by Yifan Gong, Aomei Liu, Li Zhuo, Xueyuan Xu, Hongxiao Liu Objective Disease activity is a critical indicator for monitoring the progression of ankylosing spondylitis (AS), guiding clinical decision-making, and informing treatment plans. Patient-reported outcome measures (PROMs) have gained prominence in AS clinical management. However, their potential to predict Ankylosing Spondylitis Disease Activity Score-C-reactive protein (ASDAS-CRP) remains unexplored. This study employs machine learning (ML) techniques to develop prediction models utilizing PROMs data to estimate disease activity in patients with AS. Methods We utilized data from 389 patients with AS were included sourced from the China Rheumatoid Arthritis Registry of Patients with Chinese Medicine (CERTAIN) from March 2022 to March 2024. This dataset was divided into a training set (80%) and a testing set (20%). A total of 34 variables, including clinician-recorded features and PROMs (e.g., BASDAI, BASFI, BASMI, PGA, VAS, ASAS-HI, FACIT-F, DASS-21), were employed for feature selection and assessment of feature significance using a variety of machine learning methods.
by Yifan Gong, Aomei Liu, Li Zhuo, Xueyuan Xu, Hongxiao Liu Objective Disease activity is a critical indicator for monitoring the progression of ankylosing spondylitis (AS), guiding clinical decision-making, and informing treatment plans. Patient-reported outcome measures (PROMs) have gained prominence in AS clinical management. However, their potential to predict Ankylosing Spondylitis Disease Activity Score-C-reactive protein (ASDAS-CRP) remains unexplored. This study employs machine learning (ML) techniques to develop prediction models utilizing PROMs data to estimate disease activity in patients with AS. Methods We utilized data from 389 patients with AS were included sourced from the China Rheumatoid Arthritis Registry of Patients with Chinese Medicine (CERTAIN) from March 2022 to March 2024. This dataset was divided into a training set (80%) and a testing set (20%). A total of 34 variables, including clinician-recorded features and PROMs (e.g., BASDAI, BASFI, BASMI, PGA, VAS, ASAS-HI, FACIT-F, DASS-21), were employed for feature selection and assessment of feature significance using a variety of machine learning methods. Ten models were constructed using Support Vector Machine (SVM) and K-Nearest Neighbour (KNN) classifiers in conjunction with five feature selection methods: Feature Selection with Orthogonal Regression (FSOR), Trace Ratio Criterion (TRC), Robust Feature Selection (RFS), Pearson Correlation Coefficient (PCC), and ReliefF. Model performance was evaluated based on accuracy, specificity, sensitivity, and area under the receiver operating characteristic curve (AUC-ROC). Results A total of 389 patients with AS were included in the analysis. Key characteristics assessed included Patient Global Assessment (PGA), age, and the impact of disease on daily activities. The results indicated that the FSOR+SVM model achieved the best overall performance, with an AUROC of 0.930 (95%CI: 0.87–0.99) in the validation set. Meanwhile, FSOR+SVM also exhibited the highest sensitivity (83.78%), accuracy (79.35%), and specificity (90.50%). Conclusion The machine learning model developed from PROMs data proved effective for predicting AS disease activity, showing strong agreement with clinical ASDAS-CRP measures.