Machine learning-based prediction of relapse in rheumatoid arthritis patients using data on ultrasound examination and blood test

• Published in: Scientific reports Vol. 12; no. 1; p. 7224
• Main Authors: Matsuo, Hidemasa, Kamada, Mayumi, Imamura, Akari, Shimizu, Madoka, Inagaki, Maiko, Tsuji, Yuko, Hashimoto, Motomu, Tanaka, Masao, Ito, Hiromu, Fujii, Yasutomo
• Format: Journal Article
• Language: English
• Published: England Nature Publishing Group 04.05.2022; Nature Publishing Group UK; Nature Portfolio
• Subjects: Arthritis, Rheumatoid - diagnostic imaging; Autoimmune diseases; Blood; Blood tests; Hematologic Tests; Humans; Learning algorithms; Logistic Models; Machine Learning; Microvasculature; Patients; Predictions; Recurrence; Remission; Rheumatoid arthritis; Ultrasonic imaging; Ultrasound; Wrist
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-022-11361-y

Recent effective therapies enable most rheumatoid arthritis (RA) patients to achieve remission; however, some patients experience relapse. We aimed to predict relapse in RA patients through machine learning (ML) using data on ultrasound (US) examination and blood test. Overall, 210 patients with RA in remission at baseline were dichotomized into remission (n = 150) and relapse (n = 60) based on the disease activity at 2-year follow-up. Three ML classifiers [Logistic Regression, Random Forest, and extreme gradient boosting (XGBoost)] and data on 73 features (14 US examination data, 54 blood test data, and five data on patient information) at baseline were used for predicting relapse. The best performance was obtained using the XGBoost classifier (area under the receiver operator characteristic curve (AUC) = 0.747), compared with Random Forest and Logistic Regression (AUC = 0.719 and 0.701, respectively). In the XGBoost classifier prediction, ten important features, including wrist/metatarsophalangeal superb microvascular imaging scores, were selected using the recursive feature elimination method. The performance was superior to that predicted by researcher-selected features, which are conventional prognostic markers. These results suggest that ML can provide an accurate prediction of relapse in RA patients, and the use of predictive algorithms may facilitate personalized treatment options.