Summary

The study demonstrates that the ML-powered ArthroPredict platform can accurately predict the risk of total joint replacement (TJR) in osteoarthritis patients using routine clinical data, with Random Forest, LightGBM, and XGBoost models showing the most reliable performance, offering the potential to assist clinicians in early identification and personalized intervention for high-risk patients

Abstract

Introduction

Osteoarthritis (OA) is the most common degenerative joint disease, significantly impacting physical function, quality of life, and posing a substantial global socioeconomic burden. As OA prevalence increases due to aging populations and obesity, the need for total joint replacement (TJR) is projected to rise sharply, particularly in the UK, where surgeries are expected to grow from 70,000 to 119,000 annually by 2035. Despite the effectiveness of TJR, accurately predicting which patients will progress to needing joint replacement within a specific timeframe remains a challenge in clinical practice. Addressing this need, • ArthroPredict is a novel machine learning (ML)-powered platform designed to predict the likelihood of TJR within five years of diagnosis using baseline clinical data.

Methods

This study utilized the Osteoarthritis Initiative (OAI) dataset, consisting of 4,796 patients with OA who were followed for at least five years. A total of 45 features were selected based on their clinical relevance, including demographics, medication history, comorbidities, functional scores, physical status, and responses from the 12-Item Short Form Survey (SF-12). The primary outcome was the prediction of TJR, specifically targeting the hip and knee within a five-year period. The dataset was split into training and test sets using an 80-20% split. A total of nine machine learning models were developed and evaluated for their ability to predict the 5-year likelihood of TJR in patients with osteoarthritis using baseline clinical data. Model performance was assessed using precision and area under the receiver operating characteristic curve (AUC).

Results

The rate of 5-year TJR was 12.8%. Overall, Random Forest (RF), XGBoost, and LightGBM demonstrated the most reliable predictive performance for forecasting 5-year TJR in OA patients. The RF model achieved an accuracy of 82.2%, with a predicted TJR likelihood of 27.1%, yielding an F1 score of 21.2% and an AUC of 0.642. The XGBoost model similarly performed well, with an accuracy of 84.1% and an AUC of 0.648. LightGBM showed comparable performance with an accuracy of 84.4% and an AUC of 0.658. In contrast, the NGBoost model predicted a notably higher likelihood of TJR at 21.5%, though its overall accuracy (74.7%) was slightly lower than the leading models. NGBoost achieved an AUC of 0.629. While NGBoost demonstrated higher recall than other models (31.8%), its overall performance was limited by lower precision and AUC. The Logistic Regression and Naive Bayes models exhibited the lowest predictive accuracies of 63.2% and 66.2%, respectively, with AUC scores of 0.663 and 0.669, respectively. Lastly, the Decision Tree model achieved an accuracy of 75.2%, with an AUC of 0.563.

Conclusion

Our results demonstrated that the ML-powered ArthroPredict tool can accurately predict the risk of TJR in OA patients using routine clinical data. The models, particularly RF LightGBM, and XGBoost, achieved clinically acceptable levels of accuracy, showing potential to assist clinicians in identifying high-risk patients early in their disease course. By early identification of high-risk patients using data from routine clinical practice, ArthroPredict enables timely, personalized interventions, potentially delaying surgery and alleviating the burden on healthcare systems.