Machine learning predicts clinical response to platelet-rich plasma therapy in knee osteoarthritis

The purpose of the study is to predict clinical responses to platelet-rich plasma (PRP) therapy in patients with knee osteoarthritis (KOA) before treatment and to identify the crucial efficacy predictors.

We reviewed the multidimensional clinical data of 102 KOA patients who underwent PRP therapy with a retrospective approach, including anthropometrics and blood indices. We defined the response to treatment as a reduction in the Numerical Rating Scale (NRS) pain score of ≥2 at the 6-month follow-up. After comprehensive data preprocessing, including imputation, Yeo-Johnson transformation, standardization, and balancing via SMOTETomek, 33 clinical features were retained. We then created and verified multiple machine learning models using 10-fold cross-validation (CV) on a 70% training cohort. The most effective model was subsequently validated on the 30% test cohort, and feature contributions were interpreted using SHapley Additive exPlanations (SHAP) analysis.

The Gradient Boosting Classifier showed the best overall performance among all tested models. The final model achieved a prediction accuracy of 90.3%, an area under the curve (AUC) of 0.862, and a Cohen's kappa coefficient of 0.611 on the test set, indicating high predictive consistency. The SHAP interpretability analysis revealed three biomarkers-osmotic pressure, lipoprotein(a) [Lp(a)], and uric acid-as the clinical factors most strongly associated with treatment response.

Clinical character-based machine learning models are effective in predicting PRP therapy outcomes prior to treatment. Such results provide a strong basis for introducing individualized therapeutic modalities in the management of osteoarthritis.