AirQuaNet: A Convolutional Neural Network Model With Multi-Scale Feature Learning and Attention Mechanisms for Air Quality-Based Health Impact Prediction

Abstract

Air pollution is a vital environmental and public health issue responsible for millions of early deaths worldwide every year. Predicting air quality and associated health impacts is crucial for early intervention and informed policy planning. This work presents Air Quality Network (AirQuaNet), a novel convolutional neural network (CNN) designed to predict air quality levels accurately. AirQuaNet integrates deep learning (DL) innovations, namely Multi-Scale Convolutional Blocks (MSCBs), residual connections, and self-attention mechanisms, to enhance its feature extraction capabilities and enable it to learn long-range temporal dependencies. The MSCBs employ four parallel 1D convolutional layers with different kernel sizes, enabling the model to extract multi-scale features critical for learning patterns in complex environmental data. Residual connections are employed to prevent vanishing gradient issues during training, and the self-attention mechanism dynamically weights informative inputs, improving the model's attention towards informative pollutant features. AirQuaNet was evaluated on two public datasets, the Air Quality and Health Impact Dataset and the Comprehensive Health Data for Asthma. It achieved outstanding results, with an R-2 of 0.9997 on regression tasks and a classification accuracy of 94.21%, outperforming traditional machine learning algorithms and DL baselines. These results highlight the model's robustness under diverse data environments and its ability for high generalization across varied temporal scales and types of contaminants. AirQuaNet also offers high scalability, rendering it an excellent candidate for real-time urban surveillance networks. Its ability to deal with large, heterogeneous datasets qualifies it as a top contender for environmental health forecasting, policy support, and early warning systems.