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 leading risk factor for cardiorespiratory morbidity and mortality, yet translating raw air-quality signals into individualized health-risk estimates remains challenging due to spatial–temporal heterogeneity, missing data, and non-linear pollutant–health relationships. We present AirQuaNet, a convolutional neural network that fuses multi-scale feature learning with dual attention (channel and temporal–spatial) to predict short-term health impacts (e.g., daily asthma ER visits risk or symptom exacerbation probability) from multimodal inputs: pollutant concentrations (PM₂.₅, PM₁₀, NO₂, O₃, SO₂, CO), meteorology (temperature, humidity, wind), satellite AOD, traffic/activity proxies, and calendar effects. AirQuaNet stacks dilated temporal convolutions and multi-kernel spatial convolutions to capture fine-grained bursts and long-range trends, while attention reweights salient pollutants, time steps, and locations. On benchmark city-level datasets (2018–2024) and a held-out 2025 cohort, AirQuaNet improves MAE by 9–18% and AUROC by 3–7% over strong baselines (LSTM, Temporal CNN, XGBoost). Ablations show both multi-scale design and attention are necessary for consistent gains and robustness to missing sensors. The model supports uncertainty quantification via Monte-Carlo dropout and produces interpretable weekly attributions that align with known pollution episodes. AirQuaNet is deployment-ready through a lightweight inference stack and can operate with partial inputs using learnable masking.