- The paper shows that AI integration improves early detection and subtype discrimination in cardiac amyloidosis through multimodal imaging data.
- It demonstrates high diagnostic accuracy in screening, quantification, and prognostic prediction using validated AI models across modalities.
- The work highlights current limitations, including cohort biases and the need for prospective, real-world evaluations to support clinical translation.
Artificial Intelligence in Cardiac Amyloidosis: From Detection to Multimodal Clinical Integration
Introduction
Cardiac amyloidosis (CA) exemplifies the diagnostic complexity inherent to rare infiltrative cardiomyopathies. Clinical and imaging phenotypes overlap significantly with more common etiologies of heart failure and left ventricular hypertrophy (LVH), such as hypertrophic cardiomyopathy (HCM) and hypertensive heart disease. Furthermore, management decisions in CA critically depend on accurate subtype discrimination — primarily between transthyretin amyloidosis (ATTR) and immunoglobulin light-chain (AL) disease — due to their divergent pathogenesis, prognosis, and therapeutic urgency. The application of ML and deep learning (DL) to the CA diagnostic and management pathway has rapidly evolved, targeting a spectrum of modalities: ECG, echocardiography, CMR, nuclear scintigraphy/SPECT/CT, and even electronic health record (EHR)-derived clinical data. This essay delivers an in-depth synthesis of these AI applications, emphasizing their progress, limitations, and avenues for further translational development.
The Multimodal CA Diagnostic Workflow and Its Limitations
Diagnosis of CA is inherently multimodal, integrating clinical risk enrichment, laboratory monoclonal protein evaluation, ECG, echocardiography, CMR, and nuclear imaging (predominantly bone-avid scintigraphy/SPECT/CT), with endomyocardial biopsy reserved for unresolved cases. Each modality provides complementary diagnostic contributions but also introduces specific limitations regarding sensitivity, specificity, inter-site reproducibility, and subtype specificity. For example, ECG typicalities (low voltage, conduction abnormalities) are neither sufficiently sensitive nor specific, and substantial proportions of CA patients, especially with ATTR, lack classic findings. Echocardiographic features, including increased wall thickness and relative apical-sparing strain, are similarly confounded by disease mimics.
Noninvasive diagnosis of ATTR CA is feasible via SPECT/CT imaging in the context of negative monoclonal protein testing and Perugini grade 2–3 myocardial uptake [Gillmore et al., 2016; Dorbala et al., 2021], but visual grading and region-of-interest (ROI) quantification remain protocol- and reader-dependent. CMR adds further value in tissue characterization via LGE, native T1, and ECV measurements, but cannot reliably distinguish ATTR from AL without clinical and laboratory integration. These diagnostic uncertainties and workflow-dependent inefficiencies have motivated the development of AI-based support across the entire pathway.
Task-Oriented Organization of AI Applications
This review organizes AI applications by clinical task, rather than input modality, recognizing that requirements for dataset composition, reference standards, evaluation strategy, and clinical utility differ markedly between screening, detection, quantification, prognosis, and monitoring.
Screening
AI-enabled screening uses routinely available data (ECG, echo, EHR) to identify patients who would benefit from downstream confirmatory testing. Multiple studies demonstrate robust retrospective discrimination: for example, AI-ECG models have achieved AUCs up to 0.91, with some models prospectively identifying 59% of CA cases over six months prior to clinical diagnosis [Grogan et al., 2021]. However, post-development validation highlights variability across subgroups (e.g., AUC 0.66 in Hispanic patients), echoing concerns regarding generalizability and subgroup-specific calibration [Harmon et al., 2023].
AI-augmented echocardiographic screening systems — including single-clip and video-based pipelines — have reached AUCs of ~0.93, with international validation across multiple acquisition platforms and disease-mimic comparators [Slivnick et al., 2025; Duffy et al., 2025]. Importantly, workflow-based strategies, such as staged ECG+echo paradigms, demonstrate increases in positive predictive value from 33% to as high as 77% under simulated deployment conditions [Goto et al., 2021], emphasizing the utility of sequential risk enrichment in rare disease contexts.
Detection
AI detection models, deployed in settings of established suspicion (e.g., patients referred for SPECT/CT or CMR), automate recognition of CA patterns. Single-modality nuclear imaging models have achieved near-perfect discrimination (AUC ≥0.99) in large, multicenter cohorts and show extremely high specificity and sensitivity for detecting abnormal myocardial uptake [Spielvogel et al., 2024]. These models, however, should be interpreted as standardization and triage tools rather than replacements for comprehensive diagnostic adjudication, as many rely on imaging-derived rather than patient-level reference standards.
AI detection using CMR (including deep learning and radiomics) has produced AUCs up to 0.98, but these studies are constrained by smaller cohort size, mimic-rich comparator sets, and limited external validation [Martini et al., 2020; Cockrum et al., 2024]. Novel architectures (e.g., transformer-based models) offer incremental value, particularly in equivocal cases. Subtype-aware detection remains in its infancy; ML models have shown some ability to distinguish ATTR from AL using multistage CMR feature cascades, with AUCs of ~0.92 in external multicenter datasets [Weberling et al., 2025], but clinical implementation is precluded by incomplete prospective validation and the persistent necessity of laboratory confirmation.
