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 .
The information disclosure statement (IDS) was submitted on June 10, 2024. The submission is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Specification
The disclosure is objected to because of the following informalities:
On page 12, lines 7-9, “However, if the epistemic uncertainty is high (large in 314), it means that the local model 312 does not have enough knowledge to make the prediction (enougi.e., the quantity of data is below a predetermined quantity threshold)” should read “However, if the epistemic uncertainty is high (large in 314), it means that the local model 312 does not have enough knowledge to make the prediction enough (i.e., the quantity of data is below a predetermined quantity threshold).”
Appropriate correction is required.
Claim Objections
Claims 9, 19 objected to because of the following informalities:
On claim 9, “The medical diagnosis system of claim 8, wherein the local models each include includes an encoder” should read “The medical diagnosis system of claim 8, wherein the local models each includes an encoder.”
On claim 19, “The method of claim 18, wherein the local models each include includes an encoder” should read “The method of claim 18, wherein the local models each includes an encoder.”
Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-20 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claim 1 recites “(ii) use of the local model for the medical diagnosis when the local model is confident…when predicting an ambiguous diagnosis with aleatoric uncertainty that is above the predetermined uncertainty threshold.” However, claim 2 recites “wherein when the aleatoric uncertainty is above the predetermined uncertainty threshold, both the local model and the global model abstain from prediction.” These limitations contradict one another, as claim 1 states the local model is used when the aleatoric uncertainty is above the predetermined uncertainty threshold, and claim 2 states the local model abstains from prediction, and is therefore not being used, when the aleatoric uncertainty is above the predetermined uncertainty threshold. Thus, the scope of claim 1 is rendered indefinite. For the purposes of examination, the Examiner has interpreted “(ii) use of the local model for the medical diagnosis when the local model is confident…when predicting an ambiguous diagnosis with aleatoric uncertainty that is above the predetermined uncertainty threshold” as the local model making a prediction when the aleatoric uncertainty is below the predetermined uncertainty threshold (Specification, page 11, lines 16-18, “if epistemic uncertainty is low (small in 316), local model 312 proceeds with the decision 324. Both models also consider aleatoric uncertainty 316, 334 and may abstain from prediction, 322, 342”).
Claim(s) 2-10 are rejected for at least the same reasons as claim 1 since they depend on claim 1.
Claim 11 recites “(ii) use of the local model for the medical diagnosis when the local model is confident…when predicting an ambiguous diagnosis with aleatoric uncertainty that is above the predetermined uncertainty threshold.” However, claim 12 recites “wherein when the aleatoric uncertainty determined by one local model is above the predetermined uncertainty threshold, both the one local model and the global model abstain from prediction.” These limitations contradict one another, as claim 11 states the local model is used when the aleatoric uncertainty is above the predetermined uncertainty threshold, and claim 12 states the local model abstains from prediction, and is therefore not being used, when the aleatoric uncertainty is above the predetermined uncertainty threshold. Thus, the scope of claim 11 is rendered indefinite. For the purposes of examination, the Examiner has interpreted “(ii) use of the local model for the medical diagnosis when the local model is confident…when predicting an ambiguous diagnosis with aleatoric uncertainty that is above the predetermined uncertainty threshold” as the local model making a prediction when the aleatoric uncertainty is below the predetermined uncertainty threshold (Specification, page 11, lines 16-18, “if epistemic uncertainty is low (small in 316), local model 312 proceeds with the decision 324. Both models also consider aleatoric uncertainty 316, 334 and may abstain from prediction, 322, 342”).
Claim(s) 12-20 are rejected for at least the same reasons as claim 11 since they depend on claim 11.
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.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claim(s) 1-2, 11-12 is/are rejected under 35 U.S.C. 103 as being anticipated by Kidd et al. (“Federated Learning for Sparse Bayesian Models with Applications to Electronic Health Records and Genomics”) in view of Narayan et al. (“Predicting on the Edge: Identifying Where a Larger Model Does Better”) in further view of Bhatt et al. (“Uncertainty as a Form of Transparency: Measuring, Communicating, and Using Uncertainty”).
Regarding claim 1, Kidd teaches
A medical diagnosis system in a network , comprising: a plurality of workstations for a plurality of respective medical facilities (Kidd, pages 484-485, “In this article, we study the sparse Bayesian models under the federated learning setting where data are distributed across multiple local sources (called local servers hereafter) and the goal is to perform global inference that pools information from local servers without breaching the local data privacy. Typical application includes privacy-preserved analyses of electronic health records data across multiple hospitals or medical centers where data may be limited in size in each site (hence independent analysis in each site would lack statistical power) but cannot be shared across sites due to the sensitivity of protected health information.” Kidd, page 491, “We analyze the COVID-19 clinical data electronically recorded in four Johns Hopkins’ hospitals (i.e., M = 4). Each hospital provides 100-150 patients, leading to a total sample size of 552. Due to the sensitive protected health information, data cannot be easily shared across hospitals for the purpose of statistical analyses but local computation within each hospital is feasible.”)
