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

Comprehensive Summary

Femoracetabular impingements (FAI) are a common cause of hip pain in young adults. This can lead to osteoarthritis at a young age, and the treatment often includes arthroscopic surgery to address the labrum. A trial termed FIRST, Femoracetabular Impingement RandomiSed controlled Trial, was a randomized controlled trial comparing arthroscopic osteochondroplasty to arthroscopic lavage and looked at the outcomes between these groups. This trial used pain scores as an outcome, VAS and EQ-5D, but more recently, there has bene a shift towards using MCID, the minimally clinical important difference, as a metric rather than a pain score. This study aims to train a predictive model to interpret and predict MCID scores following the treatment of FAI using data gathered from the FIRST trial. The FIRST trial included patients aged 18-50 with a total of 220 patients, and the FIRST trial used both the VAS and EQ-5D pain scale, which both range from 0-100, 100 being the best, 0 being immense pain. 110 more patients were included that were not a part of the FIRST trial, but the same markers were identified and tested. In this study, four models were trained to predict the MCID at 6 months and 12 months post-operative. 70% of the data was used for training the models, while 30% of the data was used to test the accuracy of the models. At 6 months using the VAS pain scale, the models XGBoost and logistic regression scored the best, with an AUC of .653 and .623. At 12 months, logistic regression and LASSO performed the best with an AUC of .777 and .762. For the EQ-5D pain scale, logistic regression scored the best AUC for 6 months and LASSO at 12 months. Interestingly, different models had different important data points that lead them to their prediction, where some included Outerbridge classification and age while others included traction time and BMI. In this study, logistic regression performed similarly to the learning models tested and had a similar AUC (accuracy). From this, it can be concluded that machine learning still has progress to make, and the authors note that with more data available, machine learning tends to fair better long term. However, several important key data points resulted from this study surrounding important markers in reaching MCID, such as male sex, younger age, and increased traction time, all contributing to better overall recovery.

Outcomes and Implications

The use of machine learning in determining outcomes from FAI surgeries has not surpassed logistical regression models yet. However, the authors note that with more data points, machine learning tends to fair better, so with more time and more studies about post-operative success regarding FAI surgeries, there is an opportunity for machine learning to accurately predict outcomes better than logistical regression models. However, for this study, there were important metrics that resulted in key data points that can be used to estimate the outcome in post-operative recovery, such as younger age, male sex, increased traction time, and other preoperative scores. Using MCID as a marker is important for future context, because if machine learning can accurately determine MCID scores before surgery, it can lead to results that may change the type of surgery done, ultimately decreasing burden on patients for unsuccessful surgeries and recoveries, while lending to more favorable outcomes. This measurement as well as other preoperative measurements and scores can be combined and ultimately used to estimate the overall quality of life improvement for those undergoing surgery.