Machine Learning for Hip Replacement Outcomes Shows Performance Drop in Welsh Data

Machine learning models trained on English healthcare data to predict postoperative hip replacement outcomes showed a decrease in predictive performance when evaluated on patients in Wales, according to a brief research report published in Frontiers in Digital Health. The study was conducted by researchers Aura Frizzati, Michal Pruski, Kathleen Withers, and Robert Palmer, representing NHS Wales Health Education and Improvement Wales, Cardiff and Vale University Health Board, Cardiff University, the University of Manchester, and the University of Birmingham.

The investigation examined a preliminary case for developing outcome prediction models to assess implementation challenges within NHS Wales and to evaluate whether dataset sharing from England could support local model training. The authors noted that Wales maintains an extensive collection programme for patient-reported outcome measures and that the Welsh population generally holds positive views regarding artificial intelligence in healthcare, creating potential opportunities to advance value-based healthcare through predictive analytics.

To evaluate predictive feasibility, the research team trained five distinct supervised machine learning algorithms using data from England alongside a range of clinical and demographic variables. The evaluated architectures comprised logistic regression, extreme gradient boosting, elastic net, random forest, and neural network models, each tasked with forecasting postoperative patient-reported outcomes following hip arthroplasty.

Model predictions were evaluated against predefined minimal clinically important difference thresholds across two distinct patient-reported outcome measures. These instruments included the EQ-VAS, which serves as a generic health status measure, and the Oxford Hip Score, a condition-specific evaluation tool designed for hip replacement patients.

Within the English test dataset, predictive performance remained broadly comparable across all five machine learning algorithms, with the extreme gradient boosting model achieving the highest area under the receiver operating characteristic for both outcome measures. However, when the algorithms were applied to Welsh validation data, predictive performance declined for both metrics, with the authors observing a notably more pronounced reduction in performance for the condition-specific Oxford Hip Score.

In response to these preliminary findings, the authors presented four foundational recommendations for future machine learning development within the National Health Service in Wales. These proposals include standardising the handling of missing data, prioritising transparent machine learning methodologies, utilising existing Welsh data linkage infrastructure, and establishing multidisciplinary clinical agreement regarding metric prioritisation.

The brief research report, published in the Health Informatics section of Frontiers in Digital Health, examined the operational challenges of developing predictive machine learning models for hip arthroplasty patients in Wales. A key focus of the study was evaluating the potential to support the local model training process using healthcare data from England. To assess this approach, five supervised machine learning models—logistic regression, extreme gradient boosting, elastic net, random forest, and neural net algorithms—were trained on data from England with a range of variables. The algorithms were tasked with predicting postoperative hip replacement patient-reported outcomes based on predefined minimal clinically important difference thresholds across two instruments: the generic EQ-VAS and the condition-specific Oxford Hip Score.

In testing conducted on the English test dataset, the performance of the predictive models was similar across all five machine learning algorithms. Among the evaluated approaches, extreme gradient boosting achieved the highest area under the receiver operating characteristic for both the EQ-VAS and the Oxford Hip Score. However, when the machine learning models were subsequently evaluated on Welsh patient data, predictive performance dropped across both patient-reported outcomes. Notably, the authors found that this performance drop was more pronounced for the condition-specific Oxford Hip Score than for the generic EQ-VAS, highlighting differential performance reductions across distinct outcome measures when transferring predictive models between national datasets.

Based on this preliminary case study, the authors presented four broad recommendations for future machine learning model development within the National Health Service in Wales. The proposed recommendations focus on the standardisation of the approach to missing data, a preference for transparent machine learning methodologies, the utilisation of data linkage potential in Wales, and achieving clinical agreement on metric prioritisation.

Subject of Research: Predictive machine learning models for hip replacement outcomes in Wales

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

Article References: Frizzati, A., Pruski, M., Withers, K. L., & Palmer, R. I. (2026). A preliminary investigation into operationalizing machine learning for hip replacement outcomes: predicting OHS and EQ-5D VAS in Wales. Frontiers in Digital Health, 8, Article 1973774. https://doi.org/10.3389/fdgth.2026.1973774

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1973774

Keywords: machine learning, hip arthroplasty, patient-reported outcome measures, Oxford Hip Score, EQ-VAS, predictive modeling, NHS Wales, value-based healthcare, preliminary, investigation, operationalizing, machine