Artificial Intelligence for the Diagnosis of Transthyretin Amyloid Cardiomyopathy: A Systematic Review of Machine Learning Application Authors: Eyad Jamileh, Ahmed T. Elmewafy, Muhammad Abdullah Qadeer, Zuhaib Zulfiqar, Kiran Kang, Kaveh Hosseini, Ibrahim Antoun Published: 2026 ABSTRACT BackgroundTransthyretin 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.MethodsMEDLINE, 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.ResultsTwenty‐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.ConclusionsAI 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. 1. Introduction Transthyretin amyloid cardiomyopathy (ATTR‐CM) is caused by myocardial deposition of misfolded transthyretin and commonly presents with heart failure, increased ventricular wall thickness, arrhythmia or conduction disease. Recognition is frequently delayed because its phenotype overlaps with common cardiovascular disorders and because confirmatory testing is often undertaken only after clinical suspicion has been raised [1–3]. The importance of earlier diagnosis has increased with the availability of disease‐modifying therapy, which has shifted the clinical objective from recognition of advanced disease toward identification of patients earlier in the disease course [4]. Contemporary ATTR‐CM evaluation is multimodal. Electrocardiography and echocardiography provide widely available phenotypic clues; CMR can characterise myocardial infiltration and bone‐tracer scintigraphy can establish a nonbiopsy diagnosis of ATTR‐CM when grade 2 or 3 myocardial uptake is present in the appropriate clinical context and a monoclonal protein disorder has been excluded [2, 3]. CT and PET have additional or emerging roles in selected populations. Each modality, however, depends on the recognition of sometimes subtle patterns and is affected by referral pathways, disease prevalence, image quality and operator expertise. AI and machine‐learning methods offer a potential way to identify multivariable patterns that may be difficult to recognise consistently by conventional interpretation. Previous reviews have considered AI in cardiac amyloidosis broadly, and modality‐specific systematic reviews and meta‐analyses have evaluated AI‐enhanced ECG models [5–11]. The rapidly expanding literature now includes large‐scale external validation, multimodal models, opportunistic POCUS, radiomics, synthetic‐data approaches and prospective implementation. The present review therefore focuses specifically on ATTR‐CM‐relevant diagnostic evidence across the full diagnostic pathway and emphasises methodological quality, validation design, prevalence, calibration, model transparency and cohort overlap rather than headline accuracy alone. The objective was to systematically evaluate AI and machine‐learning models used for detection, screening, classification or clinically relevant imaging phenotypes of ATTR‐CM; to compare performance across modalities and stages of validation and to identify methodological requirements for clinical translation. 2. Methods 2.1. Protocol and Reporting This systematic review was conducted in accordance with PRISMA 2020 [12] and was prospectively registered in PROSPERO (CRD420251230839) [13]. The complete PRISMA Checklist is provided in Supporting File 1. Risk of bias was assessed using an adapted QUADAS‐2 framework [14], with AI‐specific considerations informed by the QUADAS‐AI development framework [15]. AI‐specific reporting and validation characteristics were additionally mapped against selected domains from TRIPOD + AI [16] and, for imaging studies, CLAIM 2024 [17]. These reporting frameworks were used descriptively and were not converted into numerical quality scores. 