npj Antimicrobials and Resistance Biological sciences
SOLVAR: hybrid endemic-epidemic machine learning for early forecasting of high-incidence multidrug-resistant organism periods
Abstract
Healthcare-associated multidrug-resistant organism (MDRO) outbreaks require proactive infection control, yet traditional surveillance remains reactive, and machine learning (ML) models ignore transmission dynamics. We developed Stochastic Outbreak Learning and Variance-Aware Regression (SOLVAR), a hybrid framework to classify high-incidence periods 1–3 weeks ahead. SOLVAR adaptively ensembles a direct ML classifier, an ML-informed endemic-epidemic (EE) expert, and a baseline EE expert, with weights learned from walk-forward validation performance. Using 44,386 incident episodes across five MDROs from a 1700-bed hospital (2021–2026), we benchmarked SOLVAR against ML-only, standalone EE, Poisson generalized linear model, autoregressive integrated moving average, and naive baselines using rolling origin walk-forward validation over 96 weeks. SOLVAR achieved the highest area under the receiver operating characteristic curve for carbapenem-resistant Enterobacterales (CRE) and multidrug-resistant Acinetobacter baumannii at every forecast horizon (up to 0.955 and 0.718), with balanced sensitivity and specificity for a 1-week CRE high-incidence period forecast (0.769 and 0.829; negative predictive value 0.906). For highly endemic methicillin-resistant Staphylococcus aureus , SOLVAR yielded a non-degenerate classifier where ML-only and standalone EE models failed. By combining ML and EE experts through performance-based adaptive weighting and explicitly modeling overdispersion, SOLVAR supports pathogen-specific classification of high-incidence periods 1–3 weeks ahead to inform infection control planning.