gms | German Medical Science

Objectives: Among patient-reported outcome measures (PROMS), health-related quality of life (HRQoL) is increasingly used to gauge benefits of a surgical procedure. Crucially, the ability to predict change in a patient’s HRQoL before the procedure takes place could enable a reduction in unnecessary and fruitless procedures. This approach is particularly well suited for highly standardized and common procedures such as hip arthroplasty in combination with established measurements of HRQoL such as EQ-5D.Specifically, we consider using machine learning to predict HRQoL improvement one year after total hip arthroplasty (THA).

Methods: Data from 2,566 consecutive total hip arthroplasty operations performed at Clinic Volmarstein from May 13, 2015, to Jan 18th, 2021, was collected and combined with PROMS (including EQ-5D, OXHIP, HOOSPS) from patient surveys. The data included patient data (e.g., age, gender, ASA, diagnosis) as well as information on the surgical procedures (operation time, implants, surgeon etc.). HRQoL was measured via the German index score for the EQ-5D-3L framework of patient reported outcomes.

Machine learning was used in the context of a binary classification: to predict sufficient improvement (defined via minimal clinically important difference, MCID) of the EQ-5D-3L index pre-surgery to 12-month post-procedure.

Three different classification algorithms were tested: xgboost, neural networks, and logistic regression with elastic net penalty. Both permutation importance as well as SHAP values were employed to garner insights on important features for model performance.

Results and conclusion: In total, 994 operations had sufficient data for the algorithms used, 880 of which showed improvement above the MCID.

All tested algorithms showed similar levels of performance, measured by ROC AUC (performed on test data) between 0.6 and 0.7. Among important variables, preoperative EQ-5D index value and patient age had the most influence on the models.

By using modern data analytics tools such as machine learning, it is possible to process large volumes of data in medicine. This can improve the concept of “evidence-based medicine” and predict critical factors for the patient with greater certainty.

Particularly in the case of frequently performed standard interventions such as THA (with copious amounts of data available), this enables more valid preoperative assessment as well as targeted postoperative support for “at-risk” patients.

These promising approaches can be used in the future to complement medical benefit assessments prior to planned hip replacement surgery.

When used at the time of discharge, targeted support or monitoring of patients can take place to enable the best possible treatment outcome for the patient.

Additional and larger data bases are desirable to improve the performance and generalizability of the algorithms.