Electronic Health Record-Based Machine Learning Model for Predicting Disease Activity in Patients with Rheumatoid...

Background: The use of initial clinical assessments to predict therapeutic outcomes via machine learning (ML) is a promising frontier in precision medicine. The study aims to construct ML models capable of predicting disease activity in patients with rheumatoid arthritis (RA), thereby optimizing clinical decision-making and treatment selection. Methods: This multicenter retrospective study analyzed electronic health records (EHRs) from 1,864 patients with RA across 5 tertiary hospitals in China between 2017 and 2022. The dataset from Peking University People's Hospital (PKUPH) was employed as the training and internal validation cohort, whereas data from 4 other centers were used for external validation. Longitudinal variables, including demographics, laboratory indices, and medication regimens, at baseline, 3-month, and 6-month follow-up were integrated to capture dynamic disease patterns. Four ML models were trained to predict disease status 6 months post-treatment, with the primary outcome defined as clinical remission (disease activity score in 28 joints with erythrocyte sedimentation rate ≤ 2.6). Results: The final analysis included 1,629 patients from PKUPH and 235 from 4 other tertiary hospitals. In the internal validation phase, the optimal model achieved an accuracy of 95.3% and an area under the receiver operating characteristic curve (AUROC) of 0.971, with sensitivity, specificity, positive predictive, and negative predictive values of 98.1%, 84.2%, 96.1%, and 91.8%, respectively. The model exhibited generalizability in external validation, presenting an accuracy of 87.3% and an AUROC of 0.922. Furthermore, in the multiclass task of stratifying patients into remission, low, moderate, or high disease activity, the deep neural network model showed an accuracy of 68.6% and AUROC of 0.860. Conclusions: Longitudinal clinical data extracted from EHRs can be effectively leveraged to develop prognostic models. This study confirms that deep learning approaches trained on large-scale multicenter cohorts can accurately predict disease trajectories in RA, offering a valuable tool for personalized patient management.