TP3.1 Moderate Predictive Ability of Machine Learning for Achievement of Minimal Clinically Important Difference for the Pain and Healthy Utility Scores after Hip Arthroscopy: Analysis From the Femoroacetabular Impingement RandomiSed Controlled Trial (FIRST) and Embedded Prospective Cohort

Abstract

Abstract Background To predict achievement of the minimal clinically important difference (MCID) for the Visual Analogue Scale (VAS) and EuroQol-5D (EQ-5D) score at six and 12 months using both logistic regression and machine learning models. Methods This analysis used data from the Femoroacetabular Impingement RandomiSed controlled Trial (FIRST) and its embedded prospective cohort. Predictive models were trained using demographic, radiographic, and intraoperative variables using a 70:30 training split. MCID was defined using a distribution-based method. Model discrimination was assessed using area under the curve (AUC), and calibration was evaluated via slope, intercept, and Brier scores. Results 309 patients (mean ± SD age 34.0±8.7 years, 37.7% female) were included. MCID for VAS and EQ-5D was 11.4 and 9.0, respectively. MCID achievement for VAS and EQ-5D was 70% and 72%, and 52.9% and 55.5% at six and 12 months, respectively. For VAS, at six months, XGBoost and logistic regression achieved AUCs of 0.653 and 0.623, respectively. At 12 months, logistic regression and LASSO achieved AUCs of 0.722 and 0.762, respectively. Key predictors for MCID achievement VAS at six or 12 months included increased baseline VAS, decreased Outerbridge grade, male sex, increased traction time, and presence of labral tears. Logistic regression and LASSO were the two strongest performing models at six and 12 months for EQ-5D, with AUCs of 0.817 and 0.745, respectively. The most important predictor of achieving the MCID for both models was lower baseline EQ-5D scores. Conclusion Logistic regression and machine learning demonstrated moderate ability to predict achievement of the MCID for VAS and EQ-5D at six and 12 months postoperatively. For VAS, XGBoost had the highest AUC at six months, while logistic regression and LASSO were superior at 12 months. Factors contributing to MCID achievement of VAS included higher preoperative VAS scores, younger age, male sex, lower Outerbridge grade, and increased traction time. For EQ-5D, logistic regression and LASSO achieved the highest predictive performance at six and 12-months. The strongest predictor of achievement of the MCID for EQ-5D was lower preoperative scores.