Artificial intelligence versus clinicians for the diagnosis of cardiac amyloidosis: a systematic review and meta-analysis
A Agrawal, J Sethuraj, R Elangovan, K Patel, A Chauhan, K Elangovan, S Garudas, P Satish, E KrishnanAbstract
Background
Cardiac amyloidosis (CA) is a progressive infiltrative cardiomyopathy with high morbidity, mortality, and non-specific early symptoms leading to late diagnosis and limited treatment options. Early, accurate detection can improve outcomes. Artificial Intelligence(AI) holds promise for enhancing diagnostic accuracy by detecting subtle features of CA that may evade human detection, but its performance relative to clinicians remains unclear. This systematic review and meta-analysis critically appraises and quantitatively synthesizes current evidence comparing AI and clinician performance in diagnosing CA, informing its evolving role in cardiovascular practice.
Purpose
To provide the first head-to-head comparison of the accuracy of artificial intelligence algorithms versus clinicians for the diagnosis of cardiac amyloidosis across non-invasive diagnostic modalities.
Methods
We systematically searched Ovid MEDLINE, PubMed, Cochrane CENTRAL, Embase, and Web of Science from inception to June 2025 for studies directly comparing the diagnostic performance of AI algorithms and clinicians for CA. Eligible studies included those reporting accuracy and other metrics for both AI and clinician assessment using any diagnostic modality. Three reviewers independently screened titles, abstracts, and full texts, extracted data, and assessed risk of bias using the QUADAS-2 tool, resolving discrepancies by consensus. Random-effects meta-analyses and leave-one-out sensitivity analysis were performed in R, following PRISMA-DTA guidelines, to synthesize pooled accuracy estimates for AI and clinicians.
Results
Across studies assessing AI for diagnosing CA, pooled analysis showed a summary accuracy of 0.89(95% CI,0.81–0.94), with high heterogeneity(I2=85.2%; Fig 2). In contrast, clinician-based diagnostic performance yielded a pooled accuracy of 0.86(95% CI,0.73–0.93), with high heterogeneity(I2=91.5%; Fig 3). In 3 studies excluded from analysis due to limited data, AI displayed comparable diagnostic performance to clinicians. Leave-one-out sensitivity analysis demonstrated that no single study significantly altered the overall pooled diagnostic accuracy estimate.
Discussion
This synthesis underscores the comparable diagnostic performance of artificial intelligence and physicians in evaluating CA. AI can improve clinical workflow consistency and efficiency, potentially reducing interobserver variability. However, integration into routine care needs attention to algorithmic transparency, external validity, bias mitigation, and data stewardship. Despite high heterogeneity, these findings provide a robust evidentiary basis for considering AI as an adjunct to expert judgment in CA diagnosis.PRISMA flow diagramForest plots of outcomes