2.2. Eligibility Criteria Original peer‐reviewed studies were eligible if they evaluated an AI or machine‐learning model for ATTR‐CM detection, screening, classification, early prediction or diagnostic triage using ECG, echocardiography/POCUS, CT, CMR, PET, bone scintigraphy/SPECT/SPECT‐CT or a multimodal combination containing at least one of these diagnostic data sources. Broader cardiac‐amyloidosis models were eligible when ATTR‐specific diagnostic performance was separately extractable. Studies that automated imaging phenotypes directly used in ATTR‐CM screening pathways, such as high‐grade cardiac bone‐tracer uptake, were eligible even when the model endpoint was the imaging phenotype rather than final etiologic diagnosis, provided this distinction was explicitly retained in the synthesis. Studies restricted to AL amyloidosis, treatment monitoring without a diagnostic outcome, noncardiac ATTR manifestations, pure administrative‐claims/EHR case‐finding without a prespecified ECG or imaging component, reviews, editorials, conference abstracts without adequate data and reports without extractable quantitative diagnostic performance were excluded. Goto et al. [18] was included because ATTR‐specific ECG and echocardiographic diagnostic performance was separately extractable. Duffy et al. [19] was excluded because cardiac amyloidosis was modelled as a composite aetiology of left ventricular hypertrophy without separately extractable ATTR‐CM diagnostic performance. The later international EchoNet‐LVH evaluation by Duffy et al. [20] was considered supportive evidence rather than an eligible ATTR‐CM diagnostic study because its primary diagnostic metrics were reported for cardiac amyloidosis overall and subtype‐specific diagnostic performance was not separately quantified. 2.3. Search Strategy A comprehensive literature search was conducted in MEDLINE (via PubMed), Embase, CINAHL, Web of Science Core Collection and the Cochrane Central Register of Controlled Trials (CENTRAL) from database inception to August 2026. The search combined controlled vocabulary and free‐text terms relating to three principal concepts: (1) cardiac and transthyretin amyloidosis; (2) artificial intelligence, machine learning, deep learning, neural networks and radiomics and (3) relevant cardiovascular diagnostic modalities, including electrocardiography, echocardiography, ultrasound/POCUS, CT/CCTA, MRI/CMR, PET/PET‐CT, bone scintigraphy and SPECT/SPECT‐CT. Tracer‐specific terms included DPD, PYP, HMDP/HDP, florbetaben and florbetapir. No publication‐date restrictions were applied. The complete database‐specific search strategies and selected eligibility decisions for borderline reports are provided in Supporting File 2. 2.4. Study Selection and Data Extraction Two reviewers independently screened titles and abstracts and subsequently assessed potentially eligible full texts against the prespecified eligibility criteria, with disagreements resolved by discussion. Data were extracted at study level for study design, setting, recruitment period, sample size, number and prevalence of ATTR‐CM or the stated imaging‐positive endpoint, input modality, model architecture, reference standard, partitioning strategy, internal and external testing, calibration, explainability, code/model availability and diagnostic‐performance metrics, including 95% confidence intervals where reported by the original studies. Development, tuning and testing datasets were recorded separately where reported. Cohort overlap was assessed using institution, recruitment period, investigators, model lineage and cohort descriptions; potentially overlapping reports were retained when they addressed distinct model‐development, validation or implementation questions but were not treated as independent patient evidence. 