A new neural network model that detects graft ruptures and contralateral anterior cruciate ligament injuries

Abstract

AbstractPurposeThe purpose of this study was to develop a neural network model for predicting second anterior cruciate ligament (ACL) injury risk following ACL reconstruction using patient features from medical records.MethodsOf 486 consecutive patients who underwent primary unilateral ACL reconstruction, 386 patients (198 women, 188 men) with a mean age of 25.1 ± 11.6 years were included in this study. Fifty‐eight features, including demographic data, surgical, preoperative and postoperative data, were retrospectively collected from medical records, and features with an incidence of less than 5% were excluded. Finally, 14 features were used for the analysis. The multilayer perceptron was composed of four hidden layers with a rectified linear unit as activation and was trained to maximise the area under the receiver‐operating characteristic curve (auROC). Subsequently, validation was carried out through a rigorous threefold cross‐validation process. To ascertain the most efficacious combination of features with the highest auROC, a single feature with the least impact on auROC maximisation was systematically eliminated from the comprehensive variable set, ultimately resulting in the retention of a mere two variables.ResultsThe median follow‐up period was 50.5 (24–142) months. Fifty‐seven knees had a second ACL injury, with a graft rupture rate of 7.7% and a contralateral injury rate of 6.9%. The maximum auROC for predicting graft rupture was 0.81 with two features: young age and hamstring graft. Meanwhile, the maximum auROC for predicting contralateral ACL injury was 0.74 with seven features, including young age, presence of medial meniscus tear, small body mass index, hamstring graft, female sex and medial meniscus repair or treatment.ConclusionA neural network model with patient features from medical records detected graft ruptures and contralateral ACL injuries with acceptable accuracy. This model can serve as a new, useful tool in clinical practice to inform decisions about ACL reconstruction and retuning to sports postoperatively.Level of EvidenceLevel IV.