Hyperlocal forecasting of acute respiratory infections using machine learning and digital epidemiology signals: a multi-province study in Türkiye

Acute respiratory infections (ARI) are a major contributor to morbidity and seasonal pressure on health systems. Conventional surveillance systems rely on clinical and laboratory reporting, often resulting in delays that limit timely public health response. This study aimed to assess whether integrating routine surveillance data with digital search behavior and meteorological variables could improve short-term forecasting of ARI at the district level. We analyzed weekly ARI incidence across 69 districts in four climatically diverse provinces in Türkiye over a five-year period. District-level epidemiological data were combined with meteorological indicators and Google Trends search data. Multiple machine-learning models, including ridge regression, random forest, XGBoost, and ensemble methods, alongside a univariate ARIMA baseline, were developed to forecast ARI incidence one and two weeks ahead. Model performance was evaluated using R² and mean absolute percentage error (MAPE), with predefined thresholds for operational reliability. Forecasting models demonstrated moderate predictive performance, with a median R² of 0.45 for one-week-ahead predictions. Overall, 92.8% of districts met the predefined reliability threshold (MAPE < 50%). Models based on recent incidence trends, particularly ridge regression, performed comparably to or better than more complex approaches. Autoregressive features were the strongest predictors, while meteorological variables contributed limited additional explanatory value. Google Trends indicators provided complementary information: symptom-related searches were closely aligned with observed incidence, whereas broader awareness-related searches often preceded subsequent declines in case counts, hypothesized to reflect proactive behavioral shifts. A geographic gradient in epidemic timing was observed, with northern districts experiencing earlier peaks by approximately 7 weeks compared to southern districts. Short-term forecasting of ARI at the district level is feasible using simple, interpretable models that rely primarily on recent incidence data. The integration of digital behavior signals can enhance situational awareness and support timely public health decision-making. Observed regional differences in epidemic timing suggest the potential value of geographically tailored intervention strategies, including vaccination timing.