Objective To explore the occurrence of lower limb training injuries in new recruits and the predictive efficacy of various risk factors by constructing an extreme gradient boosting (XGBoost) machine learning algorithm model. Methods Through cluster sampling, 279 new recruits participating in the 2024 autumn training of a certain military unit were selected as the research subjects. Personal information, physical fitness scores before enlistment, and the occurrence of lower limb training injuries of the recruits were statistically analyzed through questionnaires. The recruits were divided into a lower limb training injury group and a non-lower limb training injury group based on whether they had lower limb training injuries during the autumn training period. Univariate analysis was used to compare the personal information, physical fitness scores, and other information between the two groups. The Least Absolute Shrinkage and Selection Operator (LASSO) was used to jointly screen statistically significant variables and a prediction model based on the XGBoost machine learning algorithm was constructed. Combined with the Shapley Additive Explanation (SHAP) feature importance analysis, the weight of each feature variable on the occurrence of lower limb training injuries was quantified, and the ROC was used to test the predictive performance of the model. Results A total of 279 questionnaires were distributed in the study, and 256 valid questionnaires were retrieved, with a recovery rate of 91.76%. Among the 256 recruits, 65 cases (25.39%) suffered from lower limb training injuries. Univariate analysis and LASSO regression identified 11 factors, including BMI, place of origin, drinking history, pre-enlistment injury history, participation in pre-service training, knowledge of training injury prevention, adequate warm-up before training, adequate relaxation after training, 3000m performance, sit-up performance, and Pittsburgh Sleep Quality Index (PSQI). The XGBoost model had an area under the curve of 0.995 (95% confidence interval: 0.989-1.000), indicating good predictive performance. The importance ranking of the explanatory variables in the mode from the highest to the lowest by SHAP values was PSQI scale, BMI, knowledge of training injury prevention, sit-up performance, place of origin, adequate relaxation after training, 3000m performance, participation in pre-service training, warm-up before training, drinking history, and pre-enlistment injury history. Conclusions The predictive model for lower limb training injuries in recruits, constructed based on the XGBoost algorithm, exhibits good predictive efficacy. This model, combined with the SHAP technique, can quantify the relative weights of risk factors on the occurrence of training injuries, and priority should be given to controlling high-risk factors during training to reduce the likelihood of injuries.