Deep learning based predictive early warning system for heat and ventilation risks in naturally ventilated hospital wards

Conference item

Hospital wards in tropical climates face persistent challenges in maintaining thermal comfort and indoor air quality, where mechanical ventilation is limited. Using over two years of 15-minute sensor data across 14 naturally-ventilated wards in two Ghanaian hospitals, together with weather data, we develop an early warning system to forecast indoor temperature, relative humidity, and CO₂ over a 24-hour horizon. We compared a multi-output XGBoost baseline with a Temporal Fusion Transformer (TFT). Results reveal a dual risk profile, where overheating and high humidity are widespread and strongly diurnally structured, whereas CO₂ exceedance is more localised and time-dependent. TFT improves temperature and humidity forecasting, while CO₂ prediction remains challenging and better served by XGBoost. Overall, the proposed approach can inform proactive interventions ahead of critical events and demonstrates the feasibility of data-driven ventilation forecasting in tropical hospitals, offering a scalable pathway toward smart, health-oriented indoor climate management in naturally ventilated facilities.

Authors: Tekler, ZD; Roberts, BM; Low, R; Cook, M

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: International Society of Indoor Air Quality and Climate
• Host title: Indoor Air 2026 Proceedings
• Publication date: 2026-10-01
• Acceptance date: 2026-04-14
• Event title: 19th International Conference of the International Society of Indoor Air Quality and Climate (INDOOR AIR 2026)
• Event location: Singapore
• Event website: https://www.indoorair2026.org/
• Event start date: 2026-06-14
• Event end date: 2026-06-18
• Language: English
• Keywords: overheating; insufficient ventilation; early warning system; machine learning; hospitals