OBJECTIVE: This study aimed to investigate the feasibility of developing machine learning models for non-invasive prediction of () infection using routinely collected adult health screening data, including demographic characteristics and clinical biomarkers, to establish a potential decision-support tool for clinical practice. METHODS: The data was sourced from the adult health examination records within the health management centers of the hospital. The Least Absolute Shrinkage and Selection Operator (LASSO) regression was employed for feature selection. Six distinct machine learning algorithms were utilized to construct the predictive models, and their performance was comprehensively evaluated. Additionally, the SHapley Additive Projection (SHAP) method was adopted to visualize the model features and the prediction results of individual cases. RESULTS: A total of 10,393 subjects were included in the dataset, with 3,278 (31.54%) having infection. After feature screening, 10 factors were selected for the prediction model. Among six machine-learning models, the Extra Trees model had the best performance, with an AUC of 0.827, Accuracy of 0.744, and Recall of 0.736. The Random Forest model also did well, with an AUC of 0.810. XGBoost attained an AUC of 0.801, indicating moderate predictive capability. SHAP analysis showed that age, WBC, ALB, gender, and wasit were the top five factors affecting infection. Higher age, WBC, wasit and lower ALB were linked to a higher infection probability. These results offer insights into infection risk factors and model performance. CONCLUSION: The Extra Trees classifier exhibited the optimal performance in predicting infections among the evaluated models. Additionally, the SHAP analysis enhanced the interpretability of the model, which offers valuable insights for early-stage clinical prediction and intervention strategies.