Non-RA screening for systemic autoimmune rheumatic diseases using a machine learning model based on physical findings and blood test data

In patients with systemic autoimmune rheumatic diseases (SARDs), particularly patients with non-rheumatoid arthritis (non-RA), the incidence of severe conditions such as interstitial pneumonia and pulmonary arterial hypertension is high. Therefore, early referral of these patients to university hospitals or other institutions capable of comprehensive medical care is desirable. However, the substantial overlap in clinical presentations across SARDs makes differentiation challenging, and objective tools to support disease classification in routine clinical practice remain limited. In this study, a machine learning based classifier was developed to distinguish Rheumatoid Arthritis (RA) from non‑RA conditions using general physical findings and routine blood test data, focusing on patients who were diagnosed with a predefined set of SARDs from a secondary‑care teaching hospital.

An observational retrospective study was conducted on 571 patients initially diagnosed with SARDs. From the hospital's clinical information database, patients' diagnoses, blood test results, and physical findings were extracted from the hospital's clinical information database to create the dataset. Physical findings were extracted as structured data using natural language processing from medical records written in a free description format by the attending physicians. For patient classification, some non-RA screening model using three classification algorithms were developed: Naive Bayes, Linear Support Vector Machines, and Random Forest.

RA and non-RA diseases share common clinical symptoms, with joint symptoms observed in over 40% of patients with the target diseases. Of the screening models, the random forest model performed the best, achieving a sensitivity of 77.3% and a specificity of 72.5%.

In this single-center retrospective cohort of patients with confirmed SARDs, a machine learning-based classifier using routinely available clinical symptoms and blood test results demonstrated promising internal performance for differentiating RA from non-RA conditions. These findings support the feasibility of machine learning-based classification within this secondary-care setting. Future studies are needed to validate the model in broader clinical populations, including primary care settings, and to evaluate its potential role in supporting referral decision-making.