This study aims to explore the predictive strength of interactions among physical examination indicators regarding diabetes risk. It specifically addresses the utilization of the LightGBM polynomial kernel model for early diabetes screening and prognosis. Methods: The study utilized the PolynomialFeatures method to derive high-order interaction data from physical examination indicators. Employing the IV feature selection model, it identified strongly predictive factors, which informed the inputs for the LightGBM polynomial kernel prediction model to predict the risk of diabetes, with the model’s predictive performance evaluated based on the AUC. Results: The LightGBM prediction model, established using high-order factors selected by the IV model for their strong predictive ability, achieved an AUC of 0.9687 (95%CI: 0.9612~0.9762). Conclusion: The LightGBM model, built on high-order interaction factors with robust predictive power, shows significant potential for diabetes risk prediction in populations undergoing physical examinations.