Identification of markers predicting clinical course in patients with Behcet disease by combination of machine learning and unbiased clustering analysis

Behçet's disease (BD) is a multisystem inflammatory disorder with diverse clinical manifestations. Identifying biomarkers predictive of clinical outcomes, such as tumor necrosis factor (TNF) inhibitor initiation and ocular inflammatory attack frequency, is critical for improving management. This study aimed to identify biomarkers predicting the clinical course of BD using peripheral blood test data and unbiased clustering combined with machine learning.

A retrospective cohort study of 238 BD patients diagnosed at Tokyo Medical University Hospital (2004-2020) was conducted. Unsupervised hierarchical clustering was applied to peripheral blood data, dividing patients into distinct groups. Machine learning techniques were used to explore biomarkers predicting the clinical course.

Cluster analysis identified four groups: Group A (low C-reactive protein), Group B (high angiotensin-converting enzyme), Group C (high anti-streptolysin O), and Group D (low neutrophil count). Group C had a higher rate of TNF inhibitor initiation (47%, p = 0.04), while Group D had fewer ocular inflammation attacks per year (1.4, p = 0.04). Logistic regression analysis identified red blood cell count (p < 0.01) and monocyte percentage (p = 0.02) as predictive biomarkers for TNF inhibitor initiation. Machine learning further confirmed mean corpuscular hemoglobin concentration (MCHC) as a significant predictor of TNF inhibitor initiation. Additionally, multiple regression analysis identified the neutrophil/lymphocyte ratio as a predictor of the number of inflammatory attacks per year (p = 0.02).

Unsupervised clustering of blood test data identified distinct BD clinical phenotypes. Monocyte percentage may predict TNF inhibitor initiation, while neutrophil/lymphocyte ratio may predict ocular inflammation frequency, highlighting pathophysiologic heterogeneity in BD.