Interpretation of machine learning classifications in clinical diagnostics using shapley values and uses thereof
Shapley values (SVs) have become an important tool to further the goal of explainability of machine learning (ML) models. However, the computational load of exact SV calculations increases exponentially with the number of attributes. Hence, the calculation of SVs for models incorporating large numbers of interpretable attributes is problematic. Molecular diagnostic tests typically seek to leverage information from hundreds or thousands of attributes, often using training sets with fewer instances. Methods are described for evaluate SVs using Monte Carlo sampling or exact calculation in polynomial time (i.e., reasonably quickly and efficiently) using the architecture of a ML model designed for robust molecular test generation, and without requiring classifier retraining.
1 . A method executing within a programmed computer to predict a patient outcome with a trained machine learning classifier, comprising the steps of:
(a) training the trained machine learning classifier arranged as a logistical combination of atomic classifiers with drop-out regularization, wherein the drop-out regularization comprises randomly selecting a small fraction of the atomic classifiers as a result of carrying out a dropout from the set of atomic classifiers;
(b) storing a set of values for attributes associated with a patient to the programmed computer to access an electronic health record of the patient or from one or more measurements obtained from the patient or sample obtained therefrom, or both;
(c) executing the trained machine learning model to classify the set of values with the trained machine learning classifier and generate the prediction using the programmed computer to classify the set of values with the trained machine learning classifier;
(d) calculating a relative contribution of some or all of the attributes to the prediction to generate an explanation of the prediction using the programmed computer; wherein calculating the relative contribution comprises either calculating exactly or estimating Shapley values for the attributes and wherein calculating the relative contribution comprises selecting subsets of the attributes to calculate the Shapley values by drawing a number of attributes associated with the patient for a subset and then randomly picking the attributes used within the subset, for dropout iterations; and
(e) displaying a report using the programmed computer, wherein the report comprises (1) the prediction of risk of future adverse event for the patient while hospitalized, (2) data representing the calculation of the relative contribution of some or all of the attributes from step (d), either in text or graphical format, and wherein the report is for planning or adjusting a treatment for the patient.
2 . The method of claim 1 , wherein step (b) comprises performing a physical measurement on the sample obtained from the patient, and wherein the report comprises one or more of a first comment on the prediction and a second comment on the data.
3 . The method of claim 1 , wherein step (b) comprises performing a physical measurement comprising mass spectrometry.
4 . The method of claim 2 , wherein the physical measurement comprises a genomic or proteomic assay.
5 . The method of claim 1 , wherein the patient comprises a hospitalized patient and wherein the attributes comprise clinical and demographic data and findings obtained at admission to a hospital and wherein the prediction comprises a prediction of risk of future adverse event for the patient while hospitalized.
6 . The method of claim 1 , wherein the trained machine learning classifier comprises a hierarchical arrangement of a binary classifier and one or more child classifiers, wherein the method further comprises the step of calculating the Shapley values for one or more of the attributes for predictions generated by both the binary classifier and the one or more child classifiers, and wherein the atomic classifiers of the trained machine learning classifier are k-nearest neighbor classifiers.
7 . The method of claim 1 ,
wherein the Shapley values for the one or more of the attributes are calculated by a Monte Carlo sampling method, and
wherein step (e) comprises presenting the Shapley values using one or more radar plots around a comment corresponding to the prediction and one or more of the Shapley values.
8 . The method of claim 1 , wherein the Shapley values for the one or more of the attributes are calculated as averages for multiple drop-out iterations.
9 . The method of claim 6 , wherein calculating the relative contribution of step (d) is performed in accordance with one of:
using subset sampling using least squares,
evaluating pairwise differences of the Shapley values, and
decomposing the Shapley values for drop-out iterations.
10 . The method of claim 5 , further comprising the step of using the report generated in step (e) to plan or adjust a treatment for the patient.
11 . The method of claim 1 , further comprising the step of using the report generated in step (e) to plan or adjust a treatment for the patient.
12 . The method of claim 8 , wherein calculating the relative contribution is performed by calculating one of the Shapley values as an average for multiple drop-out iterations.
13 . The method of claim 7 , wherein the Shapley values are calculated as a solution of a linear system.