Poster #23: Rasika Venkatesh PhD

Interpretable Temporal Modeling of Longitudinal EHR Trajectories Reveals Early Predictors of Incident Atrial Fibrillation

Rasika Venkatesh, PhD[1], Marylyn D. Ritchie, PhD[2], Dokyoon Kim, PhD[1]; [1]University of Pennsylvania, [2]Medical University of South Carolina

Atrial fibrillation (AF) is the most common arrhythmia and is associated with stroke and increased mortality. Identifying individuals at elevated risk before AF onset remains challenging, particularly when risk evolves dynamically over time. Longitudinal electronic health records (EHRs), combined with genetic and other multimodal data, provide an opportunity to characterize these trajectories and improve risk prediction. We conducted a retrospective study of 13,320 individuals with incident AF and 42,508 controls from the Penn Medicine BioBank, incorporating longitudinal EHR and genetic data. Clinical encounters, diagnoses, laboratory measurements, and vital signs were represented as sequential observations in 6-month windows over up to 10 years preceding AF diagnosis. We evaluated established and machine learning approaches, including C2HEST, logistic regression (LR), and XGBoost, alongside longitudinal deep learning (DL) models (LSTM and RETAIN) designed to capture temporal dependencies and provide attention-based interpretability. Model performance was assessed using AUROC, AUPRC, balanced accuracy, F1-score, and Brier score. Longitudinal DL models demonstrated strong predictive performance (AUROC=0.93-0.96; AUPRC=0.89-0.92; Balanced Accuracy=0.88-0.91; F1-score=0.81-0.84; Brier score=0.05-0.07), outperforming C2HEST (AUROC=0.77; AUPRC=0.48), LR (0.89; 0.73), and XGBoost (0.92; 0.84). Attention analyses identified hypertension and heart failure diagnoses, as well as temporal changes in BMI, HbA1c, and systolic blood pressure, particularly 12–18 months before AF diagnosis, as important predictive signals. Sensitivity analyses supported model robustness across temporal windows and cohort subsets. Interpretable temporal modeling of longitudinal EHR data captures evolving pre-diagnostic patterns associated with incident AF and improves prediction over conventional approaches. Ongoing work incorporates imaging and genomic features and evaluates model generalizability in All of Us and UK Biobank, with the goal of developing multimodal approaches that integrate clinical trajectories with genetic and physiologic information for individualized AF risk prediction.