Poster 197. Predicting Cross-Over to Surgical Management in Patients with Nonoperatively Treated Ulnar Collateral Ligament Injuries: A Machine Learning Analysis
Objective: To develop and validate prognostic machine learning (ML) algorithms to predict progression of ulnar collateral ligament (UCL) repair or reconstruction after a failed trial of nonoperative treatment in patients who sustained UCL injuries. Methods: A total of 253 consecutive patients who presented with a UCL injury between 2017-2024 at a tertiary academic hospital were identified. The primary outcome was surgical intervention consisting of either UCL repair or reconstruction, whereas nonoperative treatment success was defined as successful return to pre-injury activities or sport. Nineteen routinely collected features were explored for predictive value. Six unique ML models were developed on the training set (n=203), internally validated on the hold-out set (n=50), and compared to a standard logistic regression model. Algorithm performance was evaluated through discrimination, calibration, and Brier score. Local interpretable model-agnostic explanations (LIME; analyses exploring dynamic variable interactions that explain individual-level risk) and partial dependency plots (describing shape of association between predictors and outcome) were constructed to provide model decision-making transparency. Results: The incidence of failure of nonoperative treatment among patients with UCL injuries was 32.8%. A combination of nine features optimized prediction accuracy: age, tear quality (sprain, partial or full), symptom onset (insidious vs. traumatic), level of sport competition, specific sport, presence of concomitant injuries, timing of injury relative to play (in-season vs. out-of-season), history of prior elbow injuries, and throwing athlete status. The best performing model used a Random Forest architecture (Figure 1; Area Under the Receiver Operator Cureve, AUROC: 0.79; calibration intercept: -0.30; calibration slope: 0.67; Brier score: 0.018) and was superior to logistic regression (AUROC: 0.33; calibration intercept: 2.29; calibration slope: -0.43; Brier score: 0.53; Table 1). Investigation of model transparency demonstrated that age <16 years, full-thickness tears, traumatic symptom onset, non-recreational competition levels, participation in specific sports (contact sports, tennis, or baseball pitchers), in-season timing of injury, presence of concomitant injuries, history of prior elbow injury, and being a throwing athlete increased the risk of requiring surgery (Figures 2-3). An interactive prognostic application was developed: https://orthoapps.shinyapps.io/UCLINJURY/ Conclusions: Among patients with UCL injuries, several prognostic factors were identified that are associated with failure of nonoperative treatment. A prognostic ML algorithm was developed and validated based on these features to provide real-time personalized assessment of surgical risk. Use of this model and knowledge of these risk factors may allow for more informed assessments of surgical candidacy in patients with UCL injuries and identify patients with a high likelihood of necessitating surgical intervention.