Early identification of difficult-to-manage axial spondyloarthritis using machine-learning decision trees
Manuel Juárez-García1,2, Chamaida Plasencia-Rodríguez3,4, Diego Benavent5,6
1Rheumatology Department, Hospital Universitario La Paz, Madrid 28046, Spain La Paz Research Institute, Madrid 28046, Spain.
Background:
Axial spondyloarthritis (axSpA) is an inflammatory disease in which, despite expanding therapeutic options, a substantial proportion of patients do not achieve the desired treatment target, highlighting the emerging concept of difficult-to-manage (D2M) axSpA.
Objectives:
To identify characteristics and predictive factors of D2M axSpA and to develop machine learning models for early identification.
Design:
Longitudinal observational cohort study with external validation.
Methods:
Patients with axSpA from the SpA-Paz cohort initiating a first biological or targeted synthetic disease-modifying antirheumatic drug (b/tsDMARD) between 2004 and 2019 were included. D2M was defined as failure of ⩾2 b/tsDMARDs, and very good responders (GR) as retention of the first b/tsDMARD ⩾3 years or discontinuation due to improvement. Baseline clinical data and baseline/6-month disease activity measures were collected. Factors associated with D2M were assessed using descriptive, comparative, and logistic regression analyses. Classification and Regression Tree (CART) models were developed and externally validated with the REGISPONSERBIO registry.
Results:
Of 311 patients initiating b/tsDMARDs, 101 were included (42 D2M, 59 GR), with a D2M prevalence of 13.5%. D2M patients were more often smokers, Human Leukocyte Antigen B27 (HLA-B27) negative, and had higher rates of enthesitis and comorbidities. Baseline Axial Spondyloarthritis Disease Activity Score (ASDAS) and Bath Ankylosing Spondylitis Disease Activity Index (BASDAI) did not differ between groups, but after 6 months D2M patients showed higher disease activity (ASDAS 2.8 vs 1.6, BASDAI 5.4 vs 3.3; both p < 0.001). Multivariable models identified ASDAS, or BASDAI plus C-reactive protein (all at 6 months), as predictors of D2M. CART models achieved areas under the receiver operating characteristic curve of 0.70 (95% confidence interval (CI) 0.46-0.93; ASDAS model) and 0.76 (95% CI 0.55-0.97; BASDAI model), with external validation confirming discrimination.
Conclusion:
D2M axSpA affects approximately 1 in 8 patients initiating advanced therapy and is associated with smoking, HLA-B27 negativity, enthesitis, comorbidities, and poor 6-month response to first-line b/tsDMARDs. CART models using routine clinical data may support early identification.