Exploratory short-term forecasting of respiratory hospital admissions in lisbon using air-quality and meteorological data

This is the authors' abstract. We don't add key points for this paper.

Air pollution and meteorological variability are key drivers of short-term fluctuations in hospital admissions (HAs) for respiratory diseases, yet most existing studies focus mainly on retrospective associations with past hospitalizations or on same-day predictions, which limits their operational value for proactive healthcare planning. This study evaluates different machine-learning (ML) and deep-learning (DL) forecasting frameworks to predict daily HAs up to three days ahead using routinely available air-quality and meteorological data. Three preprocessing scenarios are evaluated to assess the role of temporal and environmental information: three-day sequence forecasting (S1), independent same-day modeling without historical context (S2), where three consecutive days are modeled as independent observations rather than as a temporal sequence, and three-day sequence forecasting with one-week historical information (S3). The input feature space includes air quality, multiple meteorological variables, and calendar-based indicators, together with lagged and rolling-window features designed to capture short-term environmental memory. The sequence-based HGB and RNN framework (S1 and S3) consistently delivered strong three-day-ahead forecasts, whereas the same-day model without historical context (S2) showed substantially weaker generalization, highlighting the importance of short-term temporal memory for predicting HAs. The multi-day recurrent models demonstrated high stability and accuracy across consecutive forecast horizons, with close agreement between predicted and observed admissions under normal operating conditions and increased uncertainty only during rare peak-demand events. These results demonstrate the strength of the proposed framework in capturing the dynamic, environmentally driven patterns of respiratory hospital demand and its suitability for operational healthcare forecasting.