where each workstation performs medical diagnosis using medical data that is unique to the respective medical facility (Kidd, pages 484-485, “In this article, we study the sparse Bayesian models under the federated learning setting where data are distributed across multiple local sources (called local servers hereafter) and the goal is to perform global inference that pools information from local servers without breaching the local data privacy. Typical application includes privacy-preserved analyses of electronic health records data across multiple hospitals or medical centers where data may be limited in size in each site (hence independent analysis in each site would lack statistical power) but cannot be shared across sites due to the sensitivity of protected health information.”)
a central server, connected to communicate with the plurality of workstations, for maintaining a global model for the medical diagnosis (Kidd, page 485, “Essentially, the idea is to train models locally on each server using data that are only available to that server and then send model information (instead of any private data) to a central server for aggregation. The central server subsequently sends the aggregated model information back to local servers.”)
wherein each of the workstations maintains a local model for the medical diagnosis (Kidd, pages 484-485, “In this article, we study the sparse Bayesian models under the federated learning setting where data are distributed across multiple local sources (called local servers hereafter) and the goal is to perform global inference that pools information from local servers without breaching the local data privacy. Typical application includes privacy-preserved analyses of electronic health records data across multiple hospitals or medical centers where data may be limited in size in each site (hence independent analysis in each site would lack statistical power) but cannot be shared across sites due to the sensitivity of protected health information.”)
Kidd does not, but Narayan teaches
wherein each of the workstations includes a selector configured to switch between: (i) use of the global model for the medical diagnosis only if the local model has high epistemic uncertainty about the diagnosis at a given input point (Narayan, page 7, “Switcher model curves visualize how much performance we can gain by toggling between a small and large model based on uncertainty. A natural way to visualize and further evaluate the significance of our findings regarding model uncertainty and larger model improvement is with the use of switcher models. Switcher models, for the purpose of this paper, consist of a two-model ensemble. The smaller model always evaluates an example. If it is uncertain about an example, it passes the example to the larger model.” Examiner notes the local model is the smaller model and the global model is the larger model.)
(ii) use of the local model for the medical diagnosis when the local model is confident, either in predicting a particular medical diagnosis (Narayan, page 7, “Switcher models, for the purpose of this paper, consist of a two-model ensemble. The smaller model always evaluates an example. If it is uncertain about an example, it passes the example to the larger model.”)
Kidd and Narayan both disclose a federated learning environment and thus are analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd to incorporate a switcher model to switch between a local and global model based on uncertainty like Narayan. Doing so would have been advantageous because “in cases where evaluating the large model in every case is expensive or otherwise underesirable, a simple switcher model can capture all or nearly all of the performance while saving substantial compute resources. Even more interestingly, we see multiple cases where the hybrid switcher model does better than the big model!” (Narayan, page 7).
Kidd and Narayan do not, but Bhatt teaches
wherein the local model has high epistemic uncertainty above a predetermined uncertainty threshold that is based on a quantity of data about a particular input point that is less than a predetermined quantity (Bhatt, page 3, “While there can be many sources of uncertainty [190], herein we focus on those that we can quantify in ML models: aleatoric uncertainty (also known as indirect uncertainty) and epistemic uncertainty (also known as direct uncertainty) [41, 42, 59].” Bhatt, page 3, “Epistemic uncertainty can be reduced by collecting more data in input regions where the training dataset was sparse.” Bhatt, page 2, “In our cancer diagnostic scenario, we may want our automated system to abstain from making a prediction and instead request the assistance of a medical professional when its uncertainty is above a certain threshold.” Examiner notes the epistemic uncertainty is reduced by collecting more data about the input point and thus it is based on the quantity of data at that point. Examiner further notes the system abstains from prediction when the uncertainty is above a threshold, and therefore only predicts when the uncertainty is less than a predetermined quantity.)
(ii) use of the local model for the medical diagnosis when the local model is confident…when predicting an ambiguous diagnosis with aleatoric uncertainty that is above the predetermined uncertainty threshold (Bhatt, page 3, “While there can be many sources of uncertainty [190], herein we focus on those that we can quantify in ML models: aleatoric uncertainty (also known as indirect uncertainty) and epistemic uncertainty (also known as direct uncertainty) [41, 42, 59].”Bhatt, page 2, “In our cancer diagnostic scenario, we may want our automated system to abstain from making a prediction and instead request the assistance of a medical professional when its uncertainty is above a certain threshold.” Examiner notes that this limitation is being interpreted as the local model predicting only when the aleatoric uncertainty is below the predetermined uncertainty threshold. Examiner further notes the model only abstains from prediction when the uncertainty is above a threshold, and thus predicts when the uncertainty is below the threshold.)
Kidd and Narayan both disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt utilizes a machine learning model to perform medical diagnosis guided by quantifying uncertainty and thus is analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd and Narayan to utilize an uncertainty threshold based on the quantity of data about a particular input like Bhatt. Doing so would have been advantageous because “when well-calibrated and well-communicated, uncertainty can act as a useful form of transparency for stakeholders” (Bhatt, page 2). Furthermore, “Uncertainty, if not properly quantified and considered in model development, can endanger efforts to assess model fairness” and without uncertainty the model could “be under-specified and present larger error rates for younger patients (Bhatt, page 2). Additionally, “a doctor could make better use of an automated decision making system by observing its uncertainty estimates before lever aging the model’s output in making a diagnosis” (Bhatt, page 2).
