PREDICTING PATIENT-REPORTED OUTCOME MEASURES FOLLOWING HIP AND KNEE ARTHROPLASTY USING SUPERVISED MACHINE LEARNING

The role of machine learning (ML) in orthopaedics is expanding, with potential applications in analysing patient data and making predictive assessments. This project utilised NHS PROMs data from primary hip and knee replacements to identify key trends and assess whether this data can reliably predict recovery rates based on patient-reported outcomes and comorbidities. The study used the NHS England national PROMs dataset for the 2018/2019 financial year, comprising 37,725 primary hip replacement entries and 43,639 primary knee replacement entries, each containing 139 variables. Key outcome measures included the Oxford Hip Score (OHS), Oxford Knee Score (OKS), EQ5D-3L, and EQ5D-Visual Analogue Scale (VAS). Python, with the scikit-learn (sklearn) library, was used to program and develop the ML models. Three supervised ML algorithms - SGD, CART, and SVMs - were tested independently. For primary hip replacements, the accuracy scores showed that SGD achieved 70%, CART 68%, and SVM 48%, while for knee replacements, SGD achieved 65%, CART 67%, and SVM 54%. The ROC-AUC analysis further indicated AUC values for hip replacements as 0.72 for SGD, 0.70 for CART, and 0.59 for SVM. While for knee replacements AUC was 0.59 for SGD, 0.68 for CART, and 0.52 for SVM. In precision-recall analysis, the average precision scores were 0.89 for hips and 0.79 for knees with SGD, 0.88 for hips and 0.84 for knees with CART, and 0.82 for hips and 0.74 for knees with SVM. Interestingly, feature importance analysis highlighted that preoperative anxiety and depression explained 28.1% of the variance in knee replacement outcomes, but 4.7% for hip replacements. Preoperative disability was the strongest predictor of hip replacement outcomes, explaining 5.8% of the variance. ML models showed higher predictive accuracy for primary hip replacements compared to knee replacements. While this study demonstrates the potential of ML for predicting patient-reported outcomes, limitations in the reliability of current national PROMs data hinder its broader application. The lack of data validation within the national dataset raises concerns about accuracy, making it difficult to fully trust the model's predictions without verification from individual organisations.