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

TL;DR

DeepTherm tackles the lack of heat-related mortality data by introducing a modular dual-prediction framework that separately forecasts all-cause and baseline mortality to infer excess mortality and predict deadly heatwaves. The system combines a Transformer-based all-cause predictor, a Quasi-Poisson baseline predictor, and a decision module with adjustable thresholds, and it leverages near-future synoptic typing data to anticipate heatwave events. Evaluated across 12 Spanish provinces and their capitals (1995–2023), DeepTherm demonstrates robust, regionally diverse performance with strong L1 detection and meaningful L2 results, while showing stability across years and demographic groups. It outperforms baselines that do not utilize historical all-cause mortality data, and its modular design supports flexible deployment under data availability and decision-cost constraints.

Abstract

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.

Figures (7)

• Figure 1: DeepTherm is capable of using all-cause mortality history with non-mortality records to yield early warnings for deadly heatwaves. Existing studies of deadly heatwaves remain on projecting non-mortality variables to heat-related mortality, which cannot use historical all-cause mortality records and need heat-related mortality for calibration. Our system, DeepTherm, bypasses this challenge with a novelly proposed dual-prediction pipeline. Using historical all-cause mortality data, historical non-mortality data, and near-future synoptic weather-typing data, DeepTherm isolates heat-related and baseline mortality components, enabling deadly heatwave identification without requiring heat-related mortality data. Furthermore, DeepTherm determines whether an incoming heatwave will be deadly via a flexible thresholding strategy, capable of balancing false alarms and missed alarms.
• Figure 2: A visualization of DeepTherm's performance on Spain (mainland). Figure (a) (b) shows the map chart noting the provincial prediction accuracies, and Figure (c) (d) notes the city-level prediction accuracies. Provinces outside Iberian Peninsula are hidden for brevity.
• Figure 3: DeepTherm demonstrates consistent performance across most years, with improved accuracy as more data becomes available. The figures present annual heatwave prediction results (smoothed using a three-year sliding window) for city-level and provincial data from 1997 to 2023 across 12 selected Spanish cities, evaluated by Precision and Recall metrics. Cities are grouped by climate zone: 1) Continental, 2) Mediterranean, and 3) Atlantic. DeepTherm maintains stable performance across most years while showing gradual improvement over time, except during the COVID-19 pandemic period. Notably, despite the significant challenges of mortality prediction during the pandemic, DeepTherm achieves strong performance in recent years, particularly in Continental regions with more frequent heatwave occurrences. This suggests the model's particular effectiveness in areas with higher heatwave prevalence, highlighting its potential for broad impact in vulnerable regions.
• Figure 4: DeepTherm demonstrates consistent performance across both younger (under 65) and older (65+) population groups in all cities. We evaluated DeepTherm by replacing all-age mortality records with age-specific mortality data (Under-65 and 65+ populations). Figures (a)-(d) present the average precision and recall for both age groups across all cities and regions, while (e)-(l) display city-level and provincial variations in performance metrics (averaged across L1 and L2 heatwaves) compared to all-age population predictions. The results show that DeepTherm maintains comparable performance when predicting heatwave impacts for both age groups relative to all-age predictions.
• Figure 5: Comparing the prediction results of DeepTherm to previous deadly heatwave prediction methods that project non-mortality variables to excess mortality ratio and cannot use historical all-cause mortality records.DeepTherm achieves significantly better recall while maintaining superior precision, implying DeepTherm predicted more missing alarms without causing more false alarms. This highlights the advantage of capability inside using historical mortality data, which is introduced by our novelly designed dual-prediction pipeline.