A preliminary investigation into operationalizing machine learning for hip replacement outcomes: predicting OHS and EQ-5D VAS in Wales

Résumé

Wales has an extensive program for collecting patient-reported outcome (PRO) measures data, and as a country with a population that generally views artificial intelligence use in healthcare positively, the development of machine learning (ML) models for patient outcome prediction offers an opportunity to advance value-based healthcare. This study discusses a preliminary case for the development of a predictive model for patients undergoing hip arthroplasty to assess challenges in creating such models in Wales, and the potential to support the training process with data from England. Five supervised ML models (logistic regression, extreme gradient boosting, elastic net, random forest, and neural net) were trained on data from England with a range of variables to predict postoperative hip replacement PROs based on predefined minimal clinically important difference thresholds for both generic (EQ-VAS) and condition-specific (Oxford Hip Score) PROs. The performance of the models on the English test set was similar across algorithms, with extreme gradient boosting achieving the highest area under the receiver operating characteristic for both PROs. Although the performance of the ML model dropped for both PROs when tested on Welsh data, the drop was more pronounced for the condition-specific Oxford Hip Score. Based on this case, four broad recommendations are presented for future ML model development for the National Health Service in Wales: standardization of the approach to missing data, preference for transparent ML methodologies, utilization of data linkage potential in Wales, and clinical agreement on metric prioritization.