Patients visits forecasting in emergency departments: Self-adapting LSTM models for evolving data distributions.

BACKGROUND AND

OBJECTIVE

Overcrowding in emergency departments (EDs) remains a critical challenge in healthcare systems. Accurate forecasting of patient visits is essential to optimize the allocation of human and material resources. However, traditional forecasting models often fail when abrupt distribution shifts occur, such as during pandemics or organizational changes, because they require frequent manual retraining. The objective of this study is to design a deep learning-based forecasting framework that maintains accuracy under evolving conditions without manual intervention or full retraining.

METHODS

An adaptive forecasting approach is proposed, built on a Seq2Seq LSTM architecture, enhanced with two independent strategies: Per-Sequence Scaling (PSS), which preserves sequence-specific statistics, and Continual Learning (CL), which incrementally updates the model with new data through replay buffers. These methods are evaluated separately against a standard LSTM, ARIMA, and Prophet. Historical daily ED visit data covering pre-COVID, COVID, and post-COVID phases are used. Performance is assessed across 1-day, 7-day, and 30-day horizons using MSE, MAE, MAPE, and MASE.

RESULTS

The adaptive LSTM significantly reduces forecasting errors compared to both statistical and deep learning baselines on short and medium horizons (1-day and 1-week), achieving MASE values consistently below 1. For 1-month forecasts, performance converges towards the seasonal naïve benchmark, with ARIMA occasionally achieving superior accuracy during abrupt COVID-related shifts. Importantly, the adaptive LSTM regains superiority in post-COVID conditions and maintains stable accuracy by continuously adjusting to new distributions without requiring full retraining, a property particularly valuable under hospital privacy and IT constraints, thereby ensuring robustness across changing regimes and reducing operational burden in hospital settings.

CONCLUSIONS

The adaptive Seq2Seq LSTM framework offers a robust and operationally viable solution for ED forecasting under evolving conditions. It consistently outperforms Prophet and ARIMA once data distributions stabilize, while only temporarily lagging behind ARIMA during structural shocks. Its ability to self-adjust incrementally removes the need for manual reconfiguration when data distributions change, making it particularly suitable for real-time hospital deployment under privacy and IT constraints. Future applications may extend beyond ED forecasting to other hospital departments facing similar variability in demand.