Evaluation of centre‐specific machine learning models in predicting 2‐year outcomes of hip arthroscopy for mixed femoracetabular impingement syndrome

TL;DR

This study developed machine learning models to predict long-term outcomes of hip surgery for femoracetabular impingement syndrome based on preoperative factors.

Contribution

The study demonstrates that robust machine learning models can be built with limited center-specific data to predict patient outcomes after hip arthroscopy.

Findings

The random forest model showed the best performance with an AUROC of 0.99 and a C-index of 0.95.

Preoperative symptom duration, HOS-ADL, hip joint space, and alpha angle were key predictors of outcomes.

Machine learning models proved feasible for outcome prediction even with limited center-specific data.

Abstract

To construct a centre‐specific machine learning (ML) prediction model based on preoperative factors. It was hypothesised that the ML prediction model would accurately predict whether patient‐reported outcome scores (PROs) over at least 2 years would reach the minimal clinically important difference (MCID).

A retrospective analysis was performed on mixed‐type femoroacetabular impingement syndrome (FAIS) patients who had hip arthroscopy at our institution between 2016 and 2018. The primary outcome was the rate of achieving MCID in PROs assessed at least 2 years after surgery, PROs included the hip outcome score‐activities of daily living (HOS‐ADL), modified Harris Hip Score (mHHS), visual analogue scale (VAS) for pain and international hip outcome tool‐12 (iHOT‐12), assessed at a minimum of 2 years postoperatively. Preoperative patient features were selected using the least absolute shrinkage and selection operator (LASSO) algorithm. Three ML models were constructed using balanced sample data and optimal feature subsets: logistic regression (LR), support vector machine (SVM) and random forest (RF). Model performance was assessed using the area under the receiver operating characteristic curve (AUROC) and the concordance index (C‐index). Model interpretations were conducted using the SHapley Additive explanation (SHAP) method.

A total of 210 patients (48.1% female) were included. The LR, SVM, RF models had AUROC 0.76 (0.61–0.83), 0.89 (0.80–0.94), 0.99 (0.98–1.00), respectively, and C‐index 0.74 (0.65–0.82), 0.86 (0.81–0.90), 0.95 (0.93–0.96), respectively. Preoperative symptom duration, preoperative HOS‐ADL, hip joint space and preoperative alpha angle were identified as the most important predictors.

Among the three ML prediction models, RF performed best in predicting whether PROs reached MCID, demonstrating excellent discriminative ability, calibration and robustness. This indicates that individualised and robust ML prediction models for outcome prediction based on preoperative factors are feasible even with limited amounts of centre‐specific data.

Level III.

Figures