A Machine Learning-Based Framework for Real-Time Environmental Health Risk Prediction and Spatial Air Quality Intelligence in Urban-Industrial Ecosystems

International audience The high rate of industrialization and urbanization has enhanced the pollution of the environment and the occurrence of occupational health hazards especially in urban-industrial ecosystems where people are subjected to cumulative environmental and occupational hazards. Traditional monitoring is mostly reactive and inhibited in the ability to capture nonlinear, high frequency and spatially heterogeneous exposure patterns. The paper presents and assesses a machine learning-supported framework of real-time predictions of environmental health risks and spatial air quality intelligence. The architecture integrates the multi-sources of data (environment, weather, work-related exposures, demographic risk factors, etc.) into a single predictive pipeline. A high-resolution, 30-day, controlled simulation (5 minutes each) was created in five representative areas (industrial, high-traffic urban, residential, suburban, and rural) and produced 43,200 spatiotemporal records. There was an implementation of ensemble models (Random Forest, Gradient Boosting, XGBoost) and a Long Short-Term Memory (LSTM) network with regard to pollutant prediction and Health Risk Index (HRI) estimation. The regression output reported a high predictive ability, XGBoost (R² = 0.92), Random Forest (R² = 0.91), and Gradient Boosting (R² = 0.89), and Gradient Boosting (R²), and LSTM exhibiting the best performance in terms of temporal modelling accuracy, which indicates the significance of the consideration of the sequential dependency. The application of Risk classification (Low, Moderate, High, Critical) was found to have 93 percent accuracy with F1-scores of over 0.90 and AUC values of between 0.91 and 0.95. Temporal realism was observed by characteristic bimodal peaks in PM2.5 day by day simulations, which were in agreement with the cycles of traffic and industry. Interpretability and policy relevance were also added by incorporating explainable Artificial Intelligence methods, especially SHAP. The developed framework takes the current environmental analytics a step forward and integrates predictive modeling, spatial risk mapping, and stakeholder-oriented intelligence and offers the proactive governance of the whole community of health-related issues and sustainable urban-industrial growth. The current text fits into the developing literature of combining environmental science, population health, and artificial intelligence. The study offers a novel method to forecast the environmental health risks of urban-industrial ecosystems by creating a machine learning-based model to integrate the environmental, occupational, meteorological, and demographic factors. It (Integration of ensemble learning models and Long Short-Term Memory (LSTM) networks with a composite Health Risk Index (HRI)) presents a viable tool of real-time environmental risk assessment and decision support. Also, the transparency and policy relevance are improved by the addition of explainable artificial intelligence (XAI) methods, thus the framework is useful to researchers, environmental authorities, and urban planners working on the development of smart cities sustainably and with health considerations.