engleski

FORECASTING INJURY AMONG ATHLETIC AND NON-ATHLETIC YOUTH: USAGE OF THE ARTIFICIAL INTELLIGENCE METHODS

Introduction: This study aimed to predict injuries in one year with application of machine learning (ML) methodology using various risk factors in a representative sample of adolescents. Methods: This research is a part of the CRO-PALS study conducted in a representative sample of urban youth in Zagreb (Croatia). Analyses for this study are based on 558 adolescents from the CRO-PALS cohort (age:16-17 years). Risk factors that were measured at baseline included: sex, age, body mass index, body fat percentage, moderate-to-vigorous physical activity (MVPA), training hours per week (only for athletic participants), socioeconomic status (SES) and functional movement operationalized through a total score of the Functional Movement Screen test. The details on study protocols were described elsewhere (Štefan et al., 2018). Data on injury occurrence were collected with a computerized self-reported questionnaire one year after the baseline measurements. ML analyses, as a form of interfacial intelligence, were conducted separately for athletic (n=193) and non-athletic (n=365) group of children. The model with the highest value of the area under the ROC curve (AUC) was selected as an estimate of the best predictive accuracy. In addition, values of sensitivity, specificity, and OR (95% CI) are provided. Results: Among athletic and non-athletic participants injury incidence during a 1-yr period was 25.4% and 11.2%, respectively. Within athletic and non-athletic participants, k-nearest neighbors (kNN) showed the highest value of AUC (0.64 and 0.61), and was thus chosen as the model with the highest predictive accuracy. Within athletic youth, kNN-150 showed sensitivity of 0.52, specificity of 0.76, and OR (CI 95%) of 3.42 (1.73-6.76). Among non- athletic participants, kNN-100 exhibited sensitivity of 0.83, specificity of 0.45, and OR (CI 95%) of 3.93(1.69-9.14). Conclusion: The results of this study suggest that with given predictors: sex, age, BMI, body fat percentage, MVPA, training hours per week, SES, and functional movement, kNN exhibited the best accuracy over other models. However, in the context of prediction accuracy, kNN method exhibited poor predictive accuracy for injury incidence among both athletic and non-athletic participants. Future studies should try to include more predictors in ML models to improve the accuracy of injury prediction among adolescents.

puberty, AI, movement patterns, musculoskeletal conditions, musculoskeletal injury, sport trauma, adolescence

nije evidentirano