a direct solution to the inefficiency of processing large feature sets. Chang identifies that "training and/or execution of machine learning models with large numbers of features... typically require more memory, computational resources, and time" ([0004]) and can lead to "difficulty scaling and/or meeting the latency requirement" ([0005]). Faced with these exact challenges in Yang's largescale data system, the artisan would have recognized that applying Chang's method for removing less impactful features offers the benefit of a more streamlined and scalable prediction system.
Yang teaches generating predictive models for each elapsed period, describing a process where "an (i)th generation model is trained using the extracted features and outcomes" for specific time windows (e.g., "January 1, 2012 to December 31, 2012") to output predictions (Yang, paras. [0044-0046], [0049]).
However, Yang fails to disclose generating aggregated data including a pair (X', y') where y' indicates whether a second output obtained with a subset is identical to the first output.
Chang teaches the Missing Element in bold, describing a feature analysis process that generates aggregated data including a pair (X', y') where X' identifies a varied subset of the types of feature value (modified feature values) and y' indicates whether a second output obtained with a subset is identical to the first output (rank-biased overlap). Specifically, Chang discloses obtaining "original rankings" (first output) and "modified rankings" (second output) produced from "modified feature values" (varied subset X') (Chang, paras. 0043-0046). Chang further teaches calculating a metric y' where "the measure of rank similarity includes a rank-biased overlap... the value of the rank-biased overlap falls in the range [0, 1], where... 1 indicates that S and T are identical" (Chang, para. 0048). Chang explicitly teaches aggregating this data, stating that "rank-biased overlaps between modified rankings associated with each feature... are aggregated into an importance score" (Chang, para. 0093).
It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention to combine the teachings of Yang with Chang because both references address the technical challenge of managing computational resources and efficiency in machine learning systems processing large datasets (Yang, paras. [0006], [0039]; Chang, paras. [0004]-[0006]). By integrating Chang's feature removal framework into Yang's sliding window system, a PHOSITA would be able to periodically analyze the "impact of a corresponding feature on the entity rankings" (Chang, para. [0017]) for each elapsed period in Yang, allowing the system to identify and remove low-impact features to streamline the training of the "(i)th generation model" (Yang, para. [0049]).
A person of ordinary skill in the art would have been motivated to integrate the aggregated data generation including the (X', y') pair from Chang into the system of Yang to achieve the benefit of reducing "latency, processor usage, memory usage... and/or other types of resource overhead," as Chang teaches that this process "reduces latency... during retrieval and calculation of features inputted into the machine learning models and subsequent execution of the machine learning models" (Chang, para. [0015]).
Reasonable Expectation of Success A PHOSITA would have had a reasonable expectation of success in combining the references because the modification required only ordinary skill and routine experimentation. Chang explicitly discloses that the feature removal framework is compatible with various models, including "tree-based models, deep learning models... and/or other types of machine learning models" (Chang, para. 0014).
Yang teaches training predictive models for each elapsed period, specifically describing that "an (i)th generation model is trained using the extracted features and outcomes" for specific time windows to predict adverse events (Yang, paras. [0046], [0049]).
However, Yang fails to disclose training a binary determination model using the aggregated data.
Chang teaches the Missing Element in bold, describing a process of using aggregated importance scores (aggregated data) to make a binary decision regarding feature selection. Specifically, Chang discloses a "simplification apparatus 202" that "uses importance scores 230 and a feature removal threshold 232 to identify a set of high-importance features 238" (Chang, para. [0051]). This identification process constitutes a binary determination model because it classifies features into two distinct categories: those to keep (high-importance) and those to remove (low-importance) based on whether they meet the threshold. Chang further teaches "trains a simplified version 214 of the machine learning model using only high-importance features 238" identified by this binary determination (Chang, para. [0054]).
