Development and Internal Validation of Machine Learning Algorithms for Predicting Subsequent Contralateral Slipped Capital Femoral Epiphysis in Patients With Unilateral Slips

Abstract

Background: Controversy remains about whether to pin the contralateral side in cases of unilateral slipped capital femoral epiphysis (SCFE). Machine learning (ML) algorithms can be leveraged to identify complex, nonlinear patterns in data and allow for more accurate predictions on which patients may need a prophylactic pin. The goal of this study was to use ML to develop a predictive model for contralateral SCFE using previously established risk factors as well as emerging radiographic measures. Methods: This retrospective study from two large centers included all patients aged <18 years treated for unilateral SCFE between 2000 and 2022 with at least 18 months’ follow-up. Those with incomplete records or bilateral SCFE were excluded. Demographic, clinical, and radiographic data were collected. Patients were divided into a unilateral pin or contralateral slip cohort. Significant relationships between groups were used for feature selection in the predictive model. Receiver operator characteristic curves were generated to determine optimal thresholds for significant continuous variables. ML models were compared using area under the curve (AUC) and weighted F1score. Results: There were 604 patients who met inclusion criteria. Of these, 389 (64%) patients needed no further intervention after unilateral pinning, 100 (17%) developed a contralateral slip, and 115 (19%) were prophylactically pinned. Compared to those who needed no further intervention, patients with a subsequent slip were significantly younger (12.2 vs 12.6 yrs, P = .038) and less covered (lateral central edge angle [LCEA]: 27.1° vs 29.3°, P = .002) with a significantly lower epiphyseal cupping ratio (ECR: 0.19 vs 0.23, P < .001). Optimal thresholds for predicting contralateral SCFE were an ECR <0.23 (AUC = 0.67), age <12.9 yrs (AUC = 0.57), and LCEA <30.2 (AUC = 0.60) with a maximum risk of 28.7% chance of SCFE when all 3 factors were present. Predictors selected for the final ML model included age, skeletal maturity, sex, LCEA, and ECR. Our final model reported an AUC of 0.66 and a weighted F1-score of 0.77. Conclusions: ECR is a significant radiographic predictor of contralateral SCFE. Our preliminary ML model shows early promise in predicting contralateral SCFE, but more data are required. Key Concepts: (1) This study is, to our knowledge, the first to apply emerging machine learning (ML) techniques to address the indications for prophylactic pinning of the unaffected side in patients presenting with unilateral slipped capital femoral epiphysis (SCFE). (2) Our preliminary ML model predicts development of a contralateral slip with reasonable accuracy. Our final model reported an area under the curve of 0.66 and a weighted F1-score of 0.77. (3) Epiphyseal cupping ratio is the most significant radiographic predictor of contralateral SCFE. Level of Evidence: III

Keywords

• Machine learning
• Slipped capital femoral epiphysis (SCFE)
• Contralateral
• Prophylactic
• Prediction