Climate-Informed Machine Learning for Predicting Heat-Related Health Risks: Towards Resilient and Personalized Early Warning Systems

Severe heat events cause over 489,000 deaths each year worldwide, with forecasts suggesting a 2.3–4.5 times rise in heat-related illnesses by 2050 if climate change remains unmanaged. In Nigeria, more than 75,000 hospitalizations in 2023 were due to heat stress, dehydration, and heart-related issues, impacting vulnerable populations like the elderly, outdoor workers, and children. Existing heat-health alert systems are primarily focused on the population and generic, frequently missing personalized and practical recommendations. To fill this gap, this research introduces a climate-aware machine learning (ML) framework that combines environmental factors (temperature: 28–45 °C, humidity: 20–85%, PM₂.₅: 15–110 µg/m³) with personal health data (heart rate variability at 1 Hz, blood pressure every 15 minutes, oxygen saturation 92–99%, and daily activity levels) to assess heat-related illness risk on an individual basis. Utilizing a longitudinal dataset comprising 52,846 records from 2,300 patients in southwestern Nigeria (2021–2024), we evaluated traditional classifiers (Logistic Regression, Random Forest) in comparison to sophisticated sequence models (LSTM, Bi-LSTM, Transformer). The Transformer model recorded the highest performance, attaining 92.7% accuracy, 0.88 precision, 0.91 recall, 0.89 F1-score, and a ROC–AUC of 0.94, surpassing LSTM (89.4% accuracy, 0.90 ROC–AUC) and Random Forest (84.1% accuracy, 0.82 ROC–AUC). Temporal analysis showed that increasing input sequence lengths from 10 to 50 time steps enhanced prediction accuracy by 3.8–4.1% across models, highlighting the significance of long-term observation. Even with increased training expenses (Transformer: 21.6 minutes compared to Logistic Regression: 1.4 minutes), the inference latency stayed under 2.3 ms per sample, making it suitable for real-time use in mobile and wearable applications. Simulation analyses additionally project that early interventions guided by this framework might decrease hospital admissions by 28–32% and avert as many as 11,500 heat-related fatalities each year among high-risk populations in Nigeria. This research highlights the capability of climate-aware, patient-focused ML models to enhance health by aligning with SDG 3 (Health) and SDG 13 (Climate Action).