It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention to combine the teachings of Yang with Chang because both references address the need to optimize machine learning models for processing large healthcare or professional datasets (Yang, paras. [0006], [0039]; Chang, paras. [0004]-[0006]). By integrating Chang's feature selection logic into Yang's periodic model updates, a PHOSITA would be able to apply a "binary determination model" (the threshold-based selection logic) to the "aggregated data" (importance scores derived from the current period's data) to systematically refine the feature set for each "elapsed period" (Yang's windows).
A person of ordinary skill in the art would have been motivated to integrate the binary determination model from Chang into the system of Yang to achieve the benefit of creating a "simplified version" of the model that "involves significantly less complexity and/or resource overhead than baseline version," as Chang teaches that this process allows the system to "process requests for scores or rankings... more quickly and/or handle greater volumes of these requests" (Chang, paras. [0021], [0070]).
A PHOSITA would have had a reasonable expectation of success in combining the references because the modification required only ordinary skill and routine experimentation. Chang explicitly confirms that the feature removal framework, including the binary determination logic, can be applied to "tree-based models, deep learning models... and/or other types of machine learning models" (Chang, para. [0014]), which are compatible with the predictive algorithms used in Yang (Yang, para. [0049]).
Yang teaches analyzing data across multiple patients, describing the use of "historical claim feed data is made up of a large number of claims associated with... the healthcare provider entity's patients" to identify predictive variables and assign weights (Yang, paras. [0039], [0050]). However, Yang fails to disclose explicitly determining a required number of priority feature-value types based on outputs of a binary determination model.
Chang teaches the Missing Element in bold, describing a process of calculating a specific integer threshold for feature selection. Specifically, Chang discloses that the "simplification apparatus 202 sets feature removal threshold 232 to an estimated number of features to be removed" or identifies "a predetermined number of features... that indicate the largest impact" (Chang, paras. 0051, 0053). This calculation of a specific number (e.g., "50 features") constitutes determining a required number of priority feature-value types (high-importance features) based on the outputs of the binary determination model (the thresholding logic that accepts/rejects features based on importance scores) (Chang, para. 0051).
It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention to combine the teachings of Yang with Chang because both references operate within the field of data analytics involving large user/patient populations (Yang, "patients"; Chang, "entities" such as users) (Yang, para. 0039; Chang, para. 0024). A PHOSITA utilizing Yang's system to monitor patient conditions would be motivated to integrate Chang's methodology for calculating a specific feature count threshold to manage the computational load of the predictive model.
A person of ordinary skill in the art would have been motivated to integrate the determination of a required number of priority features from Chang into the system of Yang to achieve the benefit of "meeting the latency requirement" and ensuring the model does not exceed a "target resource overhead," as Chang teaches calculating this number specifically to "lower the resource overhead to a target resource overhead" (Chang, paras. [0005], [0019]).
A PHOSITA would have had a reasonable expectation of success in combining the references because the modification required only ordinary skill and routine experimentation. Chang describes the calculation of the feature number as a mathematical operation based on resource metrics (e.g., "dividing the difference between baseline resource overhead... and target resource overhead") (Chang, para. [0053]), which is a standard engineering calculation compatible with the computational architecture of Yang.
Examiner interprets from above applicant language limitation by applying mpep 2111 rules, as a dynamic update process where the data acquisition policy (priority types) for a current or preceding time step is modified based on information derived from the requirements of a subsequent time step, because under the Broadest Reasonable Interpretation (BRI), "resetting" encompasses updating a policy or instruction set, and "based on... a later elapsed period" encompasses algorithmic feedback loops where future states influence previous state parameters (e.g., backpropagation or reinforcement learning updates).
Yang and Chang describe a time-series prediction system and a feature importance calculation that is read in the above applicant language interpreted limitation, because Yang teaches the structure of "elapsed periods" (sliding windows) and "models" generated for each period (Yang, [0046], [0053]), while Chang teaches calculating "importance scores" (order of priority) and determining a "required number" (feature removal threshold) to optimize the models (Chang '562, [0051], [0053]). (Reference, See at least Yang 0010, 0046; Chang 0046, 0051)
However does not describe the specific mechanism of resetting the priority of the earlier period by explicitly inserting features required by the later period (the "look-ahead" or retroactive update mechanism).
