Predictive Modeling of Acute Respiratory Infections: A Machine Learning Approach Using Real-World Data from Primary Care in Croatia

Timely forecasting of respiratory disease activity can support more proactive planning of primary healthcare resources. The aim of this paper is to develop a machine learning framework for forecasting acute respiratory infection (ARI) and influenza-like illness (ILI) volumes using routinely collected primary care real-world data from Croatia. We used aggregated and pseudonymized EHRs from the Croatian Central Health Information System (CEZIH), covering an 11-year period (ISO 2015-W01 to 2025-W52), to construct weekly county-level ARI and ILI series based on ICD-10 codes. The predictor set included historical disease incidence, antimicrobial prescriptions, demographic and population data, localized meteorological and air-quality variables, seasonal features, and spatial information from neighbouring counties. Five machine learning regressors were used for one- to four-week-ahead forecasting, using the period 2024–2025 for testing. XGBoost consistently achieved the best performance for both ARI and ILI. Models predicting week-to-week changes outperformed those predicting absolute case counts in most scenarios. For ARI, the best XGBoost models achieved MAPE values of 11.19%, 13.98%, 15.53%, and 16.23% for one- to four-week horizons, respectively, compared with 14.96%, 19.06%, 22.74%, and 25.19% for ILI. Seasonal variables were the most influential predictors, and prediction intervals widened with increasing forecasting horizon. Our results demonstrate the potential of primary care EHR data and machine learning to support surveillance and proactive healthcare resource planning by public health authorities and medical professionals in Croatia.