Comparative Evaluation of LSTM, GRU, and Bi-LSTM for 30-Day Heart Failure Readmission and Incident Atrial Fibrillation Prediction from Longitudinal Electronic Health Records

Recurrent neural networks have become the default choice for sequence modeling in longitudinal electronic health records, yet practitioners have limited empirical guidance on which RNN variant to select for cardiovascular early-warning tasks. This study reports a controlled comparison of long short-term memory (LSTM), gated recurrent unit (GRU), and bidirectional LSTM (Bi-LSTM) on two clinically meaningful endpoints: 30-day all-cause readmission among heart failure patients evaluated at discharge, and 6-month incident atrial fibrillation among at-risk adults evaluated from a fixed retrospective horizon anchored to the index hospitalization. The primary cohort is drawn from MIMIC-IV v3.1 (49,283 heart failure index admissions and 38,712 at-risk patients after exclusions), with eICU-CRD v2.0 used as a multi-center external validation cohort. Each variant is trained under a shared architectural search space, identical preprocessing pipelines, missing-value handling strategies, and class-rebalancing schemes. On the heart failure task, Bi-LSTM achieves the highest internal AUROC of 0.764 (95% CI: 0.752--0.776), exceeding LSTM by 1.6 percentage points and GRU by 1.3 percentage points, with consistent gains in DeLong testing. On the atrial fibrillation task, GRU attains an AUROC of 0.798 with roughly 25% fewer parameters and 24% shorter training time than LSTM. Performance gaps among the three variants narrow under external validation, with all models losing approximately 3-4 AUROC points (range: 3.4--4.2). Findings support task-conditioned variant selection rather than a universal recommendation, and we caution that the observed absolute AUROC differences, although statistically significant, are modest and require prospective decision-curve evaluation before any clinical deployment claim is justified.