Chang et al. (2019) describe that is read in the above applicant language interpreted limitation a method of "Dynamic Measurement Scheduling" where the measurement policy (priority types) for an earlier time step (t) is learned and updated based on the predictive gain and requirements of a later time step (t+1), because Chang et al. discloses training a Deep Q-Network (RL agent) where the policy is updated to "maximize this model's performance" by minimizing the "Bellman-equation square error," which mathematically updates the value of an earlier action (what to measure at t) based on the "Q-target" (requirements) of the later state (t+1). (Reference, See at least Chang et al. (2019), Abstract ("scheduling strategically-timed measurements"), Section 3.2 ("Q-target"), Algorithm 3 ("Update... using L"))
The combination of Yang + Chang+ Chang et al. (2019) applications make obvious the full limitation because it applies the specific "measurement scheduling" logic of Chang et al. (2019) to the "time-series windows" of Yang, using the "importance scores" of Chang. Specifically, Chang et al. (2019) teaches that to minimize cost in a time-series (Yang), one must look at the future reward (Later Period) to decide the current policy (Earlier Period). (Reference, See at least Chang et al. (2019) Section 1 ("answer this question... by scheduling strategically-timed measurements"))
A skilled Artisan in the art who read Yang and Chang application, would combine Chang et al. (2019) with Yang and Chang, because all three references reside in the same field of healthcare informatics and machine learning, and explicitly address the same problem of handling "redundant and expensive screening procedures" and optimizing "measurement costs" in longitudinal patient data. (Reference, See at least Chang et al. (2019) Introduction; Yang [0006], Yang, 0003-0004, Chang, 0004, 0017)
A skilled Artisan would be motivated to combine Yang + Chang+ Chang et al. (2019) with expected predictive result, because applying the dynamic scheduling policy of Chang et al. (2019) to the predictive windows of Yang provides the predictable benefit of reducing the total number of measurements required (cost saving) while maintaining the data necessary for accurate future predictions (predictive gain), as explicitly taught by Chang et al. (2019). Chang et al. (2019) provides explicit evidence of this predictable result, stating: "In a real-world ICU mortality prediction task (MIMIC3), our policies reduce the total number of measurements by 31% or improve predictive gain by a factor of 3 as compared to physicians". (Reference, See at least Chang et al. (2019) Abstract ("reduce the total number of measurements by 31%... or improve predictive gain"))
Yang teaches a system communicatively coupled to "end-user systems" (user terminals) that allows users to "monitor patient conditions" and predicts events for future time windows (Yang, paras. [0036], [0046]). However, Yang fails to disclose explicitly outputting acquisition-instruction data to a user terminal to cause the acquisition of priority feature-value types for a subsequent elapsed period.
Chang teaches the Missing Element in bold, describing the deployment of a specific model configuration that dictates future data collection. Specifically, Chang discloses that the "management apparatus 206 replaces... execution of baseline version... with execution of simplified version 214" in the "production environment" (Chang, para. [0059]). This deployment constitutes outputting acquisition-instruction data (the simplified model configuration) because it instructs the system to acquire (described as retrieve... and/or calculate") only the priority feature-value types ("high-importance features") for subsequent elapsed periods ("real-time or near-real-time basis" execution) (Chang, paras. [0021], [0054]).
It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention to combine the teachings of Yang with Chang because both references operate within networked computing environments processing large datasets for users (Yang, para. [0036]; Chang, para. [0005]). A PHOSITA utilizing Yang's system would be motivated to integrate Chang's deployment mechanism to ensure that the user terminals or production systems only expend resources collecting data identified as critical by the model's training phase.
A person of ordinary skill in the art would have been motivated to integrate the outputting of acquisition-instruction data (the simplified model deployment) from Chang into the system of Yang to achieve the benefit of reducing "resource overhead... during retrieval and calculation of features inputted into the machine learning models," Chang, 0015, as Chang teaches that executing the simplified version with only priority features "reduces latency" and "memory usage" for future requests (Chang, para. 0039, 0015).
