Title: Leveraging Electronic Health Records to Enhance COVID-19 Prognosis through Interpretable and Data-driven Methods: A Focus on Cardiovascular Risk
Abstract
Background: The COVID-19 pandemic has exacerbated cardiovascular (CV) risk and disrupted the delivery of CV care globally. Patients with severe COVID-19 are at increased risk of developing major adverse cardiovascular events (MACE), and the risk factors are multifaceted. However, limited studies have evaluated the effects of pandemic throughout its course or developed clinical prediction tools to identify individuals at high risk. Therefore, the goal of this dissertation was to evaluate COVID-19-related longitudinal changes in CV health utilization, develop and compare predictive modeling approaches, and investigate updating strategies to mitigate performance drift over time. Methods: We conducted retrospective multicenter studies using electronic health record (EHR) data from the UMass Memorial Health system and National COVID Cohort Collaborative (N3C). In first study, we used a multi-phase interrupted time series with autoregressive integrated moving average (ITS-ARIMA) and generalized linear autoregressive moving average (GLARMA) models to examine weekly CV hospitalizations and outpatient visits across pre-pandemic, lockdown, post-restriction, and vaccination periods. The second and third studies developed and validated models predicting MACE or mortality in hospitalized COVID-19 patients, and compared logistic regression (LR), machine learning (ML), and Bayesian network (BN) approaches with stepwise data inputs. Model updating methods for BN models were evaluated. Results: Cardiovascular hospitalizations declined sharply during lockdown, rebounded post-restriction, and increased 13% in later period, while outpatient visits fell more steeply but later exceeded baseline by 20%. Recovery patterns differed across CV subtypes and subpopulations. Among all modeling techniques, LR, Bayesian additive regression tree (BART), and hill-climbing BN (HC-BN) showed good discrimination (AUCs: 0.73-0.79), calibration, and clinical usefulness. The top five key predictors across all models were age group, history of heart failure, disease severity, aspartate aminotransferase (AST) and troponin. In temporal validation, LR and HC-BN demonstrated relatively robust discrimination, and HC-BN was the only model that maintained good calibration compared with other methods. TAN exhibited improved discrimination after parameter updating, while other BNs showed minimal gains. Conclusion: The pandemic caused profound but heterogeneous disruptions in CV care, which underscores the need to target at-risk groups to ensure continuity care. BNs are powerful tools for transparent, data-driven prediction, and informed decision support.