Regarding claim 2, Kidd, Narayan, and Bhatt teach The medical diagnosis system of claim 1. Bhatt further teaches
wherein the local models of each of the workstations and the global model each determine an aleatoric uncertainty for the medical diagnosis, wherein when the aleatoric uncertainty is above the predetermined uncertainty threshold, both the local model and the global model abstain from prediction (Bhatt, page 3, “While there can be many sources of uncertainty [190], herein we focus on those that we can quantify in ML models: aleatoric uncertainty (also known as indirect uncertainty) and epistemic uncertainty (also known as direct uncertainty) [41, 42, 59].”Bhatt, page 2, “In our cancer diagnostic scenario, we may want our automated system to abstain from making a prediction and instead request the assistance of a medical professional when its uncertainty is above a certain threshold.”)
Kidd and Narayan both disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt utilizes a machine learning model to perform medical diagnosis guided by quantifying uncertainty and thus is analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd and Narayan to abstain from prediction if aleatoric uncertainty is above a predetermined threshold like Bhatt. Doing so would have been advantageous because “Uncertainty, if not properly quantified and considered in model development, can endanger efforts to assess model fairness” and without uncertainty the model could “be under-specified and present larger error rates for younger patients (Bhatt, page 2).
Regarding claim 11, Kidd teaches
A method of medical diagnosis in a network including a plurality of workstations for a plurality of respective medical facilities (Kidd, pages 484-485, “In this article, we study the sparse Bayesian models under the federated learning setting where data are distributed across multiple local sources (called local servers hereafter) and the goal is to perform global inference that pools information from local servers without breaching the local data privacy. Typical application includes privacy-preserved analyses of electronic health records data across multiple hospitals or medical centers where data may be limited in size in each site (hence independent analysis in each site would lack statistical power) but cannot be shared across sites due to the sensitivity of protected health information.” Kidd, page 491, “We analyze the COVID-19 clinical data electronically recorded in four Johns Hopkins’ hospitals (i.e., M = 4). Each hospital provides 100-150 patients, leading to a total sample size of 552. Due to the sensitive protected health information, data cannot be easily shared across hospitals for the purpose of statistical analyses but local computation within each hospital is feasible.”)
where each workstation maintains a local model for medical diagnosis (Kidd, pages 484-485, “In this article, we study the sparse Bayesian models under the federated learning setting where data are distributed across multiple local sources (called local servers hereafter) and the goal is to perform global inference that pools information from local servers without breaching the local data privacy. Typical application includes privacy-preserved analyses of electronic health records data across multiple hospitals or medical centers where data may be limited in size in each site (hence independent analysis in each site would lack statistical power) but cannot be shared across sites due to the sensitivity of protected health information.”)
and a central server, connected to communicate with the plurality of workstations, for maintaining a global model for the medical diagnosis (Kidd, page 485, “Essentially, the idea is to train models locally on each server using data that are only available to that server and then send model information (instead of any private data) to a central server for aggregation. The central server subsequently sends the aggregated model information back to local servers.”)
the method comprising: performing, in the plurality of workstations, the medical diagnosis using medical data that is unique to the respective medical facility (Kidd, pages 484-485, “In this article, we study the sparse Bayesian models under the federated learning setting where data are distributed across multiple local sources (called local servers hereafter) and the goal is to perform global inference that pools information from local servers without breaching the local data privacy. Typical application includes privacy-preserved analyses of electronic health records data across multiple hospitals or medical centers where data may be limited in size in each site (hence independent analysis in each site would lack statistical power) but cannot be shared across sites due to the sensitivity of protected health information.”)
Kidd does not, but Narayan teaches
switching, in each of the workstations, between (i) use of the global model for the medical diagnosis only if the local model has high epistemic uncertainty about the diagnosis at a given input point (Narayan, page 7, “Switcher model curves visualize how much performance we can gain by toggling between a small and large model based on uncertainty. A natural way to visualize and further evaluate the significance of our findings regarding model uncertainty and larger model improvement is with the use of switcher models. Switcher models, for the purpose of this paper, consist of a two-model ensemble. The smaller model always evaluates an example. If it is uncertain about an example, it passes the example to the larger model.” Examiner notes the local model is the smaller model and the global model is the larger model.)
(ii) use of the local model for the medical diagnosis when the local model is confident, either in predicting a particular medical diagnosis (Narayan, page 7, “Switcher models, for the purpose of this paper, consist of a two-model ensemble. The smaller model always evaluates an example. If it is uncertain about an example, it passes the example to the larger model.”)
Kidd and Narayan both disclose a federated learning environment and thus are analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd to incorporate a switcher model to switch between a local and global model based on uncertainty like Narayan. Doing so would have been advantageous because “in cases where evaluating the large model in every case is expensive or otherwise underesirable, a simple switcher model can capture all or nearly all of the performance while saving substantial compute resources. Even more interestingly, we see multiple cases where the hybrid switcher model does better than the big model!” (Narayan, page 7).
