Artificial Intelligence for the Diagnosis of Transthyretin Amyloid Cardiomyopathy: A Systematic Review of Machine Learning Application.

BACKGROUND

Transthyretin amyloid cardiomyopathy (ATTR-CM) remains under-recognised, and earlier identification is increasingly important with disease-modifying therapy. Artificial intelligence (AI) applied to routine cardiovascular data may support screening, triage and diagnostic interpretation.

METHODS

MEDLINE, Embase, CINAHL, Web of Science and CENTRAL were searched from inception to August 2026 for studies evaluating AI/machine learning for ATTR-CM detection, screening or classification, including broader cardiac-amyloidosis models with separately extractable ATTR-specific performance and models detecting imaging phenotypes relevant to ATTR-CM screening. Risk of bias was assessed using an adapted QUADAS-2 framework, with AI-specific considerations informed by QUADAS-AI development work; selected TRIPOD + AI and CLAIM domains were assessed descriptively. Owing to substantial clinical and methodological heterogeneity, meta-analysis was not performed.

RESULTS

Twenty-eight primary reports were included, representing > 120,000 report-level participant entries, although overlapping cohorts precluded estimation of a unique-patient total. Performance varied by modality and task. AUCs ranged approximately 0.55-0.97 for ECG-containing models, 0.72-1.00 for echocardiography/POCUS and 0.70-0.85 for CT-based single-modality approaches; CMR differentiation of ATTR from AL achieved an AUC of 0.92. Nuclear-imaging models often showed high discrimination, including external-testing AUCs of 0.925-1.000, although several targeted tracer-uptake phenotypes rather than definitive ATTR-CM. Recurrent concerns included enriched populations, limited event counts, internal-only testing, threshold optimisation and incomplete calibration reporting.

CONCLUSIONS

AI shows promise for ATTR-CM screening and automated interpretation, but reported discrimination should not be equated with real-world clinical performance. Prospective multicentre validation, calibration at realistic prevalence, transparent reporting and workflow-level evaluation are required before routine implementation.