Explainable Multimodal Temporal Learning for Air-Pollution-Associated Respiratory Health Monitoring: Study Protocol, Proposed MTAF-Net Framework, and Preliminary Longitudinal Dataset Analysis
Abstract
Air pollution is an important environmental risk factor for respiratory morbidity, particularly among individuals with asthma, chronic obstructive pulmonary disease (COPD), and chronic bronchitis. This study proposes an explainable Multimodal Temporal Attention Fusion Network (MTAF-Net) for longitudinal respiratory-health monitoring by integrating environmental exposure, wearable physiological measurements, and clinical information. The MTAF-Respire Dataset contains 360,000 longitudinal observations from 500 participants recorded at six-hour intervals from January to June 2026. The dataset includes participant-level clinical characteristics and repeated PM₂.₅ and PM₁₀ measurements, with NO₂ available for approximately 49.3% of observations. Current data completeness, however, limits predictive modelling because wearable physiological, symptom, engineered temporal-exposure, and prospective deterioration variables are not populated. Descriptive analysis showed mean PM₂.₅ exposure of 46.30 µg/m³, mean PM₁₀ exposure of 81.03 µg/m³, and mean FEV1/FVC of 0.707. PM₂.₅, PM₁₀, and NO₂ showed strong correlations, while participant-level associations between pollutant exposure and baseline pulmonary function were weak. The proposed MTAF-Net is designed to use seven-day historical information to predict respiratory deterioration over 24-hour, 48-hour, 72-hour, and 7-day horizons and provide global, local, and temporal explanations. No predictive model training, evaluation, or explainability analysis was performed in this study, because the deterioration-outcome labels and wearable physiological data required for supervised modelling are not yet populated in the analytical file. The present study instead establishes the dataset structure, reports the available descriptive and correlational findings, and defines the modelling protocol and framework to be used for subsequent prospective predictive evaluation once the required variables become available.