Abstract 15358: Comparative Performance of Recurrent Heart Failure Prediction Models: Incorporating Social Determinants and Applying Machine Learning

Abstract

Introduction: Improving performance of predictive algorithms for 30-d readmission following heart failure (HF) hospitalization may enhance clinical management decisions. Applying novel data (social determinants of health; SDOH) and novel methods (machine learning) have been proposed. We compared performance to predict 30-d HF readmission using models with and without SDOH and under traditional and machine learning techniques. Hypothesis and Methods: We used data on 27,071 admissions within Emory Healthcare (mean age: 66 , 48% female, 51% black). We categorized 69 covariates as patient-level clinical (e.g., diastolic blood pressure), patient-level demographic (e.g. gender, age), and zip code SDOH (e.g. percent unemployed). Logistic regression under generalized linear model (GLM) and gradient boosted model (GBM) frameworks were applied to test all combinations of covariate sets. We evaluated performance using 5-fold cross-validation calculating area under the curve (AUC; discrimination) and mean absolute error (MAE; calibration) statistics. Results: 30-d readmission occurred in 17.7% of hospital admissions. Under GLM, model performance was similar across all covariate sets (AUCs .500 - .502). Under GBM, model performance measures were higher than GLM but similar across covariate sets (.500-.664). Notably, SDOH improved model discrimination (AUC .606 clinical only, AUC .664 clinical+SDOH) in the full dataset (“in-sample”) runs. Conclusions: Patient clinical data provided the richest information for prediction of HF readmission, with incremental gains through the addition of area SDOH data such as zip code level unemployment. Furthermore, these gains were only realized in GBM highlighting the marginal utility in flexible modeling methods. While non-clinical data sources can provide a rich understanding of the social mechanisms producing disease, their utility for clinical prediction may be outcome-specific.