364 Identification of Symptom-Based Phenotypes in PASC Patients through Bipartite Network Analysis: Implications for Patient Triage and Precision Treatment Strategies

Abstract

OBJECTIVES/GOALS: Approximately 10% of COVID-19 patients experience multiple symptoms weeks and months after the acute phase of infection. Our goal was to use advanced machine learning methods to identify PASC phenotypes based on their symptom profiles, and their association with critical adverse outcomes, with the goal of designing future targeted interventions. METHODS/STUDY POPULATION: Data. All COVID-19 outpatients from 12 University of Minnesota hospitals and 60 clinics. Independent variables consisted of 20 CDC-defined PASC symptoms extracted from clinical notes using NLP. Covariates included demographics, and outcomes included New Psychological Diagnostic Evaluation, and Number of PASC Hospital Visits (>=5). Cases (n=3235) consisted of patients with at least one symptom, and controls (n=3034) consisted of patients with no symptoms. Method. (1) Used bipartite network analysis and modularity maximization to identify patient-symptom biclusters. (2) Used multivariable logistic regression (adjusted for demographics and corrected through Bonferroni) to measure the odds ratio of each patient bicluster to adverse outcomes, compared to controls, and to each of the other biclusters. RESULTS/ANTICIPATED RESULTS: The analysis identified 6 PASC phenotypes ( http://www.skbhavnani.com/DIVA/Images/Fig-1-PASC-Network.jpg ), which was statistically significant compared to 1000 random permutations of the data (PASC=.31, Random Median=.27, z=11, P