Although artificial intelligence models of sports-related knee injuries have shown high performance in early studies, with some achieving strong discriminatory ability, most lack external validation and routine clinical use is still in the research phase.
The findings highlight both the rapid advances and current limitations of artificial intelligence in sports medicine, particularly in prediction, imaging, and recovery modeling.
Machine learning models that analyzed biomechanical, physiological, and demographic data showed variable performance in predicting injury risk in athletes. In one study of 791 female athletes, a support vector machine model achieved an average area under the curve of 0.63 in predicting anterior cruciate ligament (ACL) injuries, indicating modest predictive performance despite comprehensive inputs.
In contrast, tree-based models showed better performance under certain conditions. A random forest model predicting medial tibial stress syndrome achieved an area under the curve of 0.98 and a classification accuracy of 0.96 in a military cohort. External validation in a separate cohort showed an area under the curve of 0.95 and a precision of 0.94.
Deep learning models applied to magnetic resonance imaging achieved diagnostic accuracy ranging from 55% to nearly 100% for detection of ACL and meniscal tears. These systems can automate tear detection, localization, and grading, and have demonstrated performance comparable to professional radiologists in some settings when used as an auxiliary tool. Multimodal approaches that combine imaging and clinical data have shown improved predictive performance for certain applications, particularly for postoperative outcomes.
The machine learning model also predicted postoperative outcomes and recovery trajectory. In one study of 680 patients, a model predicting graft failure after ACL reconstruction achieved area under the curve values ranging from 0.71 to 0.85, with knee hyperextension identified as a significant predictor.
For return to sport, a random forest model of 102 athletes achieved an area under the curve of 0.952 using early postoperative functional measures such as hop test, balance score, and muscle strength metrics. Younger age, greater muscle strength, and lower BMI were associated with improved outcomes.
Across domains, most models were derived from small unicentric or highly specific cohorts, limiting generalizability. Data sets were often unbalanced, underrepresenting female athletes and non-elite athletes, and often relying on region-specific populations. Many models also lacked external validation. Much of the current evidence is based on retrospective studies, with limited prospective validation.
The researchers also highlighted interpretability challenges and technical and regulatory barriers to integration into clinical workflows, as many models function as “black box” systems without transparent inferences.
In a narrative review of studies identified through PubMed, Embase, MEDLINE, and Web of Science, researchers evaluated the application of artificial intelligence and machine learning across injury prediction, diagnosis, prognosis, and rehabilitation in sports-related knee injuries.
“AI has the potential to transform the management of sports-related knee injuries through more predictive, personalized and precise care,” lead researcher Saran Singh Gill, from the Department of Surgery and Oncology at Imperial College London, UK, and colleagues wrote.
The research team also included Nasir Karma and Chinmay Madhukar Gupte.
Chinmay Madhukar Gupte is the President-Elect of the British Association of Knee Surgeons, and the researchers reported that they had no known competing financial or personal interests in connection with this review.
Source: knee
