Aiming at the problems of heterogeneous variables, more missing variables, class imbalance and complex risk boundaries in epidemiological data, this paper constructed an optimized support vector machine risk prediction model for multi-source monitoring information. In this study, demographic characteristics, physical signs, laboratory results, exposure information and dynamic change characteristics were incorporated into the unified computing framework, and the identification ability of the model for high-dimensional nonlinear risk patterns was improved through missing repair, feature screening, category reweighting, kernel parameter joint optimization and probability calibration. The experimental results based on structured epidemiological samples show that the Accuracy of the model is 89.47%, the Recall is 90.10%, the AUC is 0.931, the Brier Score is reduced to 0.098, and the average inference delay is 138 ms. The proposed method achieves a good balance between prediction accuracy, risk ranking and deployment efficiency, and can provide computational support for primary screening, key population identification and regional public health early warning.
