Inferring daily lethal heat mortality from weekly death records using machine learning

Extreme heat has a complex and delayed effect on human mortality operating across sub-daily to weekly timescales. Many large-scale mortality datasets are reported at weekly resolution, with stratified age brackets. However, temporal aggregation in prediction can obscure short-lived lethal heat episodes and lead to underestimation of heat-related mortality. Methods for estimating temperature–mortality relationships from temporally aggregated data have been explored within statistical frameworks, which remain the standard approach in environmental epidemiology, but these approaches constrain the form of the risk function and limit the flexibility of predictor representations.

We propose a machine-learning framework that enables daily mortality prediction with no restriction on the temporal resolution of the training mortality dataset. The method learns a heat-related risk function conditioned on lagged sequences of recent daily meteorological conditions and regional socio-environmental characteristics. Weekly expected deaths are decomposed into daily estimates which are multiplied with the learned risk factors to get the model’s daily predicted deaths. The daily death predictions for a week are summed to a weekly total to match the available temporal resolution of the death dataset. The gradients of the learned risk factor propagate through this aggregation step, allowing the model to learn temporally resolved mortality responses without requiring daily death labels.

The framework is trained using weekly NUTS-3 Eurostat mortality data with five-year age stratification, together with high-resolution MSWX daily reanalysis meteorology. Validation is performed using a French 2019 individual-level daily mortality dataset, which reports spatial, age, and sex information for all registered deaths in France, enabling direct evaluation of predicted daily deaths aggregated to consistent spatial and age resolutions.

We expect this approach to recover intra-week variability in mortality associated with short-duration temperature signals, outperform uniform or heuristic temporal disaggregation methods, and improve attribution of lethal heat events. By linking daily climate exposure to weekly mortality records without requiring more fine-grained data collection, this method expands the analytical value of existing mortality datasets and supports more timely assessment of lethal heat risk.