To explore the machine learning model of chronic pain after total knee arthroplasty (TKA) and analyze its influencing factors.

278 patients with knee osteoarthritis who received TKA in our hospital from October 2021 to May 2024 were selected and randomly divided into training set (195 cases) and test set (83 cases) according to the ratio of 7:3. The pain Numerical Rating Scale (NRS) was used to evaluate the pain before surgery and at discharge. The anxiety and depression Scale (HADS) was used to evaluate the emotional status of the patients. The optimal hyperparameters of the training set were obtained by using the 5-fold cross-validation method, and logistic regression (LR), random forest (RF), extreme gradient Lift (XGBoost) and support vector machine (SVM) models were constructed. Validation sets are used for internal validation. Accuracy, sensitivity, specificity, recall rate, F1 value and area under the curve (AUC) were used to evaluate the predictive performance of the models, and the models with the best performance were compared. Shapley additive interpretation (SHAP) algorithm was used to evaluate important risk variables, and the influence of clinical characteristics on the models was analyzed.

Among 278 patients with knee osteoarthritis treated with TKA, 87 case had chronic pain after surgery (31.29%). Compared with the pain-free group, the patients in the training intensive pain group with diabetes mellitus, the proportion of postoperative thrombosis in lower extremity vein, preoperative HADS score, preoperative NRS score and postoperative CRP level were significantly higher (P<0.05). There were no significant differences between the two groups in terms of age, gender, body mass index (BMI), comorbidities of hypertension, coronary heart disease, disease duration, ASA grading, preoperative knee joint tenderness, surgical site, surgical time, length of hospital stay, and NRS score at discharge (P>0.05). The comparison showed that the AUC of LR, RF, SVM and XGBoost models were 0.725, 0.945, 0.780 and 0.884, respectively, all greater than 0.7, and the AUC and accuracy of RF model were the largest, suggesting that the prediction performance of RF model was better than the other three models. In the optimal model, SHAP algorithm was used to find that postoperative CRP, preoperative HADS, preoperative NRS score, postoperative venous thrombosis of lower extremity, and diabetes mellitus were important factors affecting chronic pain after TKA.

All prediction models of chronic pain after TKA based on machine learning algorithm show good prediction performance, among which RF model has the best comprehensive prediction efficiency, and its risk factors have important guiding significance for clinical prevention and treatment of chronic pain.