2.5. Risk of Bias and AI‐Methodology Appraisal Risk‐of‐bias judgements were made across the four QUADAS‐2 domains of patient selection, index test/model conduct, reference standard and flow/timing, and were classified as low, high or unclear risk of bias. AI‐specific considerations were incorporated into the assessment, including data partitioning, model complexity relative to event counts, threshold optimisation, external testing and verification. Applicability concerns, where assessed, were classified separately as low, high or unclear. Attention was paid to case‐control enrichment, low event counts relative to model complexity, patient‐level versus image‐level splitting, reuse of tuning data for threshold selection, absence of independent external testing and incomplete verification. In parallel, selected TRIPOD + AI/CLAIM items were extracted to describe architecture rationale, data partitioning, internal testing, external testing, calibration, explainability and code/model availability. CLAIM terminology of internal testing and external testing is used where possible to avoid the ambiguity of the term validation. 2.6. Data Synthesis and Meta‐Analysis Feasibility Formal meta‐analysis was not undertaken because the included reports did not estimate a sufficiently comparable diagnostic quantity across clinically and methodologically similar populations. Target labels varied among definitive ATTR‐CM, mixed cardiac amyloidosis with ATTR‐specific subgroup analysis and high‐grade tracer uptake suggestive of cardiac amyloidosis. Modalities and model inputs differed substantially, thresholds were study‐specific and frequently optimised internally and many reports provided AUCs without compatible 2 × 2 diagnostic data or variance estimates at a prespecified threshold. Furthermore, development, internal‐testing, external‐testing and implementation cohorts represented different stages of model evaluation, while several reports shared source populations or model lineages, limiting statistical independence. A single pooled AUC or bivariate sensitivity/specificity estimate was therefore considered clinically inappropriate. Results were synthesised narratively by modality and validation stage, with ranges of reported performance presented where appropriate. 3. Results 3.1. Study Selection and Evidence Base The database search yielded 534 records. After removal of duplicates, 268 titles and abstracts were screened. Of these, 34 full‐text articles and conference abstracts were assessed for eligibility. Following a detailed review, 28 primary reports were identified for the current synthesis (Figure 1). The included reports spanned 2021–2026 and together represented more than 120,000 report‐level participant entries. This aggregate is intentionally not reported as a unique‐patient denominator because several publications shared institutions, model lineages or derivative cohorts and some studies reported multiple development and testing datasets (Tables 1 and 2). PRISMA flowchart. TABLE 1Characteristics and principal findings of included reports. StudyDesign/setting N and target eventsModality/AI approachTesting strategyMain performanceGrogan et al. [21]Retrospective matched case‐control; Mayo Clinic4995; 2541 CA cases (mixed AL/ATTR; selected prevalence 50.9%)12‐lead ECG; deep neural network60/20/20 patient‐level train/internal‐test/holdoutAUC 0.91 (95% CI 0.90–0.93); PPV 0.86; 84% of holdout CA detected; signal preceded diagnosis in 59% by > 6 monthsHarmon et al. [22]Post‐development temporal testing; Mayo Clinic7040; 440 CA, including 201 ATTRwt and 20 ATTRv12‐lead ECG; previously developed AI‐ECGPost‐development temporal testing using fixed prior thresholdOverall AUC 0.84 (95% CI NR); ATTRwt AUC 0.84 (0.81–0.87); ATTRv AUC 0.76 (0.63–0.89); sensitivity 64.3%; specificity 90.4% at prior thresholdGoto et al. [18]Retrospective multicentre development and external testingECG cohorts > 9000; 324 ATTR cases across main ECG sites; separate echocardiography cohortsECG CNN and echocardiography deep learningHeld‐out derivation‐site testing plus external academic centresATTR‐specific ECG: BWH AUC 0.92 (95% CI 0.91–0.94), MGH 0.87 (0.84–0.90), UCSF 0.97 (0.95–0.98). ATTR‐specific echo: BWH 0.97 (0.96–0.98), MGH 0.94 (0.89–0.98), UCSF 1.00 (0.99–1.00), Northwestern 1.00 (1.00–1.00), Keio 0.96 (0.91–0.98)Oikonomou et al., 2025 (POCUS) [23]Retrospective multicentre; YNHHS + Mount Sinai38,751 POCUS patients; YNHHS included 48 confirmed ATTR‐CM cases with prior/follow‐up POCUSPOCUS video CNNPatient‐level development split; independent internal and external ED cohortsATTR‐CM AUROC 0.92 (95% CI NR) at YNHHS and 0.99 (95% CI NR) at Mount Sinai; 23/48 (47.9%) YNHHS cases screened positive a median 1.9 years before diagnosisOikonomou et al. (ECG/TTE) [24]Retrospective two‐health‐system longitudinal study1790; 286 ATTR‐CM (112/984 YNHHS; 174/806 HMH)AI‐ECG and AI‐Echo CNNsHeld‐out testing and geographically distinct external cohortAI‐Echo AUROC 0.93 (95% CI NR); AI‐ECG 0.91 (95% CI NR); probabilities diverged up to 3 years before nuclear diagnosisGarmany et al. [25]Retrospective clinical evaluation; tertiary referral cohort598; 181 ATTR‐CA (30.3%); 28 AL excludedPreviously developed AI‐Echo and AI‐ECGNo retraining; predefined thresholds in scintigraphy‐referred cohortAI‐Echo AUC 0.93 (0.91–0.95), sensitivity 86%, specificity 85%; AI‐ECG AUC 0.79 (0.76–0.83), sensitivity 80%, specificity 64%Gonzalez‐Lopez et al. [26]Retrospective single‐centre ATTR referral cohort2901; 585 ATTR‐CM (20.2%); 20,754 ECGsTime‐slicing CNN (Willem AI)Patient‐level cross‐validation plus held‐out test; no external siteTest AUC 0.88 (0.85–0.91), sensitivity 80.7%, specificity 78.5%; ATTRv AUC 0.91 (95% CI NR), ATTRwt 0.88 (NR)Hourmozdi et al. [27]External comparative testing in heart‐failure population3368; 176 confirmed ATTR‐CM (5.2%)EchoNet‐LVH and EchoGo Amyloidosis; compared with claims RF and Mayo scoreIndependent external cohort; prespecified algorithmsEchoNet‐LVH AUC 0.88 (95% CI NR); EchoGo 0.92 (NR); Mayo score 0.79 (NR); claims RF 0.49 (NR); fairness audit performedSlivnick et al. [28]Multisite development with international external testingDevelopment 2612 (52% CA); external 2719 (597 CA; 22%)Single apical four‐chamber video CNN ensembleFive‐fold development; global external testing across 18 sitesExternal AUROC 0.93 (95% CI NR); sensitivity 85%; specificity 93% after 13% uncertain outputs; ATTRwt sensitivity 85%; ATTRv 86%; PYP‐referral subgroup AUROC 0.86 (NR)Fraix et al. [29]Retrospective derivation plus external multicentre testing260 (111 ATTR‐CM; 42.7%) + external 454 (57% ATTR‐CM)Decision‐tree ML using GLS, RVFWT, apical sparing and LV mass (GRAAL)Internal cross‐validation; external multicentre cohortDerivation AUC 0.90 (0.86–0.94); external AUC 0.83 (0.80–0.87); sensitivity 87%; specificity 65%Mori et al. [30]Four‐centre echocardiographic radiomics study517; 261 ATTR‐CA (229 derivation + 32 external; selected case‐control)Echocardiographic radiomics/logistic modelsLeave‐one‐out CV in three centres; external fourth‐centre testing, n = 64Internal AUC 0.97–0.99 (95% CI NR) for full‐ROI models; external AUC 0.95 (NR); accuracy 87%Weberling et al. [31]Multicentre CMR cohort from 56 institutions400; 116 ATTR, 95 AL, 94 HCM, 95 healthyMultiparametric CMR ML cascadeRandom train/test plus sensitivity analysis using internal training/external referralsATTR vs AL AUC 0.92 (95% CI NR); HCM vs amyloidosis 0.99 (NR); healthy vs cardiac disease 1.00 (NR); ATTR‐vs‐AL fell to 0.88 (NR) using demographics + imaging onlyShiri et al. [32]Prospective single‐centre severe‐AS cohort263; 27 ATTR‐CM (10.3%)Tabular ML across ECG, echo, 4D‐CT strain/radiomics and clinical data70/30 split repeated across 100 random seeds; no external siteECG < 0.68; CT radiomics 0.70–0.76; echo 0.79; CT strain 0.85; multimodal 0.84; 95% CIs NR for these estimatesAntonopoulos et al. [33]Multicentre severe‐AS CT‐radiomics cohorts589 across study arms; ATTR event counts varied by armCTA myocardial radiomics (AmyloidRS) + clinical variablesTraining/validation cohorts by study armAmyloidRS c‐index 0.88 (95% CI NR) in training and 0.84 (NR) in validation; combined clinical + radiomics balanced accuracy 0.984Nizam et al. [34]Retrospective two‐centre study; severe‐AS/TAVR workflow816; 127 ATTR‐CM (15.6%)Multimodal CT + echo + ECG neural networkSeparate TAVR workflow test cohort; authors note not true geographic external testingMultimodal AUROC 0.85 (0.74–0.93); sensitivity 73.3%; specificity 82.9%; NPV 0.96; CT 0.70 (95% CI NR), echo 0.72 (NR), ECG 0.55 (NR); Brier score 0.081Ramos‐Polo et al. [35]Prospective single‐centre CRONOS‐ATTR124; 95 ATTR‐CM (76.6%: 30 ATTRv, 65 ATTRwt)XGBoost using clinical + ECG + echo AI‐derived featuresInternal model development onlyAUC 0.84 (95% CI NR); SHAP explanations; ECG + echo‐only model materially weakerSantarelli et al. [36]Prospective single‐centre PET study47; 28 CA + 19 controls; subtype counts reported inconsistently across abstract/tableEarly‐frame [18F]‐florbetaben PET CNNCross‐validation/internal testing onlyATTR sensitivity 0.935 (95% CI NR); specificity 0.897 (NR); accuracy 0.972 (NR); no external testingBargagna et al. [37]Single‐centre PET neural architecture‐search study47; 15 ATTR, 13 AL, 19 controls; same source cohort as Santarelli[18F]‐florbetaben PET; NAS‐NetPatient‐level splitting; repeated internal cross‐validation; no external siteATTR sensitivity 93.3% ± 7.8%, specificity 78.0% ± 2.9%, accuracy 70.9% ± 3.7%; values reported as dispersion estimates rather than 95% CIsHong et al. [38]Retrospective single‐centre multimodal nuclear study50; 15 ATTR (30%), 12 AL, 23 non‐CALogistic ML using 11C‐PiB PET/CT + 99mTc‐DPD + clinical/imaging features100‐fold stratified Monte Carlo CV; no external siteML subtype AUC 0.94 (0.93–0.95), sensitivity 0.83, specificity 0.94; rationale‐based ATTR AUC 0.95 (95% CI NR); prognostic adjusted HR 2.43Halme et al. [39]Retrospective multicentre Finnish scintigraphy cohort1334; 47 Perugini grade ≥ 2 (3.5%)2D bone‐scintigraphy CNNsInternal cross‐validation; no independent external testingAUC > 0.88 (95% CI NR) for grade < 2 vs ≥ 2; AUC 0.94 (NR) for grade 3 vs lower gradesDelbarre et al. [40]Retrospective two‐centre development/external‐testing study4681; 383 Perugini ≥ 2 (281/3048 development; 102/1633 external)Whole‐body bone‐scintigraphy CNNFive‐fold internal CV at Henri Mondor; independent Lille external cohortAUC 0.999 (95% CI NR) internally and externally; external sensitivity 96.1%; specificity 99.5%; estimated PPV falls substantially at low prevalenceSpielvogel et al. [41]Retrospective international multicentre cross‐tracer study16,241 patients/19,401 scans; 433 high‐grade uptake cases across cohortsWhole‐body 99mTc scintigraphy deep learningAustrian development; independent UK, Chinese and Italian testingAUC 1.000 (95% CI NR) internally; external AUC 0.925–1.000 (CIs NR); AI outperformed readers in audit; adjusted mortality HR 1.44Salimi et al. [42]Retrospective multicentre multi‐tracer/multi‐scanner study3747 across six datasets; labelled development and three external labelled cohorts plus 3215 unlabelled scansBone‐scintigraphy deep learning detection and severity scoringExternal datasets using HDP, PYP and DPDExternal detection AUC 0.956–0.999 (95% CIs NR); severity AUC 0.889–0.992 (CIs NR); Grad‐CAM/saliency usedMiller et al. [43]Retrospective PYP SPECT/CT diagnostic/prognostic study299; 83 ATTR‐CA (27.8%)DL CT segmentation enabling automated SPECT quantificationInternal cohort onlyCPA AUC 0.989 (0.974–1.000); VOI 0.988 (95% CI NR); TBR 0.979 (NR); quantitative measures predicted CV death/HF admissionBhattaru et al. [44]Retrospective two‐centre SPECT/CT study170 (77 primary + 93 validation); 42/77 primary had abnormal myocardial uptakeU‐net automated myocardial segmentation; PYP/HDP SPECT/CTIndependent second‐centre testingDiagnostic accuracy 100% (95% CI NR) in primary and 98% (NR) in validation; HBP ratio 3.1 ± 1.4 vs 1.4 ± 0.2Haberl et al. [45]Multicentre