Kidd and Narayan do not, but Bhatt teaches
wherein the local model has high epistemic uncertainty above a predetermined uncertainty threshold that is based on a quantity of data about a particular input point that is less than a predetermined quantity (Bhatt, page 3, “While there can be many sources of uncertainty [190], herein we focus on those that we can quantify in ML models: aleatoric uncertainty (also known as indirect uncertainty) and epistemic uncertainty (also known as direct uncertainty) [41, 42, 59].” Bhatt, page 3, “Epistemic uncertainty can be reduced by collecting more data in input regions where the training dataset was sparse.” Bhatt, page 2, “In our cancer diagnostic scenario, we may want our automated system to abstain from making a prediction and instead request the assistance of a medical professional when its uncertainty is above a certain threshold.” Examiner notes the epistemic uncertainty is reduced by collecting more data about the input point and thus it is based on the quantity of data at that point. Examiner further notes the system abstains from prediction when the uncertainty is above a threshold, and therefore only predicts when the uncertainty is less than a predetermined quantity.)
(ii) use of the local model for the medical diagnosis when the local model is confident…when predicting an ambiguous diagnosis with aleatoric uncertainty that is above the predetermined uncertainty threshold (Bhatt, page 3, “While there can be many sources of uncertainty [190], herein we focus on those that we can quantify in ML models: aleatoric uncertainty (also known as indirect uncertainty) and epistemic uncertainty (also known as direct uncertainty) [41, 42, 59].”Bhatt, page 2, “In our cancer diagnostic scenario, we may want our automated system to abstain from making a prediction and instead request the assistance of a medical professional when its uncertainty is above a certain threshold.” Examiner notes that this limitation is being interpreted as the local model predicting only when the aleatoric uncertainty is below the predetermined uncertainty threshold. Examiner further notes the model only abstains from prediction when the uncertainty is above a threshold, and thus predicts when the uncertainty is below the threshold.)
Kidd and Narayan both disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt utilizes a machine learning model to perform medical diagnosis guided by quantifying uncertainty and thus is analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd and Narayan to utilize an epistemic uncertainty threshold based on the quantity of data about a particular input like Bhatt. Doing so would have been advantageous because “when well-calibrated and well-communicated, uncertainty can act as a useful form of transparency for stakeholders” (Bhatt, page 2). Furthermore, “Uncertainty, if not properly quantified and considered in model development, can endanger efforts to assess model fairness” and without uncertainty the model could “be under-specified and present larger error rates for younger patients (Bhatt, page 2). Additionally, “a doctor could make better use of an automated decision making system by observing its uncertainty estimates before lever aging the model’s output in making a diagnosis” (Bhatt, page 2).
Regarding claim 12, Kidd, Narayan, and Bhatt teach The method of claim 11. Bhatt further teaches
further comprising determining, by each of the local models and the global model, an aleatoric uncertainty for the medical diagnosis, wherein when the aleatoric uncertainty determined by one local model is above the predetermined uncertainty threshold, both the one local model and the global model abstain from prediction (Bhatt, page 3, “While there can be many sources of uncertainty [190], herein we focus on those that we can quantify in ML models: aleatoric uncertainty (also known as indirect uncertainty) and epistemic uncertainty (also known as direct uncertainty) [41, 42, 59].”Bhatt, page 2, “In our cancer diagnostic scenario, we may want our automated system to abstain from making a prediction and instead request the assistance of a medical professional when its uncertainty is above a certain threshold.”)
Kidd and Narayan both disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt utilizes a machine learning model to perform medical diagnosis guided by quantifying uncertainty and thus is analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd and Narayan to abstain from prediction if aleatoric uncertainty is above a predetermined threshold like Bhatt. Doing so would have been advantageous because “Uncertainty, if not properly quantified and considered in model development, can endanger efforts to assess model fairness” and without uncertainty the model could “be under-specified and present larger error rates for younger patients (Bhatt, page 2).
Claim(s) 3, 13 is/are rejected under 35 U.S.C. 103 as being anticipated by Kidd in view of Narayan in further view of Bhatt in further view of Dolezal et al. (“Uncertainty-informed deep learning models enable high-confidence predictions for digital histopathology”).
Regarding claim 3, Kidd, Narayan, and Bhatt teach The medical diagnosis system of claim 1. Kidd, Narayan, and Bhatt do not, but Dolezal teaches
wherein each of the workstations is configured to determine the predetermined uncertainty threshold based on a calibration dataset (Dolezal, page 2, “In this work, we describe a clinically-oriented method for determining slide-level confidence using the Bayesian approach to estimating uncertainty, with uncertainty thresholds determined from training data.” Dolezal, page 2, “UQ-enabled models were trained in cross validation for dataset sizes ≥ 100 slides, with uncertainty and prediction thresholds determined through nested cross-validation. AUROCs within the high-confidence UQ cohorts are greater than the non-UQ AUROCs for all dataset sizes ≥ 100 slides with α=0.05, reaching 0.981 ±0.004 at maximum dataset size.” Dolezal, page 11, “Thus, we use a nested cross-validation strategy in which optimal thresholds are determined from within training data only, and then applied to vali dation data (Fig. 1d). Within a given outer cross-fold training dataset, data is segmented into five nested cross-folds, and optimal θtile and θslide are determined for each of the five inner cross-fold validation datasets.” Examiner notes the calibration dataset is the training dataset, which is used to determine an uncertainty threshold.)
