Abstract 16360: Myocardial Injury and Rapid Unsupervised Topological Clustering of Electronic Health Record Features Predict Death and Length of Stay During a COVID-19 Hospitalization

Abstract

Background: Elevated troponin levels, a biomarker of myocardial injury (MI), has emerged as a predictor of acute COVID-19 outcomes. However, a lack of routine use in most patients makes it difficult to evaluate as a stand-alone predictor. We aimed to combine troponin levels and a new rapid, unbiased topological clustering pipeline to test for improvements in clinical risk stratification during COVID hospitalizations. Methods: SlicerDicer (SD) is an intuitive, anonymous clinical electronic health record (EHR) data exploration tool that helps characterize clinical outcomes in hospital cohorts. EHR data and outcomes were extracted with SD and clustered with Mapper Plus (MP), a clustering tool that identifies clinically relevant disjoint clusters. Initial clusters were further stratified by troponin level. Measured outcomes were in-hospital mortality rates (MR) and in-hospital length of stay (LOS). Results: 2692 patient records (2020-2021) with an overall MR of 6.3% and 97 clinical predictors were analyzed. MP identified 3 clusters with significant differences in MR (C1- 25.8%, C2- 4.2%, C3- 0.23%, P<0.001) and LOS in hours (C1- 290, C2- 132, C3- 8.3, P<0.001). Further stratification based on the troponin level reclassified the 3 MP clusters into 4 sub-clusters with distinctive MR (SC1- 32.6%, SC2- 17.3%, SC3- 2.3%, SC4-0.09%, P<0.001) based on MI presence. Conclusions: Using SD-based anonymous EHR data and unsupervised topological clustering with MP during a COVID hospitalization provides a low-effort, robust pipeline that predicts in-hospital MR and LOS. When MP clusters were re-stratified by the presence of MI, new sub-clusters (4) showed reclassification of both high- and low-risk patients. This risk classification schema provides a new pragmatic and computational framework for selecting hospitalized patients with COVID that can benefit from further in-hospital therapies and mechanistic studies on the role of myocardial injury in COVID-19 prognosis.