Integrating ARIMA and Deep Learning Models for Counterfactual Evaluation of Nirsevimab's Early Impact on RSV Infant Admissions.

This study presents a hybrid modelling framework that integrates classical statistical methods (ARIMA) with deep learning architectures to enhance longitudinal forecasting of hospital admissions. We conducted a preliminary, hospital-based evaluation of the early impact of nirsevimab (Beyfortus®), a monoclonal antibody against respiratory syncytial virus (RSV), on infant hospitalisations at the Hôpitaux Universitaires de Genève (HUG), Switzerland. Weekly RSV-positive admissions (April 2021-October 2025) were analysed using ARIMA-Fourier and hybrid deep learning models trained on pre-intervention data (January 2022-October 2024). Estimated reductions in RSV-related hospitalisations were 44% for infants under 12 weeks (ARIMA 44.3%, 95% CI [21.7-143]) and 39% for those under 12 months (ARIMA 39.5%, 95% CI [35-280]). Hybrid deep learning models yielded consistent reductions (35-48%), confirming predominantly linear RSV dynamics and reinforcing confidence in the observed post-nirsevimab decline. These preliminary findings demonstrate the potential of hybrid statistical-machine learning frameworks for early, interpretable, and robust evaluation of vaccine impact using real-world hospital data.