Modular Deep-Learning-Based Early Warning System for Deadly Heatwave Prediction

Severe heatwaves in urban areas significantly threaten public health, calling for establishing early warning strategies. Despite predicting occurrence of heatwaves and attributing historical mortality, predicting an incoming deadly heatwave remains a challenge due to the difficulty in defining and estimating heat-related mortality. Furthermore, establishing an early warning system imposes additional requirements, including data availability, spatial and temporal robustness, and decision costs. To address these challenges, we propose DeepTherm, a modular early warning system for deadly heatwave prediction without requiring heat-related mortality history. By highlighting the flexibility of deep learning, DeepTherm employs a dual-prediction pipeline, disentangling baseline mortality in the absence of heatwaves and other irregular events from all-cause mortality. We evaluated DeepTherm on real-world data across Spain. Results demonstrate consistent, robust, and accurate performance across diverse regions, time periods, and population groups while allowing trade-off between missed alarms and false alarms.

💡 Research Summary

This paper introduces “DeepTherm,” a modular deep-learning-based early warning system designed to predict deadly heatwaves, addressing a critical gap in public health preparedness. The core challenge lies in the difficulty of obtaining precise heat-related mortality data, which is essential for forecasting the lethality of an impending heatwave. Existing systems often rely on temperature thresholds or require historical heat-mortality data for calibration, limiting their practicality and accuracy.

DeepTherm innovatively circumvents this data scarcity through a novel dual-prediction pipeline. The system operates by concurrently forecasting two key metrics: (1) all-cause mortality using a complex deep learning model (based on Transformer architecture) that analyzes short-term historical data on mortality and non-mortality factors (e.g., temperature, humidity), and (2) baseline mortality (the expected mortality in the absence of irregular events like heatwaves) using a simpler Quasi-Poisson regression model trained on long-term data. The difference between these two predictions yields an estimate of excess mortality, which serves as a proxy for heat-related impact. This estimate is then used to classify an incoming heatwave (identified using near-future synoptic weather-typing data) into three levels: Level 0 (non-deadly), Level 1 (excess mortality >15% of baseline), and Level 2 (excess mortality >30% of baseline).

The system is evaluated extensively on real-world data from 12 provinces and their capital cities across Spain, covering different time periods (provinces: 2015-2023, cities: 1995-2023). The evaluation simulates a real-time operational setting, where the model is retrained annually with newly available data. Results demonstrate that DeepTherm achieves robust and accurate performance. For provincial Level 1 heatwave prediction, it attained an average accuracy of 77.0%, precision of 80.6%, and recall of 87.4%. Performance was consistently strong across diverse geographic regions with varying climate conditions, different age groups (younger and older populations), and evaluation years. The system performed particularly well in regions with higher heatwave frequency. Furthermore, DeepTherm incorporates a flexible decision module that allows policymakers to adjust the alarm threshold, enabling a customizable trade-off between the costs of false alarms and missed detections based on operational needs.

In summary, DeepTherm presents a practical and effective solution for deadly heatwave prediction without requiring hard-to-obtain heat-specific mortality data. Its modular design, combining the pattern recognition power of deep learning with the stability of statistical modeling for baseline estimation, offers significant advantages in terms of robustness, generalizability, and operational flexibility for public health early warning systems.