Spatio-Temporal Machine Learning Framework for Urban Health Risk Evaluation

Abstract

Air Pollution is a foremost health concern in megacities such as Delhi, where fine particulate matter (PM₂.₅) levels every so often exceed international safety guidelines. While air quality indices offer a general idea of air quality levels, these do not account for differences in individual exposure levels due to localized factors such as congestion levels. Therefore, this study focuses on a innovative exposure-oriented air quality index which is based on Urban Health Risk Index (UHRI), that basically is an amalgamation of high-resolution PM₂.₅ forecasting along with congestion-driven exposure augmentation factors. To develop this framework, hourly CPCB data and high-resolution traffic queue density data were processed to develop time-aware autoregressive/rolling features using gradient boosting techniques such as XGBoost. Strong performance was observed, achieving 13.3 µg/m³ root mean square error (R² = 0.9966) for PM₂.₅ concentration forecasting and 0.0299 units root mean square error (R² = 0.955) for traffic density prediction. A mechanistic congestion multiplier, accounting for emission characteristics and street canyon effects, was then used to develop amplification factors between 1.0 and 2.3. UHRI is developed based on the fusion of pollutant forecasts and congestion multipliers. This allows for the estimation of health risk in a microenvironment-sensitive manner beyond city-level air quality indicators. The experimental results demonstrate that even under moderate congestion conditions, health risks due to PM2.5 exposure are substantially increased. Moreover, in severe gridlock conditions, health risks are even increased by more than double compared to ambient values. The proposed framework is likely to be useful as an understandable and viable decision support system for health management in cities. Future work may include spatial modeling, uncertainty analysis, and even intelligent traffic control system incorporation.