Machine learning-based methods in diagnosing cardiac amyloidosis: a meta-analysis

Background: -fold amyloid, often presenting as left ventricular hypertrophy. Early nonspecific symptoms lead to frequent misdiagnosis as hypertrophic cardiomyopathy, delaying care for this progressive disease. While machine learning (ML) has been applied to the diagnosis of CA, systematic evidence of its accuracy remains lacking, hindering the development of intelligent detection tools. Objectives: To explore the diagnostic accuracy of ML, providing evidence-based data to advance smart detection tools for CA. Methods: We searched the Cochrane Library, PubMed, Embase, and Web of Science up to September 25, 2025, adhering to PRISMA 2020 guidelines. Study quality was evaluated using the QUADAS-2 instrument. Subgroup analyses were stratified by disease type [light chain CA (AL-CA), transthyretin CA (ATTR-CA)] and imaging modality (echocardiography) to explore sources of heterogeneity and assess diagnostic performance across different clinical scenarios. Results: The current meta-analysis incorporated 30 studies. In validation sets, ML for overall CA showed sensitivity 0.87 [95% confidence interval (CI) 0.83-0.91], specificity 0.88 (95% CI: 0.81-0.92), positive likelihood ratio (PLR) 7.0 (95% CI: 4.4-11.4), negative likelihood ratio (NLR) 0.14 (95% CI: 0.10-0.20), and SROC AUC 0.93 (95% CI: 0.91-0.95). For AL-CA, ML demonstrated sensitivity 0.85 (95% CI: 0.76-0.91), specificity 0.82 (95% CI: 0.75-0.87), PLR 4.8 (95% CI: 3.4-6.7), NLR 0.18 (95% CI: 0.11-0.30), and SROC AUC 0.88 (95% CI: 0.85-0.91). For ATTR-CA, ML revealed sensitivity 0.84 (95% CI: 0.77-0.89), specificity 0.85 (95% CI: 0.78-0.91), PLR 5.7 (95% CI: 3.6-9.2), NLR 0.19 (95% CI: 0.12-0.28), and SROC AUC 0.91 (95% CI: 0.88-0.93). Echocardiography-only ML models showed sensitivity 0.83 (95% CI: 0.81-0.85), specificity 0.86 (95% CI: 0.82-0.89), PLR 5.9 (95% CI: 4.4-7.9), NLR 0.20 (95% CI: 0.17-0.23), and SROC AUC 0.88 (95% CI: 0.85-0.91). Conclusions: ML demonstrates favorable diagnostic accuracy for CA. Nevertheless, the aggregated findings warrant cautious interpretation owing to inherent methodological limitations in the existing evidence. Future investigations incorporating diverse cases from broader geographic regions are needed to further validate the diagnostic performance of ML for CA and to advance the subsequent development of assessment tools based on artificial intelligence. Systematic Review Registration: PROSPERO CRD42024536601.