Construction And Analysis Of Non-Contact Injury Risk Prediction Model Based On Machine Learning

Objective:(1)A longitudinal observational study of non-contact injuries in adolescent female basketball players in Fujian Province was conducted to explore the influencing factors of non-contact injury risk in terms of demographic,training load,subjective perceived wellness,and sport performance assessment.(2)Based on the two modeling ideas of group generality and individual specificity,we used multiple machine learning algorithms for model construction and proposed a non-contact injury risk prediction model applicable to both groups and individuals.Predictive models were used to determine the relationship between important variables and the risk of non-contact injury.Method: A prospective longitudinal observational study was used as the study design to collect data on demographic data,training load,subjective perceived wellness,sports quality assessment and injury status of eighteen young female basketball players from Fujian Province during their preparation for the 2021 National Games.Generalized estimating equations were used to screen athletes’ demographic data,training load,subjective perceived wellness,athletic ability assessment for potential variables associated with non-contact injury risk.Based on both group generality and individual specificity modeling ideas,we used logistic regression,decision trees,random forests and extreme gradient boosters for model construction of factors influencing non-contact injury risk,and global attribution analysis and local attribution analysis of important variables using Shapley additivity characteristics explanation and partial dependency plots.Results:(1)A total of nine potential influential variables were screened by generalized estimating equations.These were athlete skill level(OR=5.122,p<0.05),training load monotonicity(OR=1.626,p<0.01),acute chronic workload ratio(OR=1.945,p<0.1),four-week cumulative fatigue(OR=0.441,p<0.1),four-week cumulative sleep quality(OR=0.429,p<0.05),four-week cumulative stress level(OR=0.382,p<0.01),four-week cumulative training desire(OR=0.106,p<0.1),four-week cumulative Hooper index(OR=0.721,p<0.05)and assisted jump touch height off the ground(OR=1.147,p<0.01).(2)The non-contact injury prediction model constructed with the extreme gradient boosting model had good internal validity(precision: 0.6662,recall: 0.8943,F2-score:0.8349,AUC: 0.9253)and external validity(precision: 0.6522,recall: 0.9375,F2-score:0.8621,AUC: 0.9447).The training load monotonicity,four-week cumulative fatigue,four-week cumulative stress level and assisted jump touch height off the ground were found to have high weights in the non-contact injury risk prediction model by the feature attribution analysis method.Two-dimensional interaction partial dependence plots of significant variables showed that the risk of non-contact injury was significantly higher when the training load monotonicity was increased and the four-week cumulative fatigue was decreased,the running jump touch height off the ground was increased and the four-week cumulative fatigue was decreased,the training load monotonicity was increased and the running jump touch height off the ground was increased,and the four-week cumulative fatigue and the four-week cumulative stress level value were decreased simultaneously.(3)The individualized non-contact injury risk prediction model constructed based on individual optimal aggregation time sliding window and prediction time sliding window had very good internal validity(precision: 0.8451,recall:0.9683,F2-score: 0.9091,AUC: 0.9507).The results were much better than the injury risk prediction model constructed based on the group level.Also,the important variables and threshold values of non-contact injury risk for each athlete were determined based on the individualized model.Conclusion:(1)The potential influencing factors of the risk of non-contact injury among youth female basketball players in Fujian Province were athlete skill level,training load monotonicity,acute chronic workload ratio,four-week cumulative fatigue,four-week cumulative sleep quality,four-week cumulative stress level,four-week cumulative training desire,four-week cumulative Hooper index and assisted jump touch height off the ground.The athlete’s skill level,training load monotonicity,acute chronic workload ratio,and assisted jump touch height off the ground were positively correlated with the risk of non-contact injury.Four-week cumulative fatigue,four-week cumulative sleep quality,four-week cumulative stress level,four-week cumulative training desire and four-week cumulative Hooper index were negatively associated with the risk of non-contact injury.(2)The non-contact injury risk prediction model for adolescent female basketball players constructed by applying machine learning methods had high specificity,sensitivity and interpretation;four important variables of non-contact injury risk were identified by the feature attribution analysis method.training load monotonicity,four-week cumulative fatigue,four-week cumulative stress level and assisted jump height off the ground,respectively.Two-dimensional interaction partial dependence plots of significant variables showed significant interactions between training load monotonicity,perimeter cumulative fatigue,perimeter cumulative stress level and assisted jump touch height off the ground.(3)The individualized non-contact injury risk prediction model constructed based on individual athlete data has higher specificity and accuracy in identifying individual injury risk factors,and the method has potential applications and development prospects in sports injury prevention.