Multi-task CNN–LSTM architecture with hybrid PSO–GWO optimization and SHAP explainability for joint prediction of urban air quality and health impact

Health risks and accurate air quality prediction play a crucial role in the environmental surveillance and protection of the public health. Current methods are limited to single task classification of Air Quality Index (AQI) with few features and do not consider estimating the health impacts or multi-modal environmental data. The study proposes a new multi-task CNN-LSTM deep learning model, which simultaneously predicts the category of AQI and the health impact scale score with two output heads. The three different data modalities, including pollution indicators, meteorological variables, and anthropogenic activity factors, are combined into one multi-modal data type in the framework. The branches of the CNN are used to draw the interactions between spatial features of each mode, whereas the LSTM layers simulate the temporal environmental relationship. A hybrid PSO-GWO algorithm is dynamically adjusting model hyperparameters, and SHAP is multi-level explainable. The proposed model shows AQI classification accuracy of 95.30%, macro ROC-AUC of 0.9919, and health impact prediction R2 = 0.8945. SHAP indicates CO, proximity to industries, and NO2 as main drivers of AQI, whereas PM10, industrial proximity, and CO are predominant for health risk. The suggested framework presents an explainable, and implementable resolution to smart environmental monitoring and population health decision-making.