A PHOSITA would have had a reasonable expectation of success in combining the references because the modification required only ordinary skill and routine experimentation. Chang describes the deployment of the simplified version as a standard software management operation ("replaces... execution," "ramping up") (Chang, paras. [0059], [0086]), which is technically compatible with the networked architecture described in Yang.
Note: Claim 1 is also rejected under a similar analysis above it is substantially similar to claim 10.
Yang + Chang + Chang et al. (2019) teaches, Claim 2.
The information processing device according to claim 1, wherein the condition of the patient includes bio-information of the patient. (Yang, paragraph abstract, [0033]).
Yang teaches, a method for healthcare predictive analysis where the system collects "historical claim feed data regarding its patients" (Yang, Abstract), and that this data includes "basic medical data at the time of the relevant visit (e.g., weight, height, blood pressure)"
Yang + Chang+ Chang et al. (2019) teaches, Claim 4.
The information processing device according to claim 3, wherein the at least one processor is configured to execute the instructions to collect the second output that is obtained when each set of a varied number and/or combination of some types of feature value in the predetermined number of types of feature value is input to the model of each elapsed period.
Chang teaches varying the combination of features. The system generates "multiple sets of modified feature values" by replacing or swapping feature values and then "applies [the] baseline version to each set of modified feature values... to produce a corresponding set of modified rankings" (i.e., outputs) ([0044], [0046]). Chang teaches varying the number of features. The framework's core purpose is to identify a "number of features to be removed" ([0019], [0076]), identify the features with the "lowest importance scores" ([0019]), and train a "simplified version" of the model on the smaller, remaining feature set ([0018], [0047]).
Yang + Chang+ Chang et al. (2019) teaches, teaches Claim 11.
The information processing method according to claim 10, further comprising collecting the second output that is obtained when each varied set of some types of feature value in the predetermined number of types of feature value is input to the model of each elapsed period.
Chang describes detailed method for calculating feature importance by generating and testing multiple varied feature sets to determine their impact on a model's output. The process begins when the system obtains "multiple sets of feature values... and original rankings outputted by the machine learning model from the sets of feature values" ([0043], [0088-0089]). For a given feature, the system then generates a "modified set of feature values" by replacing its original value with a "default 'missing' value" or swapping it with a value from another record ([0044], [0090]). This modified set is then inputted into the baseline model to "produce a modified ranking" ([0046], [0091]). By systematically doing this for different features and across many data records, the system generates numerous "varied sets" of feature values and their corresponding outputs. The impact of each feature is then quantified by calculating a "rank-biased overlap between the modified ranking and the original ranking" ([0091], [0047]-[0048]) and aggregating these values into an "importance score for the feature" ([0050], [0093]).
Yang + Chang+ Chang et al. (2019) teaches, Claim 12.
The information processing method according to claim 11, further comprising:
collecting the aggregated data including an indication as to whether or not the first output that is obtained from a model of an elapsed period is identical to the second output that is obtained when each varied set of some types of feature value in the predetermined number of types of feature value is input to the same model corresponding to the elapsed period; Chang's "feature removal framework" (Title, [0001]) directly teaches this process. For the first part, Chang's method is based on comparing an "original ranking" (a first output) with a "modified ranking" (а second output) ([0046, 0043]). This comparison generates an "indication" of similarity. Chang explicitly teaches calculating a "measure of rank similarity between each modified ranking... and an original ranking", such as a "rank-biased overlap" ([0046-0050]). These rank-biased overlap values, calculated across many data records, constitute the claimed "aggregated data" that indicates the degree to which the outputs are identical.
and setting, on a basis of the aggregated data, types to be associated with the model of each elapsed period.
Chang teaches setting the feature types based on this aggregated data. Chang aggregates the rank biased overlaps into an "importance score for the feature" ([0050]). Based on these scores, the framework "identifies a subset of features with lowest important scores for removal" and trains a "simplified version of the machine learning model ... remaining features that exclude the ide