Using Ensemble and Multi-Model Learning to Predict Longitudinal Hospitalization of Cardiovascular Patients and Warn of Pandemic Risks: A Retrospective Cohort Study with External Validations

Pandemic surveillance stands as a pivotal global public health priority. However, it remains unclear whether shifts in medical-seeking behavior during major public health events can serve as early warning signals for broader societal changes. We developed an ensemble method on cardiovascular patients’ electronic medical record (EHR), combined with socioeconomic factors to predict admission behavior of those patients and warn of shifts in medical demand during the early stage of pandemic. We extracted records of 40,000 patients of Wuhan Union Hospital, excluding minors and low-quality data. The internal discovery cohort (n = 37,071) and validation cohort (n = 39,915) was extracted by this strategy. The two cohorts span 2010 to 2024, covering both the pre-pandemic and post-pandemic periods. The analysis is informed by intra-dataset variability and aligned with known external events. In the Same-day Hospitalization Classification Task , the model achieved an average AUC of 0.85 on discovery cohort. In Regression of Time to Next Admission Task , the model yielded a R² of 0.81 and a MAE of 24.6 days. Through temporal evaluation and feature analysis we identified irregular patient flow patterns associated with medical reform, and detected anomalous traffic as early as May 2019. Post-pandemic data confirmed the association between actual admissions and COVID-19 infection, and our model tended to optimistically estimate admission. Long-term monitoring of cardiovascular patients reveals predictable patterns and clear temporal regularities. Ensemble learning approaches can capture population-level behavioral shifts and offer a promising strategy for epidemic surveillance and control.