Upper extremity injuries in baseball players demand advanced prevention. Our study analyzed clinical features using machine learning techniques to provide precise and individualized injury risk assessment and prediction. We recruited 98 baseball players and collected data on glenohumeral internal/external rotation, posterior capsule thickness, supraspinatus tendon thickness, acromiohumeral distance, and occupation ratio. Players were monitored for upper extremity injuries throughout a baseball season. We evaluated the predictive accuracy of these clinical variables using five models: Glenohumeral Internal Rotation Deficit (GIRD), Logistic Regression, Random Forest, CatBoost, and Support Vector Machine. SHapley Additive exPlanation (SHAP) analysis was used to clarify each feature's role in injury prediction. During the season, 28 players experienced injuries. CatBoost (accuracy: 0.70 ± 0.05
AUC: 0.66 ± 0.05) and logistic regression (accuracy: 0.63 ± 0.07
AUC: 0.64 ± 0.08) excelled in bootstrapped evaluations and performed well in independent tests, with CatBoost maintaining an accuracy of 0.70 and an AUC of 0.62. Including GIRD had a negligible effect on CatBoost's accuracy. This integration with SHAP analyses enables a better understanding of each clinical feature's role in predicting injuries, laying the foundation for personalized injury prevention strategies. With these novel approaches, overall and individualized injury prediction can be enhanced, and future research in sports medicine can be advanced.