A Geo‐ <scp>AI</scp> ‐Based Approach for Spatiotemporal Prediction of Infectious Diseases With Cluster Distribution Pattern in Urban Environments

ABSTRACT Susceptibility prediction of infectious diseases is one of the important spatiotemporal issues in epidemiological disease management, especially in highly populated urban areas. This paper aims to accurately predict the spatiotemporal dynamics of contagious diseases, assuming a cluster‐based propagation pattern. COVID‐19, as one of these types, is the focus of the research. The main contribution is the proposal of a deep learning CNN‐based framework that utilizes two cluster‐based approaches, namely the attention‐aware (AT) mechanism and radial basis function (RBF), and a comparison with regression‐based approaches. Seven experiments have been designed, including OLS, GWR, LSTM, CNN, CNN/RBF, CNN/AT, and CNN/RBF/AT. The proposed algorithms have been implemented in Tehran, the capital of Iran, for 142,000 COVID‐19 occurrences over 12 months (February 20, 2021, to February 20, 2022). The evaluation process indicated the significant spatio‐temporal accuracy enhancement of CNN/AT/RBF (RMSE = 0.0087, R = 0.9994, AUC = 0.985, F ‐Score = 0.9781) relative to OLS (RMSE = 0.81, R = 0.6254, AUC = 0.6954, F ‐Score = 0.5241), GWR (RMSE = 0.024, R = 0.7111, AUC = 0.741, F ‐Score = 0.605), CNN/RBF (RMSE = 0.0082, R = 0.9857, AUC = 0.985, F ‐Score = 0.9655), CNN/AT (RMSE = 0.0101, R = 0.9802, AUC = 0.931, F ‐Score = 0.9155), CNN (RMSE = 0.0152, R = 0.8711, AUC = 0.841, F ‐Score = 0.7025), and LSTM (RMSE = 0.0098, R = 0.9745, AUC = 0.925, F ‐Score = 0.8824). Finally, to reveal the intrinsic mechanism of prediction, the one‐at‐a‐time sensitivity analysis and the explainability technique SHapley Additive exPlanations (SHAP) were used, and it identified the most significant impact of age, population density, gender, and air pollution factors on disease spread, respectively.