Mortality and morbidity associated with winter weather conditions revealed by machine learning models

Context: Machine learning (ML) models are being increasingly used to predict the health impacts of environmental exposures. Although cold temperatures account for a larger share of the disease burden than hot temperatures, the use of ML to assess how winter weather conditions affect mortality and morbidity remains limited. Objective: In this study, we compared ML models to predict cases of mortality and morbidity associated with winter conditions during the coldest months of the year in five regions of Quebec, Canada. Methods: ML models, including Least Absolute Shrinkage and Selection Operator (Lasso), Decision Tree (DT), Random Forest (RF), Light Gradient Boosting Model (LGBM) and Multi-Layer Perceptron (MLP), were calibrated to predict all-cause mortality, cardiovascular, and respiratory hospitalizations during November to March months, 2000 to 2019. Lagged weather variables (e.g., temperature, precipitation, humidity, snowfall, freezing rain) and air pollutants (O3, PM2.5, NO2, SO2 and CO) at up to 21 days were considered, as well as calendar variables (day of week, holiday, month, year). Feature importance and SHapley Additive exPlanations (SHAP) were used to reveal the most important predictors. Results: On the selected test winter years, RF and LGBM performed better than the other models across locations to predict cardiorespiratory hospitalizations. For mortality, while RF was slightly more performant in Montreal, Lasso and calendar-only models were highly competitive in the other locations. Lagged temperatures, precipitation and air pollutants were all important to predict the three health outcomes during winter, with some differences across outcomes. Calendar variables also played a major role, especially for cardiovascular hospitalizations. Discussion: Public health authorities could benefit from ML models to improve their understanding of winter weather condition associated with different health outcomes. Further studies should compare more thoroughly the performance of ML and non-ML approaches, and extend this associative study to a more causal study design.