A Multicenter Explainable Machine Learning Analysis of Autoimmune Disease Comorbidity in Ankylosing Spondylitis

Jichong Zhu¹Chengqian Huang²Yang Lin³Tianyou Chen⁴Lei Ren³Jiarui Chen¹Jiang Xue¹Hao Li¹Hong Cheng⁵Xinli Zhan¹Chong Liu^1*

• ¹First Affiliated Hospital, Guangxi Medical University, Nanning, China
• ²The Affiliated Hospital of Youjiang Medical University for Nationalities, Youjiang, China
• ³Guilin Peoples Hospital, Guilin, China
• ⁴The Second Affiliated Hospital of Guangxi Medical University, Nanning, China
• ⁵Nanxishan Hospital of Guangxi Zhuang Autonomous Region, Guilin, China

Background: Ankylosing spondylitis (AS) frequently coexists with other autoimmune diseases, leading to increased clinical heterogeneity and diagnostic complexity. Early identification of autoimmune comorbidity in AS remains challenging in routine practice. Methods: A multicenter, retrospective, cross-sectional study was conducted, where clinical and laboratory data were collected from three independent tertiary centers between 2012 and 2025. Patients were classified into three groups: AS alone, autoimmune diseases alone, and AS with autoimmune comorbidities. Routinely available variables, including demographic characteristics, systemic inflammatory indices, hematological parameters, and liver and renal function markers, were analyzed. Multiple machine learning algorithms were developed for two clinically relevant classification tasks: AS alone vs. AS with autoimmune comorbidities, and autoimmune diseases alone vs. AS with autoimmune comorbidities. Model performance was evaluated using AUC, calibration, decision curve analysis, and clinical impact curves. SHapley Additive exPlanations (SHAP) were applied to enhance interpretability. Results: Among all models, LightGBM consistently demonstrated superior and stable performance across discrimination, calibration, and clinical utility metrics. In distinguishing AS alone from AS with autoimmune comorbidities, key contributors included age, gender, renal function–related markers (eGFR, CysC, BUN, UA), and protein and hepatobiliary indices (ALB, DBIL). In comparisons between autoimmune diseases alone and AS with autoimmune comorbidities, SHAP highlighted metabolic-and synthesis-related features (GLOB, PREALB, CHE, ALP), acid–base balance (HCO₃), and inflammatory activity (ESR). These patterns suggest that AS-associated autoimmune comorbidity represents a distinct systemic inflammatory–metabolic phenotype rather than a simple amplification of inflammation. Conclusions: Using routinely available clinical data, an explainable machine learning framework enables accurate identification and characterization of autoimmune comorbidity in AS. This approach has practical potential for early risk stratification and clinical decision support in real-world settings.