Kidd and Narayan both disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt and Dolezal utilize a machine learning model to perform medical diagnosis guided by quantifying uncertainty and thus are analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd, Narayan, and Bhatt to determine an uncertainty threshold based on a dataset like Dolezal. Doing so would have been advantageous because “This nested training strategy mitigates the data leak at the cost of requiring larger training dataset sizes” (Dolezal, page 11). Furthermore, Dolezal’s results “show that high-confidence predictions outperform predictions without uncertainty, in both cross-validation and testing on two large external datasets spanning multiple institutions” (Dolezal, page 1), wherein the uncertainty threshold is used to determine a high-confidence prediction.
Regarding claim 13, Kidd, Narayan, and Bhatt teach The method of claim 11. Kidd, Narayan, and Bhatt do not, but Dolezal teaches
further comprising determining, by each of the workstations, the predetermined uncertainty threshold based on a calibration dataset (Dolezal, page 2, “In this work, we describe a clinically-oriented method for determining slide-level confidence using the Bayesian approach to estimating uncertainty, with uncertainty thresholds determined from training data.” Dolezal, page 2, “UQ-enabled models were trained in cross validation for dataset sizes ≥ 100 slides, with uncertainty and prediction thresholds determined through nested cross-validation. AUROCs within the high-confidence UQ cohorts are greater than the non-UQ AUROCs for all dataset sizes ≥ 100 slides with α=0.05, reaching 0.981 ±0.004 at maximum dataset size.” Dolezal, page 11, “Thus, we use a nested cross-validation strategy in which optimal thresholds are determined from within training data only, and then applied to vali dation data (Fig. 1d). Within a given outer cross-fold training dataset, data is segmented into five nested cross-folds, and optimal θtile and θslide are determined for each of the five inner cross-fold validation datasets.” Examiner notes the calibration dataset is the training dataset, which is used to determine an uncertainty threshold.)
Kidd and Narayan both disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt and Dolezal utilize a machine learning model to perform medical diagnosis guided by quantifying uncertainty and thus are analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd, Narayan, and Bhatt to determine an uncertainty threshold based on a dataset like Dolezal. Doing so would have been advantageous because “This nested training strategy mitigates the data leak at the cost of requiring larger training dataset sizes” (Dolezal, page 11). Furthermore, Dolezal’s results “show that high-confidence predictions outperform predictions without uncertainty, in both cross-validation and testing on two large external datasets spanning multiple institutions” (Dolezal, page 1), wherein the uncertainty threshold is used to determine a high-confidence prediction.
Claim(s) 4, 14 is/are rejected under 35 U.S.C. 103 as being anticipated by Kidd in view of Narayan in further view of Bhatt in further view of Shuai et al. (“BalanceFL: Addressing Class Imbalance in Long-Tail Federated Learning”).
Regarding claim 4, Kidd, Narayan, and Bhatt teach The medical diagnosis system of claim 1. Kidd further teaches
wherein the local model is trained with local hospital patient data (Kidd, page 491, “We analyze the COVID-19 clinical data electronically recorded in four Johns Hopkins’ hospitals (i.e., M = 4). Each hospital provides 100-150 patients, leading to a total sample size of 552. Due to the sensitive protected health information, data cannot be easily shared across hospitals for the purpose of statistical analyses but local computation within each hospital is feasible.”)
Kidd, Narayan, and Bhatt do not, but Shuai teaches
wherein when the local hospital patient data does not fit a local data statistical distribution, the selector of a respective workstation switches to downloading trained global weights of the global model to the local model (Shuai, page 271, “In addition to global imbalance, another major challenge is local data imbalance, i.e., the sample number of every class on a node is highly uneven. An extreme case of the local data imbalance is the class missing, where some classes are not present at all while the node is still required to recognize them.” Shuai, page 275, “As analyzed in Section 4.2, class missing and class imbalance of activated classes are two major impediments to the self-balancing in step (2). First, to tackle the class missing issue, we let the local model inherit the knowledge of missing classes from the downloaded global model via a distillation loss, so as to preserve the responses on those classes to prevent the catastrophic forgetting.” Examiner notes the local data imbalance (missing classes, uneven classes) is the case in which the local data does not fit a local data statistical distribution. Examiner further notes the local model inheriting knowledge from the downloaded global model is the downloading of trained global weights of the global model to the local model.)
Kidd, Narayan, and Shuai disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt utilizes a machine learning model to perform medical diagnosis guided by quantifying uncertainty and thus are analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd, Narayan, and Bhatt to download global weights of the global model to local models when local data does not fit a statistical distribution like Shuai. Doing so would have been advantageous because “class missing and class imbalance of activated classes are two major impediments to the self-balancing” (Shuai, page 275), and “as federated learning proceeds, the server-side global model more or less has accumulated some knowledge on all classes, thus can be used as a teacher to remind the node-side local model of the knowledge on those absent classes” (Shuai, page 275).
Regarding claim 14, Kidd, Narayan, and Bhatt teach The method of claim 11. Kidd further teaches
wherein the local model is trained with local hospital patient data (Kidd, page 491, “We analyze the COVID-19 clinical data electronically recorded in four Johns Hopkins’ hospitals (i.e., M = 4). Each hospital provides 100-150 patients, leading to a total sample size of 552. Due to the sensitive protected health information, data cannot be easily shared across hospitals for the purpose of statistical analyses but local computation within each hospital is feasible.”)