AI detection in cardiac CT, particularly for opportunistic identification among patients scanned for other indications (e.g., TAVI planning or diffuse LV thickening), represents a developing frontier, but validation is currently limited to preliminary multicenter datasets.
Quantification
AI-driven quantification in SPECT/CT imaging represents a paradigmatic shift by establishing continuously distributed metrics — such as target-to-background ratio (TBR), volume of involvement (VOI), and composite cardiac pyrophosphate activity (CPA) — with direct correlations to clinical outcomes (hazard ratios 1.41–1.84 per SD increase for cardiovascular events) [Miller et al., 2024]. These interpretable volumetric biomarkers are amenable to longitudinal tracking, enabling quantification of serial changes associated with therapy or disease progression and outperforming visual or planar indices in analytic robustness and reproducibility.
Similarly, automated extraction of functional echocardiographic markers, such as left ventricular outflow tract velocity-time integral (LVOT-VTI), has permitted objective monitoring of physiological progression, with a ≥5% decrease over 12 months independently predicting all-cause mortality (HR 1.41) [Venneri et al., 2025]. However, prospective outcome studies linking these markers directly to intervention thresholds and treatment decision-making remain pending.
Prognostic Risk Prediction and Longitudinal Monitoring
Dedicated AI-based prognostic models, particularly those leveraging CMR texture and whole-heart LGE analysis, now outperform established clinical staging systems for AL-CA in internal validation (C-index 0.91 vs. 0.65 for Mayo staging; AUC 0.95 vs. 0.71 for 2.6-year survival) [Wang et al., 2025]. Diagnostic model outputs (e.g., AI-ECG scores) frequently encode disease severity and portend worse outcomes across subtypes [Amadio et al., 2025]. Longitudinal, serial quantification via SPECT/CT has demonstrated treatment-associated decreases in tracer-burden metrics and correlation with biomarker and event reduction [Spielvogel et al., 2025; Miller et al., 2026].
Despite these advances, clinical actionability is limited by the largely retrospective evidence base, potential for cohort and ascertainment bias, and a lack of harmonized response criteria.
Limitations of Current Evidence and Methodological Challenges
A central theme across all reviewed AI applications is the current maturity gradient: nuclear imaging-based CA detection and disease-burden quantification are closest to clinical translation, while subtype-aware discrimination, individualized prognosis, and treatment-response monitoring remain at formative stages. Critical obstacles include:
- Cohort design limitations: Many studies are limited by small sample sizes, enrichment for confirmed CA at advanced stage, failure to include relevant phenotypic mimics in comparator groups, and inadequate external validation.
- Lack of robust reference standards: Circularity in ground truth, particularly when AI is validated against consensus imaging reads rather than multi-disciplinary adjudication with laboratory and genetic confirmation.
- Uncertain calibration and real-world performance: AUC and discrimination alone are insufficient. Calibration, sensitivity/specificity at clinical thresholds, PPV/NPV at deployment prevalence, and robust assessment in demographically diverse and non-academic populations are required.
- Workflow integration and regulatory requirements: Meaningful utility necessitates embedding AI tools within real clinical pathways, ensuring interpretability, reproducibility, and clearly defined action points, particularly for screening and monitoring. Regulatory approval must be paired with post-market performance surveillance and bias assessment.
Implications, Roadmap, and Future Directions
AI has delivered substantial progress in standardizing image interpretation, enabling volumetric quantification, and facilitating scalable suspicion-raising tools. The next translation leap will be contingent on:
- Development of multicenter, mimic-rich, longitudinal datasets with subtype- and outcome-linked labels. Multimodal platforms that directly integrate imaging, laboratory, clinical, and genomic data streams will be necessary for robust subtype assignment, risk stratification, and personalized monitoring.
- External validation in diverse real-world populations. Projects such as multicenter collaborative benchmarks for SPECT/CT and CMR-based quantification, and prospective implementation studies in screening and longitudinal follow-up populations, should be prioritized.
- Refinement of clinically meaningful endpoints and thresholds for escalation or therapeutic modification. This includes the definition of actionable change in AI-derived biomarker trajectories, with correlation to clinical status and management outcomes.
- Regulatory and workflow integration aligned with trustworthy AI principles. Prospective clinical trials, continuous post-deployment performance monitoring, and explicit subgroup (age, sex, ethnicity, comorbidities) analysis are prerequisites for equitable implementation.
The ultimate value of AI in CA will depend on synergistic integration with established diagnostic algorithms, enhancing — rather than supplanting — the rigor and nuance of expert-driven clinical decision-making.
Conclusion
AI has rapidly advanced the diagnostic and management landscape for cardiac amyloidosis, particularly in suspicion-raising, standardized detection, and interpretable disease-burden quantification. However, strong claims regarding subtype-aware classification, prognostic risk prediction, and treatment-response monitoring must await further evidence derived from large, multicenter, externally validated, and clinically integrated studies. The field’s next phase will be defined by the translation of AI from retrospective benchmarking to prospective, workflow-embedded, and outcome-driven tools that support the evolving paradigm of precision medicine in cardiac amyloidosis.