Towards Early Warning Systems for Shigellosis in Africa: Multi-Horizon Forecasting of Incidence in Rural Western Kenya Using Statistical State-Space and Machine Learning Models

Introduction Despite the significant burden of Shigellosis, routine surveillance systems in resource-limited settings lack predictive capabilities to guide proactive public health responses. Consequently, we developed and evaluated the accuracy, uncertainty and computational cost of multi-horizon forecasting models for Shigella incidence among children <5 years in rural western Kenya using 16-year longitudinal incidence time-series. Methods We estimated monthly Shigella incidence using data from three pediatric enteric studies (2008-2024), supplemented with Kenya Health Information System during inter-study gaps. Incidence estimates incorporated sampling, healthcare-seeking, stool-testing, and pathogen-specific weights, with bootstrap-derived 95% confidence intervals. Predictors included Google Trends health-related indicators, climate, temporal trends, and seasonality. Forecasts at 1-, 3-, and 6-month horizons were evaluated using rolling-origin validation. Following horizon-specific predictor selection using Boruta, 14 state-space, machine-learning, and hybrid models were compared using point-forecast and probabilistic metrics, bootstrap confidence intervals, and Model Confidence Sets. Adaptive conformal prediction recalibrated prediction intervals, while Shapley Additive exPlanations assessed predictor contributions. Results Mean monthly Shigella incidence was 2.51 per 100 child-years (standard deviation=1.85). Shigella dynamics in Kenya can be forecast beyond seasonal persistence alone for up to six months reducing absolute errors by ~ 33%. However, optimal model architectures and predictor sets shift systematically across forecasting horizons. Temporal trend and recent incidence were the dominant predictors, while predictor importance shifted with forecast horizon toward syndromic, pharmaceutical, seasonal, and lagged environmental signals. Extreme Gradient Boosting performed best at 1 month (Root mean squared error [RMSE]=1.86), remained competitive at 3 months (RMSE=1.89), whereas exponential smoothing (ETS) performed best at 6 months (RMSE=2.05). Forecast accuracy declined with increasing horizon. ETS and TBATS provided the best-calibrated prediction intervals, whereas machine-learning models generally under-covered. Model Confidence Sets demonstrated horizon-dependent model distinguishability. Conclusion Horizon-specific model selection balancing predictive accuracy, uncertainty calibration, and computational feasibility could strengthen operational early warning systems for shigellosis.