HARMONY: Hierarchical Machine Learning Framework for Disease Outbreak Forecasting in Nursing Homes

Background: Nursing homes experience infectious disease outbreaks as sparse, clustered, and heterogeneous events. Conventional single-stage forecasting models are often poorly matched to this structure because they learn from zero-dominated outcome distributions and conflate two distinct processes: whether an outbreak will occur and how large it may become. This limits their utility for infection prevention and control (IPAC) teams that need actionable, facility-level early warning signals. Objective: This study developed and evaluated a hierarchical machine learning framework for infection outbreak forecasting in nursing homes (HARMONY) using longitudinal IPAC indicators. Methods: We conducted a retrospective longitudinal study using 1040 facility-month observations from 26 nursing homes in Ontario, Canada, collected between August 2021 and December 2024. Each observation included facility-level IPAC process indicators, vaccination measures, audit performance, education completion, regulatory compliance, and outbreak counts. HARMONY decomposed forecasting into two linked stages: a gate model for binary outbreak occurrence and a conditional intensity model for outbreak magnitude among outbreak-positive observations. Models were trained using temporal validation to preserve chronology and reduce information leakage. Candidate classifiers included Light Gradient-Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), CatBoost, Random Forest, and Tabular Prior-Data Fitted Network (TabPFN). Candidate regressors included gradient-boosting models and temporal neural architectures, including temporal convolutional networks (TCNs), gated recurrent units (GRUs), bidirectional long short-term memory (BiLSTM) networks, time-series transformers (TSTs), and temporal fusion transformers (TFTs). Performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, F1score, root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and mean absolute scaled error (MASE). SHapley Additive exPlanations (SHAP) identified predictors contributing to model outputs. Results: Outbreaks were temporally clustered and uneven across facilities, while most IPAC indicators showed weak linear correlation with outbreak counts, supporting nonlinear temporal modeling. In the occurrence stage, LightGBM achieved discrimination, with an AUC of 0.950, sensitivity of 89.1%, and specificity of 91.6%. In the magnitude stage, the TFT achieved the lowest RMSE of 0.3908, while the TCN achieved the lowest MAE of 0.1741 and MASE of 0.3471. HARMONY outperformed single-stage baseline regression models, reducing the best RMSE from 0.6817 to 0.3908 and the best MAE from 0.5563 to 0.1741. Improvement was significant using the Wilcoxon signed-rank test (P = 0.00391). SHAP analysis identified hand hygiene compliance, adenosine triphosphate audit performance, staff COVID-19 booster vaccination, resident respiratory syncytial virus vaccination, and IPAC education completion as influential predictors. Conclusions: HARMONY reframes nursing home outbreak forecasting as a gated surveillance problem rather than a single count-prediction task. By separating outbreak emergence from outbreak escalation, the framework represents zero-inflated, episodic outbreak dynamics and converts routine IPAC data into operational early warning signals for targeted outbreak preparedness and response.