Predictions of City-based Respiratory Hospital Visits: Developing and Validating a Machine Learning Model with a Novel Composite Air Pollution Index.

OBJECTIVE: City-specific tools for assessing and warning about respiratory disease risks are underdeveloped, limiting effective public health response. This study aimed to develop and validate a novel city-specific prediction framework (WHA air-LSTM) for forecasting daily respiratory outpatient visits by integrating a composite air pollution health index. METHODS: Based on over 223.7 million hospital visits across multiple megacities, we constructed and validated a five-level morbidity-driven composite air pollution index (WHA air) for each city using city-specific exposure-response relationships. An LSTM model was built using WHA air, temperature, humidity, and historical visit data to predict next-day visits. The proposed modeling framework was developed with city-level data, and it was externally validated using datasets from other cities. RESULTS: Higher WHA air levels were significantly associated with increased outpatient visits. The model demonstrated excellent predictive performance (Beijing: R 2 = 0.963, RMSE = 53.5) and effectively captured visit surges. Excluding WHA air degraded model accuracy (ΔRMSE = +44.1%). The framework maintained robust performance in external validation, confirming its transferability. CONCLUSION: The WHA air-LSTM framework provides a scalable and practical tool for city-level respiratory disease early warning by bridging environmental monitoring with clinical practice.