Granted Patent B2

US 12,562,275 · App. 17/497,852 · Granted Feb 24, 2026

Inventors: Bum Chul Kwon (Cambridge, MA); Uri Kartoun (Cambridge, MA); Shaan Syed Khurshid (Cambridge, MA); Steven Alan Lubitz (Newton, MA); Kenney Ng (Arlington, MA)

Assignees: International Business Machines Corporation; The Broad Institute, Inc.

G16H50/20G06N3/00G06N20/00G16H40/67G16H50/30G16H50/70

Abstract

Claims (69)

1 . A computer-implemented method for assessing a risk to a subpopulation of a medical condition, the method comprising:

obtaining covariates and an outcome data for a population;

partitioning the population into a plurality of subgroups, wherein each subgroup is a different comorbidity that is predictive of the medical condition;

producing outcome predictions by applying a machine learning model to the covariate data for the population;

computing performance measures of the machine learning model for each subgroup based on the outcomes predictions and the outcome data;

comparing the performance measures for at least one subgroup to the performance measures for at least one other subgroup, including producing an interactive visualization of machine learning performance of the machine learning model for each of the subgroups, wherein the visualization is displayed by a user interface, wherein the performance measures include a concordance index that measures an accuracy of the machine learning model for predicting the risk of the medical condition and a fairness metric;

identifying an outlying subgroup for which the concordance index is different than the concordance index for one or more other subgroups;

receiving, through the interactive visualization, input indicative of a modified definition of the outlying subgroup;

tracking changes in the modified definition;

computing, in real-time, updated model performance results by applying the machine learning model to the outlying subgroup based on the modified definition; and

updating the visualization of the machine learning performance of the machine learning model, including the concordance index and the fairness metric, for each of the subgroups.

2 . The method of claim 1 further comprising:

retraining the machine learning model on additional covariate and outcomes data for the outlying subgroup.

3 . The method of claim 1 further comprising:

guiding a medical apparatus in treating a member of the outlying subgroup for the medical condition in response to the retrained machine learning model predicting an outcome for that member.

4 . The method of claim 1 further comprising:

identifying population traits that differ in the outlying subgroup compared to other subgroups; and

presenting, through the interactive visualization, to a user, the population traits which differ, to indicate bias in the outlying subgroup.

5 . The method of claim 1 , further comprising:

partitioning the population by applying a clustering algorithm to the population.

6 . The method of claim 1 , further comprising:

partitioning the population by applying domain expert knowledge to the population.

7 . The method of claim 1 further comprising comparing the outcome predictions for the at least one subgroup to the outcomes predictions for the entire population.

8 . The method of claim 1 wherein the performance measures include model performance metrics including at least one of prediction accuracy, concordance index, area under the curve, calibration, or standardized hazard ratios.

9 . The method of claim 1 wherein the performance measures include bias and fairness metrics including at least one of statistical parity difference, true positive rate difference, or true negative rate difference.

10 . A computer program product comprising one or more computer readable storage media that embody computer executable instructions, which when executed by a computer cause the computer to perform a method for assessing a risk to a subpopulation of a medical condition, the method comprising:

obtaining covariates and an outcome data for a population;

partitioning the population into a plurality of subgroups, wherein each subgroup is a different comorbidity that is predictive of the medical condition;

producing outcome predictions by applying a machine learning model to the covariate data for the population;

computing performance measures of the machine learning model for each subgroup based on the outcomes predictions and the outcome data;

comparing the performance measures for at least one subgroup to the performance measures for at least one other subgroup, including producing an interactive visualization of machine learning performance of the machine learning model for each of the subgroups, wherein the visualization is displayed by a user interface, wherein the performance measures include a concordance index that measures an accuracy of the machine learning model for predicting the risk of the medical condition and a fairness metric;

identifying an outlying subgroup for which the concordance index is different than the concordance index for one or more other subgroups;

receiving, through the interactive visualization, input indicative of a modified definition of the outlying subgroup;

tracking changes in the modified definition;

computing, in real-time, updated model performance results by applying the machine learning model to the outlying subgroup based on the modified definition; and

updating the visualization of the machine learning performance of the machine learning model, including the concordance index and the fairness metric, for each of the subgroups.

11 . The computer readable storage medium of claim 10 wherein the method further comprises:

retraining the machine learning model on additional covariate and outcomes data for the outlying subgroup.

12 . The computer readable storage medium of claim 10 wherein the method further comprises:

guiding a medical apparatus in treating a member of the outlying subgroup for the medical condition in response to the retrained machine learning model predicting an outcome for that member.

13 . The computer readable storage medium of claim 10 wherein the method further comprises:

identifying population traits that differ in the outlying subgroup compared to other subgroups; and

presenting, through the interactive visualization, to a user, the population traits which differ to indicate bias may exist in the outlying subgroup.

14 . The computer readable storage medium of claim 10 wherein the method further comprises:

partitioning the population by applying a clustering algorithm to the population.

15 . The computer readable storage medium of claim 10 wherein the method further comprises:

partitioning the population by applying domain expert knowledge to the population.

16 . The computer readable storage medium of claim 10 wherein the method further comprises:

comparing the outcome predictions for the at least one subgroup to the outcomes predictions for the entire population.

17 . An apparatus comprising:

a memory embodying computer executable instructions; and

at least one processor, coupled to the memory, and operative by the computer executable instructions to perform a method for assessing a risk to a subpopulation of a medical condition, the method comprising:

obtaining covariates and an outcome data for a population;

partitioning the population into a plurality of subgroups, wherein each subgroup is a different comorbidity that is predictive of the medical condition;

producing outcome predictions by applying a machine learning model to the covariate data for the population;

computing performance measures of the machine learning model for each subgroup based on the outcomes predictions and the outcome data;

comparing the performance measures for at least one subgroup to the performance measures for at least one other subgroup, including producing an interactive visualization of machine learning performance of the machine learning model for each of the subgroups, wherein the visualization is displayed by a user interface, wherein the performance measures include a concordance index that measures an accuracy of the machine learning model for predicting the risk of the medical condition and a fairness metric;

identifying an outlying subgroup for which the concordance index is different than the concordance index for one or more other subgroups;

receiving, through the interactive visualization, input indicative of a modified definition of the outlying subgroup;

tracking changes in the modified definition;

computing, in real-time, updated model performance results by applying the machine learning model to the outlying subgroup based on the modified definition; and

updating the visualization of the machine learning performance of the machine learning model, including the concordance index and the fairness metric, for each of the subgroups.

18 . The apparatus of claim 17 wherein the method performed by the at least one processor further comprises:

retraining the machine learning model on additional covariate and outcomes data for the outlying subgroup.

19 . The apparatus of claim 17 wherein the method performed by the at least one processor further comprises:

identifying population traits that differ in the outlying subgroup compared to other subgroups; and

presenting, through the interactive visualization, to a user, the population traits which differ to indicate bias may exist in the outlying subgroup.

20 . The apparatus of claim 17 wherein the method performed by the at least one processor further comprises:

partitioning the population by applying a clustering algorithm to the population.

Assignments (3)

Continuity (1)

References Cited (18)