Kidd, Narayan, and Bhatt do not, but Shuai teaches
the method further comprising when the local hospital patient data does not fit a local data statistical distribution, switching, by a respective workstation, to downloading trained global weights of the global model to the local model (Shuai, page 271, “In addition to global imbalance, another major challenge is local data imbalance, i.e., the sample number of every class on a node is highly uneven. An extreme case of the local data imbalance is the class missing, where some classes are not present at all while the node is still required to recognize them.” Shuai, page 275, “As analyzed in Section 4.2, class missing and class imbalance of activated classes are two major impediments to the self-balancing in step (2). First, to tackle the class missing issue, we let the local model inherit the knowledge of missing classes from the downloaded global model via a distillation loss, so as to preserve the responses on those classes to prevent the catastrophic forgetting.” Examiner notes the local data imbalance (missing classes, uneven classes) is the case in which the local data does not fit a local data statistical distribution. Examiner further notes the local model inheriting knowledge from the downloaded global model is the downloading of trained global weights of the global model to the local model.)
Kidd, Narayan, and Shuai disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt utilizes a machine learning model to perform medical diagnosis guided by quantifying uncertainty and thus are analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd, Narayan, and Bhatt to download global weights of the global model to local models when local data does not fit a statistical distribution like Shuai. Doing so would have been advantageous because “class missing and class imbalance of activated classes are two major impediments to the self-balancing” (Shuai, page 275), and “as federated learning proceeds, the server-side global model more or less has accumulated some knowledge on all classes, thus can be used as a teacher to remind the node-side local model of the knowledge on those absent classes” (Shuai, page 275).
Claim(s) 5-9, 15-19 is/are rejected under 35 U.S.C. 103 as being anticipated by Kidd in view of Narayan in further view of Bhatt in further view of Burnap et al. (“Design and Evaluation of Product Aesthetics: A Human-Machine Hybrid Approach”).
Regarding claim 5, Kidd, Narayan, and Bhatt teach The medical diagnosis system of claim 1. Kidd, Narayan, and Bhatt do not, but Burnap teaches
The medical diagnosis system of claim 1, wherein the local models each are a neural network trained as a Dirichlet model (Burnap, page 10, “Our encoder imputes the missing values by estimating a multinomial classifier (assumed Dirichlet distribution, see Appendix A2).”)
including a normalization flow and a decoder mapping extracted features to a vector of class probabilities (Burnap, page 14, “We use spectral normalization as a regularization technique to control the magnitude of gradients during model training (see §5.3).” Burnap, page 9, “We use a deep neural network, 𝑓 ℎ⃗ ,𝛽⃗ , to map embeddings to the rating of interest, say a 1-to-5 rating on “appeal.”
wherein an encoder performs the feature extraction function that maps an input to a lower-dimensional embedding (Burnap, page 6, “We do both by training an encoding model to embed images in a lower-dimensional space. Embeddings reduce the dimensionality, and allow us to combine the relatively thin (and expensive for the firm) labelled training data (images with aesthetic ratings) with a much larger sample of 180,000 unlabeled training data (images without ratings).”)
Kidd and Narayan disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt utilizes a machine learning model to perform medical diagnosis guided by quantifying uncertainty and thus is also analogous to the claimed invention. Burnap utilizes a Dirichlet and encoding model to model distributions and thus is also analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd, Narayan, and Bhatt to utilize a Dirichlet neural network model, decoder, encoder, and a normalization flow like Burnap. Doing so would have been advantageous because Burnap’s model achieved a “43.9% improvement relative to a baseline and substantial improvement over both conventional machine learning models and pretrained deep learning models” (Burnap, page 2).
Regarding claim 6, Kidd, Narayan, Bhatt, and Burnap teach The medical diagnosis system of claim 5. Burnap further teaches
wherein each of the workstations comprises a training layer that performs training of the local models using a training loss function that simultaneously maximizes likelihood of training embeddings (Burnap, page 10, “We use a deep neural network, 𝑓 𝑋,𝑎⃗ ,𝛽⃗ , to map images, 𝑋, and product attributes, 𝑎⃗ , to an embedding distribution, 𝑞 (⋅ |𝑋,𝑎⃗).” Burnap, page 10, “The encoder loss function includes two terms. The first loss term rewards the encoder for estimating a 𝑞(⋅ |𝑋,𝑎⃗) that is close to the prior.” Examiner notes the loss function includes a loss term that maximizes closeness of 𝑞 (⋅ |𝑋,𝑎⃗), which is the embedding distribution.)