generative‐AI + downstream classification study15,799 patients/16,823 scans; local classifier n = 181; 6448 patients across four external sitesGenerative model + CNN uptake classifiersInternal and cross‐tracer/cross‐scanner external testingFor CA‐indicative uptake: baseline internal AUC 0.900 (95% CI NR); external AUCs 0.968 and 0.978 (CIs NR) where positive cases were present; synthetic augmentation improved AUC by mean 5 ± 4 percentage pointsBati et al. [46]Single‐centre curated PYP cohort109; 62 Perugini 2–3 positive (56.9%), 47 negativeCNN + clinical metadata; Late Fusion and CMF‐NetTrain/validation/test split plus LOO and Monte Carlo CV; no external siteEfficientNet CMF‐Net accuracy 90.9% (95% CI NR); F1 91.4% (NR); ResNet‐50 late‐fusion sensitivity 94.9% (NR); enriched scan‐positivity endpointJain et al. [47]Model development/external testing plus prospective AI‐augmented clinical programmeATTRACTnet: 799 internal + 422 external; implementation: 1471 AI‐positive, 256 eligible, 50 tested, 24 ATTR‐CMMultimodal deep learning using ECG, echo, demographics and orthopaedic codesFive‐fold internal CV, external cohort, then nonrandomised clinical implementationAUC 0.85 (95% CI NR) internally and 0.82 (NR) externally; 24/50 (48%) tested AI‐positive patients had ATTR‐CM; 21/24 (88%) initiated treatment; 2.8‐fold diagnostic‐yield increase vs historical control Note: AL, light‐chain amyloidosis; ATTR, transthyretin amyloidosis; ATTR‐CM/ATTR‐CA, transthyretin amyloid cardiomyopathy/cardiac amyloidosis; AUC/AUROC, area under the receiver‐operating‐characteristic curve; ECG, electrocardiography; HBP, heart‐to‐blood‐pool; HCM, hypertrophic cardiomyopathy; POCUS, point‐of‐care ultrasound; PYP, technetium‐99m pyrophosphate; TBR, target‐to‐background ratio; TTE, transthoracic echocardiography. Values in parentheses following AUC/AUROC estimates are 95% confidence intervals unless otherwise stated. NR indicates that a 95% confidence interval was not reported or was not available for the cited estimate. Values reported by the original studies as mean ± dispersion are retained in that form and are not presented as confidence intervals. Counts are study‐level entries and should not be summed as unique patients because of cohort/model overlap. Imaging‐uptake endpoints are identified as such and are not equated with definitive ATTR‐CM diagnosis.Abbreviations: CA, cardiac amyloidosis; CMR, cardiovascular magnetic resonance; CT, computed tomography; HF, heart failure; NPV, negative predictive value; PPV, positive predictive value; VOI, volume of involvement. Characteristics and principal findings of included reports. Note: AL, light‐chain amyloidosis; ATTR, transthyretin amyloidosis; ATTR‐CM/ATTR‐CA, transthyretin amyloid cardiomyopathy/cardiac amyloidosis; AUC/AUROC, area under the receiver‐operating‐characteristic curve; ECG, electrocardiography; HBP, heart‐to‐blood‐pool; HCM, hypertrophic cardiomyopathy; POCUS, point‐of‐care ultrasound; PYP, technetium‐99m pyrophosphate; TBR, target‐to‐background ratio; TTE, transthoracic echocardiography. Values in parentheses following AUC/AUROC estimates are 95% confidence intervals unless otherwise stated. NR indicates that a 95% confidence interval was not reported or was not available for the cited estimate. Values reported by the original studies as mean ± dispersion are retained in that form and are not presented as confidence intervals. Counts are study‐level entries and should not be summed as unique patients because of cohort/model overlap. Imaging‐uptake endpoints are identified as such and are not equated with definitive ATTR‐CM diagnosis. Abbreviations: CA, cardiac amyloidosis; CMR, cardiovascular magnetic resonance; CT, computed tomography; HF, heart failure; NPV, negative predictive value; PPV, positive predictive value; VOI, volume of involvement. TABLE 2Potential cohort and model‐lineage overlap.ReportsAssessment/implicationSantarelli