284 Generalizable Machine Learning Methods for Subtyping Individuals on National Health Databases: Case Studies Using Data from HRS, N3C, and All of Us

Abstract

OBJECTIVES/GOALS: While disease subtypes are critical for precision medicine, most projects use unipartite clustering methods such as k-means which are not fully automated, do not provide statistical significance, and are difficult to interpret. These gaps were addressed through bipartite networks and tested for generalizability on three national databases. METHODS/STUDY POPULATION: Data. All participants with self-reported stroke from the 2010 Health and Retirement Study (HRS), with cases (n=798) having one or more 8 depressive symptoms measured by the Centers for the Epidemiological Study–Depression 8 scale, and controls (n=389) with none of those symptoms. The replication data set consisted of independent identically-defined participants (cases=725, controls=190) from 1998 HRS. Method. (1) Bipartite network analysis and modularity maximization to automatically identify patient-symptom biclusters with significance. (2) Rand Index to measure the replicability of symptom co-occurrences in the replication data. (3) ExplodeLayout to visualize and interpret the subtypes. (4) R libraries to generalize the methods, upload them to CRAN, and then tested on the N3C and All of Us platforms. RESULTS/ANTICIPATED RESULTS: The analysis identified 4 depressive symptom subtypes (https://postimg.cc/Ny8YwXJW) which had significant modularity (Q=0.26, z=3.03, P DISCUSSION/SIGNIFICANCE: We developed generalizable methods to automatically identify biclusters, measure the clustering significance, and visualize the results for interpretation. These methods were successfully tested on three national level data bases. Such generalizable methods should accelerate the analysis of subtypes, and the design of targeted interventions.