and prevents an impact of a uncertain cross entropy loss on density estimation parameters of the Dirichlet model (Burnap, page 16, “This is especially true for the models in Figure 4, which feature complex deep learning models that balance multiple competing and/or adversarial loss-function terms. We must avoid unbounded “back-propagated” gradients that lead to catastrophic failures in model training (e.g., outputting images of only black pixels). We stabilize training with the spectral normalization layer in each block. This promotes Lipschitz continuity to regularize the model by dividing the raw output weights of each layer by the largest singular value of the matrix of weights. See Miyato et al. 2018. We manually enforce both soft and hard constraints on portions of the model architecture. For example, we bound the variance of the Gaussian random variables in the KL-divergence terms and use floating point safeguards to ameliorate the potential numerical instability introduced in §4 by logarithms and various 𝐿-norms.” Burnap, page 32, “The encoder neural net for 𝜇⃗ and 𝜎⃗ thus also produces 𝜋⃗ , from which we draw Dirichlet probabilities, 𝑎= 𝑞 𝜋⃗|𝑋,𝛽⃗ , using a soft-max function. We recognize 𝐸⃗ log𝑞 𝜋⃗|𝑋,𝛽⃗ as the cross-entropy for a draw of the attributes, 𝑎⃗ , from the multinomial probabilities, 𝑎. This provides the second term in the loss function below. During training, this term encourages the encoder to learn attributes, while during prediction (when we do not know attributes) this term allows us to estimate unknown product attributes, 𝑎, by sampling from the multinomial distribution indexed by 𝜋⃗ .” Examiner notes the second term in the loss function is a cross-entropy loss term, and the impact of the loss function are the unbounded backpropagated training gradients which is prevented during model training.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Kidd, Narayan, Bhatt with the teachings of Burnap for the same reasons disclosed in claim 5.
Regarding claim 7, Kidd, Narayan, Bhatt, and Burnap teach The medical diagnosis system of claim 6. Burnap further teaches
wherein the prevention by the training loss function is preventing propagation of a training gradient to the density estimation parameters of a parametric model to estimate density (Burnap, page 16, “This is especially true for the models in Figure 4, which feature complex deep learning models that balance multiple competing and/or adversarial loss-function terms. We must avoid unbounded “back-propagated” gradients that lead to catastrophic failures in model training (e.g., outputting images of only black pixels). We stabilize training with the spectral normalization layer in each block. This promotes Lipschitz continuity to regularize the model by dividing the raw output weights of each layer by the largest singular value of the matrix of weights. See Miyato et al. 2018. We manually enforce both soft and hard constraints on portions of the model architecture. For example, we bound the variance of the Gaussian random variables in the KL-divergence terms and use floating point safeguards to ameliorate the potential numerical instability introduced in §4 by logarithms and various 𝐿-norms.”)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Kidd, Narayan, Bhatt with the teachings of Burnap for the same reasons disclosed in claim 5.
Regarding claim 8, Kidd, Narayan, and Bhatt teach The medical diagnosis system of claim 1. Kidd further teaches
wherein the global model is trained as an average over parameters of the local models (Kidd, page 486, “For example, one may aggregate the estimates of θ1,...,θM using some deterministic function θ = f(θ1,...,θM) such as average for continuous parameters and majority vote for discrete parameters. Such deterministic approach is often ad hoc (e.g., lack of finite-sample theoretical justification) and generally does not propagate estimation uncertainty from local parameters θ1,...,θM to the global parameter θ.” Examiner notes the global model is trained using global parameter θ, which is obtained by aggregating the estimates of the local parameters θ1,...,θM.)
Kidd, Narayan, and Bhatt do not, but Burnap teaches
wherein the global model includes an encoder, having encoder parameters, a density model, having density parameters, and a classifier model with classifier parameters (Burnap, page 3, “Specifically, we use an encoding model to represent visual designs (images) using a smaller number of 512 abstract “features,” and train a predictive model that predicts aesthetic ratings based on those features.” Burnap, page 8, “We denote the parameters of the predictive model by 𝛽⃗P , the generative model by 𝛽⃗G , the encoding model by 𝛽⃗E , and the combined vector of parameters by 𝛽⃗ = 𝛽⃗P ,𝛽⃗G ,𝛽⃗E .” Burnap, page 10, “In practical applications, product attributes 𝑎⃗ often contain missing values. Our encoder imputes the missing values by estimating a multinomial classifier (assumed Dirichlet distribution, see Appendix A2).” Examiner notes the encoding model (encoder) estimates a Dirichlet distribution which is a continuous probability density function and thus is considered a density model. Examiner further notes the predictive model is the classifier model. Examiner further notes the encoding model parameters are denoted by 𝛽⃗E, and the predictive model parameters are denoted by 𝛽⃗P.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Kidd, Narayan, Bhatt with the teachings of Burnap for the same reasons disclosed in claim 5.
Regarding claim 9, Kidd, Narayan, Bhatt, and Burnap teach The medical diagnosis system of claim 8. Kidd further teaches
wherein, after federated learning, the local models are trained...using local medical data (Kidd, page 485, “Essentially, the idea is to train models locally on each server using data that are only available to that server and then send model information (instead of any private data) to a central server for aggregation.” Kidd, pages 484-485, “In this article, we study the sparse Bayesian models under the federated learning setting where data are distributed across multiple local sources (called local servers hereafter) and the goal is to perform global inference that pools information from local servers without breaching the local data privacy. Typical application includes privacy-preserved analyses of electronic health records data across multiple hospitals or medical centers where data may be limited in size in each site (hence independent analysis in each site would lack statistical power) but cannot be shared across sites due to the sensitivity of protected health information.”)
while keeping the encoder parameters fixed with global encoder parameter values (Kidd, page 488,
PNG
media_image1.png
271
671
media_image1.png
Greyscale
Examiner notes the encoder and global encoder are the local and global models, respectively, and the global model sends global parameter values to the local model which uses the global parameter values to perform a local update and update the local model parameters.)
