How can machine learning help prevent knee injuries of experienced athletes?

INTRODUCTION: Risk prediction of anterior cruciate ligament (ACL) injuries is highly multifactorial. As machine learning (ML) algorithms have the ability to understand complex relationships in large datasets, they seem appropriate to predict at-risk athletes based on a large number of features [1]. Thus, we compared two artificial intelligence models that can help prevent knee injuries including ACLs of experienced athletes.METHODS: 50 athletes, 24 healthy and 26 with a medical history of severe knee injuries in the last 3 years took part in our experiment. They completed an informative questionnaire with physical, psychological (KOOS, SAS), and sports questions and performed a horizontal force-velocity test on a ballistic device [2]. From each push phase, we extracted a vector con-taining 60 biomechanical and psychological features (force peak, velocity, KOOS Pain score, …). Thus, we obtained a 519 vectors database. To train our machine learning algorithms to classify athletes correctly, we split the database into a train set of 405 vectors (39 athletes) and a test set of 114 vectors (11 athletes). For interpretability reasons, we chose to train nearest-neighbors (kNN) and decision tree (DT) classifiers with and without principal component analysis (PCA). Metrics used for model evaluation were accuracy and the area under the receiver operating characteristic curve (AUC).RESULTS: The test of our models on the 114 test vectors led to an accuracy of 79% with an AUC of 0.77 for the kNN, while the DT accuracy was 64% with an AUC of 0.61. With the use of a PCA before training, the results were lower for the kNN (accuracy of 72%, AUC of 0.71), but were higher for the DT (accuracy of 77%, AUC of 0.77). Finally, the first kNN classified correctly ten out of the eleven test athletes.CONCLUSION: Unlike the kNN, the DT is sensitive to small variations in data which may explain the differences in perfor-mance as our features were highly person dependent. Our best model achieved an AUC of 0.77, surpassing the Jauhi-ainen model’s [3]. The results of the PCA might be better as the size of our database would increase. Finally, the results suggested that ML could help identify at-risk athletes and be useful in predicting knee injury risk.REFERENCES[1]J. Corban et al.,Orthop J Sports Med, 2021[2]R. Macchi et al., Front-Physiol, 2021[3]S. Jauhiainen et al., Am J Sports Med, 2022