<p>Background: Over recent years, machine-learning and radiomics approaches have been increasingly applied to support the diagnosis and clinical management of ankylosing spondylitis (AS). Nevertheless, the reported diagnostic accuracy of these models remains inconsistent, largely due to variations in imaging acquisition, feature-engineering workflows, and validation strategies across studies. A comprehensive quantitative reassessment is therefore warranted to better delineate the true diagnostic and prognostic capability of such artificial-intelligence-driven methods. Purpose: This meta-analysis aimed to quantitatively determine how well radiomics and machine-learning frameworks perform in diagnosing and predicting outcomes in AS, while identifying methodological and clinical factors that may influence model accuracy and generalizability. Methods: Adhering to the PRISMA 2020 reporting standards, systematic searches were performed in PubMed, Embase, and Web of Science up to October 2025. Eligible investigations applied ML or radiomics algorithms to MRI or X-ray data for diagnostic or prognostic purposes and provided sufficient quantitative metrics (AUC, sensitivity, specificity, or 2 × 2 tables). Random-effects and bivariate meta-analyses were conducted to pool summary estimates, followed by subgroup evaluations based on imaging modality, sequence selection, algorithm class, disease phenotype, validation strategy, and cohort size. Potential publication bias was examined through funnel-plot symmetry and Egger’s regression. Results: Twenty-two studies met the inclusion standards (13 diagnostic, 4 prognostic, and 5 biomarker analyses). MRI-based models yielded a pooled AUC of 0.856 (95% CI, 0.81–0.90) with mean sensitivity 0.79 and specificity 0.85, whereas X-ray-based deep-learning networks reached AUC 0.906 (95% CI, 0.86–0.94). Dual-sequence MRI (T1WI + T2WI) offered the highest accuracy (AUC 0.912). For prognostic modeling, average AUC was approximately 0.785 with determination coefficients (R²) between 0.90 and 0.95. Observed heterogeneity mainly reflected differences in imaging modality, algorithm selection, and validation design. Conclusion: Machine-learning and radiomics approaches demonstrate strong overall diagnostic potential for AS. Dual-sequence MRI appears particularly advantageous for detecting early inflammatory activity, while radiograph-based deep-learning tools provide rapid, scalable screening options. Prognostic applications also show encouraging predictive accuracy, supporting more individualized patient management. Future work should emphasize multicenter standardization and rigorous external validation to facilitate translation into clinical practice.                                                                                                                                                                </p>

Diagnostic and prognostic performance of artificial intelligence and radiomics in ankylosing spondylitis: a systematic review and meta-analysis

期刊论文
发表日期：

2026年1月19日

查看全文

Lei Wang,
Songyang Wang,
Xuanzhe Yang,
Yan Zhao,
Feng Zhang,
Zixiang Wu,
Xiong Zhao

摘要

Background: Over recent years, machine-learning and radiomics approaches have been increasingly applied to support the diagnosis and clinical management of ankylosing spondylitis (AS). Nevertheless, the reported diagnostic accuracy of these models remains inconsistent, largely due to variations in imaging acquisition, feature-engineering workflows, and validation strategies across studies. A comprehensive quantitative reassessment is therefore warranted to better delineate the true diagnostic and prognostic capability of such artificial-intelligence-driven methods. Purpose: This meta-analysis aimed to quantitatively determine how well radiomics and machine-learning frameworks perform in diagnosing and predicting outcomes in AS, while identifying methodological and clinical factors that may influence model accuracy and generalizability. Methods: Adhering to the PRISMA 2020 reporting standards, systematic searches were performed in PubMed, Embase, and Web of Science up to October 2025. Eligible investigations applied ML or radiomics algorithms to MRI or X-ray data for diagnostic or prognostic purposes and provided sufficient quantitative metrics (AUC, sensitivity, specificity, or 2 × 2 tables). Random-effects and bivariate meta-analyses were conducted to pool summary estimates, followed by subgroup evaluations based on imaging modality, sequence selection, algorithm class, disease phenotype, validation strategy, and cohort size. Potential publication bias was examined through funnel-plot symmetry and Egger’s regression. Results: Twenty-two studies met the inclusion standards (13 diagnostic, 4 prognostic, and 5 biomarker analyses). MRI-based models yielded a pooled AUC of 0.856 (95% CI, 0.81–0.90) with mean sensitivity 0.79 and specificity 0.85, whereas X-ray-based deep-learning networks reached AUC 0.906 (95% CI, 0.86–0.94). Dual-sequence MRI (T1WI + T2WI) offered the highest accuracy (AUC 0.912). For prognostic modeling, average AUC was approximately 0.785 with determination coefficients (R²) between 0.90 and 0.95. Observed heterogeneity mainly reflected differences in imaging modality, algorithm selection, and validation design. Conclusion: Machine-learning and radiomics approaches demonstrate strong overall diagnostic potential for AS. Dual-sequence MRI appears particularly advantageous for detecting early inflammatory activity, while radiograph-based deep-learning tools provide rapid, scalable screening options. Prognostic applications also show encouraging predictive accuracy, supporting more individualized patient management. Future work should emphasize multicenter standardization and rigorous external validation to facilitate translation into clinical practice.