Development of Machine-Learning Algorithms to Predict Attainment of Minimal Clinically Important Difference After Hip Arthroscopy for Femoroacetabular Impingement Yield Fair Performance and Limited Clinical Utility

Purpose: To determine whether machine learning (ML) techniques developed using registry data could predict which patients will achieve minimum clinically important difference (MCID) on the International Hip Outcome Tool 12 (iHOT-12) patient-reported outcome measures (PROMs) after arthroscopic management of femoroacetabular impingement syndrome (FAIS). And secondly to determine which preoperative factors contribute to the predictive power of these models.

Methods: A retrospective cohort of patients was selected from the UK's Non-Arthroplasty Hip Registry. Inclusion criteria were a diagnosis of FAIS, management via an arthroscopic procedure, and a minimum follow-up of 6 months after index surgery from August 2012 to June 2021. Exclusion criteria were for non-arthroscopic procedures and patients without FAIS. ML models were developed to predict MCID attainment. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC).

Results: In total, 1,917 patients were included. The random forest, logistic regression, neural network, support vector machine, and gradient boosting models had AUROC 0.75 (0.68-0.81), 0.69 (0.63-0.76), 0.69 (0.63-0.76), 0.70 (0.64-0.77), and 0.70 (0.64-0.77), respectively. Demographic factors and disease features did not confer a high predictive performance. Baseline PROM scores alone provided comparable predictive performance to the whole dataset models. Both EuroQoL 5-Dimension 5-Level and iHOT-12 baseline scores and iHOT-12 baseline scores alone provided AUROC of 0.74 (0.68-0.80) and 0.72 (0.65-0.78), respectively, with random forest models.

Conclusions: ML models were able to predict with fair accuracy attainment of MCID on the iHOT-12 at 6-month postoperative assessment. The most successful models used all patient variables, all baseline PROMs, and baseline iHOT-12 responses. These models are not sufficiently accurate to warrant routine use in the clinic currently.

Level of evidence: Level III, retrospective cohort design; prognostic study.