Machine Learning Applications in the Diagnosis and Prediction of Aortic Dissection: A Systematic Review and Meta‐Analysis

Background Aortic dissection is a life‐threatening condition requiring rapid and accurate diagnosis. Machine learning (ML) models have shown promise in enhancing diagnostic performance using imaging modalities, but their pooled efficacy remains unclear. This systematic review and meta‐analysis evaluates the diagnostic accuracy of ML models for detecting aortic dissection. Methods Following PRISMA guidelines, we searched PubMed, Scopus, Embase, Web of Science, and ProQuest up to March 10, 2025, identifying 775 articles. After removing duplicates, 358 articles were screened, 28 underwent full‐text review, and 21 studies were included, with 18 providing sufficient metrics for meta‐analysis. The MIDAS module in Stata was used to pool sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) for the best‐performing models, assessed via a bivariate mixed‐effects model. Risk of bias was evaluated using PROBAST + AI. Results The 21 studies included 261,477 patients (mean age 58.74 years, 66.67% male), predominantly from China (11 studies). Imaging modalities included CT angiography (10 studies) and noncontrast/contrast‐enhanced CT (five studies). The best models (18 studies, 7957 positive, 41,852 negative cases) achieved a pooled sensitivity of 0.92 (95% CI: 0.87–0.96), a specificity of 0.95 (95% CI: 0.91–0.98), and an AUROC of 0.98 (95% CI: 0.96–0.99). Deep learning models (DenseNet121, Attention U‐Net) predominated. Moderate to substantial heterogeneity in both pooled sensitivity ( I 2 = 64.85%) and specificity ( I 2 = 62.87%), likely due to diverse ML algorithms and imaging protocols, was identified. Conclusions ML models show strong diagnostic performance for detecting aortic dissection, highlighting their promise for clinical use. However, the standardization of methodologies is essential to reduce variability and promote broader clinical implementation.