Burnap further teaches
wherein the local models each include includes an encoder, having encoder parameters, a density model, having density parameters, and a classifier model with classifier parameters (Burnap, page 3, “Specifically, we use an encoding model to represent visual designs (images) using a smaller number of 512 abstract “features,” and train a predictive model that predicts aesthetic ratings based on those features.” Burnap, page 8, “We denote the parameters of the predictive model by 𝛽⃗P , the generative model by 𝛽⃗G , the encoding model by 𝛽⃗E , and the combined vector of parameters by 𝛽⃗ = 𝛽⃗P ,𝛽⃗G ,𝛽⃗E .” Burnap, page 10, “In practical applications, product attributes 𝑎⃗ often contain missing values. Our encoder imputes the missing values by estimating a multinomial classifier (assumed Dirichlet distribution, see Appendix A2).” Examiner notes the encoding model (encoder) estimates a Dirichlet distribution which is a continuous probability density function and thus is considered a density model. Examiner further notes the predictive model is the classifier model. Examiner further notes the encoding model parameters are denoted by 𝛽⃗E, and the predictive model parameters are denoted by 𝛽⃗P.)
wherein...models are trained through training the density model and the classifier model (Burnap, page 8, “We denote the parameters of the predictive model by 𝛽⃗P , the generative model by 𝛽⃗G , the encoding model by 𝛽⃗E, and the combined vector of parameters by 𝛽⃗ = 𝛽⃗P ,𝛽⃗G ,𝛽⃗E . To train the model, we minimize the combined loss function: (1) ℒ𝛽⃗ = ℒ 𝛽⃗P + ℒ 𝛽⃗G + ℒ 𝛽⃗E where ℒ 𝛽⃗P, ℒ 𝛽⃗G, and ℒ 𝛽⃗E indicate the predictive, generative, and encoding loss functions, respectively.” Examiner notes the model is trained by training the density (encoder) model and classifier (predictive) model.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Kidd, Narayan, Bhatt with the teachings of Burnap for the same reasons disclosed in claim 5.
Claim(s) 5 recites substantially similar limitations to claim(s) 15, and is/are therefore rejected under the same analysis.
Claim(s) 6 recites substantially similar limitations to claim(s) 16, and is/are therefore rejected under the same analysis.
Claim(s) 7 recites substantially similar limitations to claim(s) 17, and is/are therefore rejected under the same analysis.
Claim(s) 8 recites substantially similar limitations to claim(s) 18, and is/are therefore rejected under the same analysis.
Claim(s) 9 recites substantially similar limitations to claim(s) 19, and is/are therefore rejected under the same analysis.
Claim(s) 10, 20 is/are rejected under 35 U.S.C. 103 as being anticipated by Kidd in view of Narayan in further view of Bhatt in further view of Muthusamy et al. (“Comparison of different configurations of quantile regression in estimating predictive hydrological uncertainty”).
Regarding claim 10, Kidd, Narayan, and Bhatt teach The medical diagnosis system of claim 1. Kidd, Narayan, and Bhatt do not, but Muthusamy teaches
wherein the predetermined uncertainty threshold is quantile [0.8, 0.9] (Muthusamy, page 514, “Vsarious approaches to estimate hydrological uncertainty have been presented in the literature. One of the approaches is to use statistical post-processing to estimate predictive hydrological uncertainty, and several hydrological post processors have been explored in recent studies including Hydrological Uncertainty Processor (HUP) [3], Bayesian Model Averaging (BMA) [4], Model Conditional Processor (MCP) [2] and Quantile Regression (QR) [5,6]. Among these methods, this paper focuses on the application of QR. QR provides a simple, efficient and robust means for estimation of predictive uncertainty [5].” Muthusamy, page 517, “In order to evaluate the performance at different regimes of the flow forecasts, a series of thresholds values are defined based on the quantiles of the forecast data (0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9 and 0.95) and BS is calculated for each threshold value.”)
Kidd and Narayan disclose a federated learning environment and thus are analogous to the claimed invention. Bhatt and Muthusamy both utilize a machine learning model and uncertainty to quantify real-world data and thus are also analogous to the claimed invention. It would have been obvious to one having ordinary skill in the art prior to the effective filing date of the application to have modified Kidd, Narayan, and Bhatt to set the uncertainty threshold at quantile [0.8, 0.9] like Muthusamy. Doing so would have been advantageous because “the performance of the other stations is generally improving with increasing threshold. Actual warning level thresholds are all above the 0.9 quantile where all the stations perform better” (Muthusamy, page 517).
Claim(s) 20 recites substantially similar limitations to claim(s) 10, and is/are therefore rejected under the same analysis.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Zared O. Cohen whose telephone number is (571)270-0531. The examiner can normally be reached M-Th, 8am to 5pm ET.
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, Michelle Bechtold can be reached at (571) 431-0762. 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.
/Z.O.C./Examiner, Art Unit 2148 /MICHELLE T BECHTOLD/Supervisory Patent Examiner, Art Unit 2148
