AI-BASED CLASSIFICATION OF SPONDYLOARTHRITIDES USING THE TURKISH PATIENTS OF ASAS PERSPA DATASET: INSIGHTS ON CLINICAL FEATURES AND PATIENT OUTCOMES

Abstract

Background: Spondyloarthritides (SpA), encompassing axial spondyloarthritis, peripheral spondyloarthritis, and psoriatic arthritis, are characterized by diagnostic complexities arising from overlapping symptoms and heterogeneous clinical presentations. Despite advancements in imaging and biomarker research, diagnostic delays persist, significantly impacting disease management and patient outcomes [1, 2]. Several studies [3–5] have demonstrated the potential of artificial intelligence (AI) to assist diagnosis in rheumatology by analyzing large-scale clinical datasets to identify patterns and correlations that may not be apparent to human observers. For spondyloarthritides, AI has the potential to reduce diagnostic delays and improve classification accuracy by leveraging comprehensive datasets such as the dataset gathered by Assessment of SpondyloArthritis international Society (ASAS). Objectives: This study aims to develop an AI-based classifier to differentiate axial SpA, peripheral SpA and psoriatic arthritis using the ASAS perSpA dataset. Methods: ASAS perSpA dataset is gathered from 24 countries with 489 patients from Türkiye, being scattered across country. We employed a variety of machine learning models, including Random Forest, XGBoost, CatBoost, LightGBM, Support Vector Machines (SVM), and Neural Networks, to train a classifier model using a subset of about 740 clinical variables from the ASAS perSpA dataset. Patients are divided into two groups: the training group, being 80% (n=391) of the patients were used to train the classifier model, and the second group, the assessment group being 20% (n=98) were presented to the previously trained AI-model to be categorized into one of three SpA conditions: Axial SpA, Peripheral SpA and Psoriatic Arthritis. Model performance was assessed by correct predictions in comparison to the original dataset. Additionally, feature selection identified the top 10 variables contributing to the models' predictions. Results: Among the patients included in the study, 256 were male, and 233 were female. The average age for males was 40.24 ± 11.27 years, while for females, it was 43.61 ± 11.01 years. In axial SpA (n=391), 230 were male and 161 were female, whereas in peripheral SpA (n=96), 43 were male and 53 were female, and in psoriatic arthritis (n=80), 19 were male and 61 were female. Since the selection of patients for the training and assessment groups was conducted randomly, there were no significant differences in the demographic characteristics between the two groups. The dataset performances were as follows: Random Forest (68.42%), XGBoost (75.79%), LightGBM (81.05%), CatBoost (75.79%), SVM (82.11%), and Neural Networks (78.52%). SVM achieved the highest accuracy, followed closely by LightGBM. Table 1 highlights the top features for each SpA category. The correlation of peripheral SpA with quality-of-life scores (BASDAI, BASFI, and ASAS-HI) suggests this category has the greatest impact on patient outcomes. The absence of HLA-B27 across all categories indicates it is not a significant diagnostic factor for SpA. Additionally, the features selected by the AI model for axial SpA suggest that its diagnosis relies primarily on clinical findings rather than laboratory markers. Conclusion: This study demonstrates the feasibility of using AI to enhance SpA diagnosis with large datasets like ASAS perSpA. The model emphasizes feature selection to simplify complex data. Future work will extend analysis to data from other countries and refine the model for improved accuracy. REFERENCES: [1] Russell, M.D., et al., Diagnostic delay is common for patients with axial spondyloarthritis: results from the National Early Inflammatory Arthritis Audit. Rheumatology, 2022. 61(2): p. 734-742. [2] Seo, M.R., et al., Delayed diagnosis is linked to worse outcomes and unfavourable treatment responses in patients with axial spondyloarthritis. Clinical rheumatology, 2015. 34: p. 1397-1405. [3] McMaster, C., et al., Artificial intelligence and deep learning for rheumatologists. Arthritis & Rheumatology, 2022. 74(12): p. 1893-1905. [4] Redeker, I., et al., Identification of a machine learning-based diagnostic model for axial spondyloarthritis in rheumatological routine care using a random forest approach. RMD open, 2024. 10(4): p. e004702. [5] Sequí-Sabater, J.M. and D. Benavent, Artificial intelligence in rheumatology research: what is it good for? RMD open, 2025. 11(1): p. e004309. Table 1. Top features selected by AI

Axial SpA Peripheral SpA Psoriatic Arthritis • Loss of lumbar lordosis • Axial involvement history • Spinal vertebrotomy • Occiput-wall distance > 0 cm • Bamboo spine • Thoracic kyphosis • Lower limbs predominantly • Back pain • Sacroiliac findings • Back pain onset age < 45 • ASAS Criteria for Axial SpA • Past/present arthritis compatible with SpA • Peripheral joint disease history • Arthritis history • Last CRP value • Diagnosis Age • BASDAI score • BASFI score • Psoriasis diagnosis • Total swollen joints • Total joint tenderness • ASAS-HI score • Total tender joints • Psoriasis diagnosis • Psoriasis history • CASPAR score • Current psoriatic nail dystrophy • Dactylitis diagnosis • Radiographic Destructive arthropathy of the distal interphalangeal joints • Radiographic evidence of juxta-articular new bone formation appearing as illdefined ossification near joint margins on plain radiographies of the hand or foot • Age • Past/present arthritis compatible with SpA • Last CRP value Acknowledgements: NIL. Disclosure of Interests: None declared.