Enhancing Tenosynovitis Diagnosis With XGBoost: A Machine Learning Approach Using Ultrasound Data

Rheumatoid arthritis (RA) is a systemic autoimmune disorder characterized by chronic inflammation and progressive joint destruction. Tenosynovitis is one of the early manifestations of RA. This study aims to develop a machine learning (ML)-based model using ultrasound (US) radiomics to objectively diagnosis tenosynovitis in RA patients, thereby facilitating accurate evaluation of RA.

This study included a total of 1496 grayscale US images of the wrist extensor tendons, wrist flexor tendons, and finger flexor tendons from 152 patients with RA. Radiomic features were extracted from the US images. The radiomic features and a total of 10 clinically relevant features were selected to train machine learning model. To avoid data leakage, the dataset was partitioned at the patient level with 80% allocated to the training set and 20% to the test set. The model was trained for 100 epochs with an early stopping strategy (halted if no test set performance improvement was observed for 10 consecutive epochs) to prevent overfitting. Comparative experiments were conducted against two conventional ML methods: Support Vector Machine (SVM) and Random Forest (RF). Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, accuracy, and the area under the receiver operating characteristic curve (AUC) of the model were calculated. Calibration curves and decision curve analysis (DCA) were employed to assess the model's clinical utility, and the average per-case processing time was measured to verify real-time applicability.

The dataset comprised 1196 images for training and 300 images for testing. In the training set and the test set, the AUC reached 0.969 and 0.914 respectively. The model demonstrated robust performance on the test set, with a sensitivity of 0.881, specificity of 0.788, PPV of 0.505, NPV of 0.964, F1 score of 0.642 and accuracy of 0.807. The model also outperformed SVM (AUCs: 0.909 and 0.881) and RF (AUCs: 0.955 and 0.880) on the training and test sets, respectively. The average processing time per case was 0.8 s, which meets the real-time operational requirements of clinical settings. Calibration curves indicated excellent agreement between predicted and observed outcomes, while DCA confirmed the model's clinical applicability and predictive accuracy across both training and test datasets.

The XGBoost-based machine learning model developed in this study could effectively diagnosed tenosynovitis on US imaging for RA patients, outperforming SVM and RF. Owing to its real-time processing capability, the model has potential to facilitate the accurate evaluation of RA.