Outcome Prediction Model Following Proximal Femoral Osteotomy in Legg-Calvé-Perthes Disease Using Machine Learning Algorithms

Abstract

Background: The purpose of the current study was (1) to analyze various factors that may be associated with the outcomes of Legg-Calvé-Perthes disease (LCPD), and (2) to develop and internally validate machine learning algorithms capable of providing patient-specific predictions of which patients with LCPD will achieve relevant improvement in radiologic outcomes after proximal femoral varus osteotomy (PFVO). We examined several variables, previously identified as factors, that may influence the outcome of LCPD and developed a machine learning algorithm based on them. Methods: In this retrospective study, we analyzed patients aged older than  6 years at the time of LCPD diagnosis who underwent PFVO at our institution between 1979 and 2015. Univariate and multivariate logistic regression analyses were used to examine the effects of variables on the sphericity of the femoral head at skeletal maturity, including age at onset, sex, stage at operation, extent of epiphyseal involvement and collapse, presence of specific epiphyseal, metaphyseal, and acetabular changes, and postoperative neck shaft angle (NSA). Recursive feature selection was used to identify the combination of variables from an initial pool of 13 features that optimized the model performance. Five machine learning algorithms [extreme gradient boosting (XGBoost), multilayer perception, support vector machine, elastic-net penalized logistic regression, and random forest) were trained using 5-fold cross-validation 3 times and applied to an independent testing set of patients. Results: Ninety patients with LCPD who underwent PFVO were included in this study. The mean age at diagnosis was 7.93 (range, 6.0 to 12.33) years. The average follow-up period was 10.11 (range, 5.25 to 22.92) years. A combination of 8 variables, optimized algorithm performance, and specific cutoffs were found to decrease the likelihood of achieving the 1 or 2 Stulberg classification: age at onset ≤ 8.06, lateral classification ≤ B, 12.40 < preoperative migration percentage (MP) ≤ 22.85, Catterall classification ≤ 2, 117.4 < postoperative NSA ≤ 122.90, −10.8 < postoperative MP ≤ 6.5, 139.65 < preoperative NSA ≤ 144.67, and operation at stage 1. The XGBoost model demonstrated the best performance (F1 score: 0.78; area under the curve: 0.84). Conclusions: The XGBoost machine learning algorithm achieved the best performance in predicting the postoperative radiologic outcomes in patients with LCPD who underwent PFVO. In our population, age at onset ≤ 8.06, lateral classification ≤ B, 12.40 < preoperative MP ≤ 22.85, Catterall classification ≤ 2, 117.4 < postoperative NSA ≤ 122.90, −10.8 < postoperative MP ≤ 6.5, 139.65 < preoperative NSA ≤ 144.67, and operation at an early stage had the likelihood of achieving the spherical femoral head for the patients with LCPD that underwent PFVO. After external validation, the online application of this model may enhance shared decision-making. Level of Evidence: Level III—retrospective cohort study.