et al. [36]/Bargagna et al. [37]Same Pisa/Fondazione Toscana G. Monasterio [18F]‐florbetaben cohort; both use 47 subjects. Treat as derivative analyses, not independent participantsSpielvogel et al. [41]/Haberl et al. [45]Both include Vienna bone‐scintigraphy data and overlapping external centres; Haberl explicitly reuses related cross‐centre infrastructure. Report separately but do not sum as independent evidenceGrogan et al. [21]/Harmon et al. [22]Same Mayo AI‐ECG model lineage. Harmon is post‐development temporal evaluation of the previously developed model; cohorts are temporally distinct but algorithms are not independentGrogan et al./Harmon et al. [21, 22]/Garmany et al. [25]Mayo health‐system/model lineage. Garmany evaluates previously developed AI‐ECG and AI‐Echo in a scintigraphy‐referred cohort. Participant overlap with earlier Mayo reports should be regarded as possible unless ruled out by dates/identifiersSlivnick et al. [28]/Garmany et al. [25]/Hourmozdi et al. [27]Related Ultromics/EchoGo/EchoNet model families are evaluated across different cohorts. These are valuable external evaluations but not independent model‐development programmesOikonomou et al. (POCUS) [23]/Oikonomou et al. (ECG‐TTE) [24]Both include Yale‐New Haven Health System data, but ascertainment pathways differ (ED POCUS vs nuclear amyloid referral). Potential participant overlap cannot be excludedShiri et al. [32]/Antonopoulos et al. [33]No evidence of cohort overlap identified: studies were conducted in geographically distinct health systems (Bern vs Greece/Cyprus)Hong et al. [38]/Spielvogel‐Haberl programme [41, 45]Shared Vienna investigators and institutional data sources; direct participant overlap was not established but should be flagged as possible Potential cohort and model‐lineage overlap. The evidence bases included ECG models, echocardiography and POCUS, CMR, CT/radiomics, PET, bone scintigraphy/SPECT‐CT and multimodal systems. Sample sizes ranged from 47 participants in early PET studies to more than 16,000 patients in large‐scale scintigraphy studies and 38,751 patients in real‐world POCUS testing. ATTR‐CM prevalence varied from approximately 2%–5% in broad or prevalence‐targeted screening cohorts to more than 50% in deliberately enriched case‐control or imaging datasets, underscoring why predictive values and apparent classification accuracy cannot be compared directly across reports. 3.2. Electrocardiography‐Based AI ECG‐based discrimination was heterogeneous rather than uniformly high. In the Mayo AI‐ECG study, Grogan et al. reported an AUC of 0.91 for the detection of mixed AL/ATTR cardiac amyloidosis [21]. In postdevelopmental temporal testing of the same model, Harmon et al. observed lower overall discrimination (AUC 0.84), with AUC 0.84 for ATTRwt and 0.76 for ATTRv, while specificity remained 90.4% at the previously defined threshold [22]. Goto et al. reported ATTR‐specific ECG AUCs ranging from 0.87 to 0.97 across academic‐centre cohorts [18]. Oikonomou et al. reported an AI‐ECG AUROC of 0.91 in a longitudinal ATTR‐CM nuclear‐testing cohort [24]. More recent dedicated ATTR studies reinforce the importance of the intended‐use population. Gonzalez‐Lopez et al. evaluated 20,754 ECGs from 2901 individuals, including 585 ATTR‐CM cases, and reported a held‐out AUC of 0.88 with 80.7% sensitivity and 78.5% specificity; discrimination was similar for ATTRv (0.91) and ATTRwt (0.88), although sensitivity was lower in asymptomatic and NYHA class I presentations [26]. In a high‐risk scintigraphy‐referred cohort, Garmany et al. reported an AUC of 0.79 for AI‐ECG, materially lower than the paired AI‐Echo model [25]. In the severe‐AS multimodal study by Nizam et al., ECG alone performed close to chance (AUROC 0.55) [34]. Across ECG‐containing analyses, reported AUCs therefore ranged from approximately 0.55 to 0.97, with the lower bounds emphasising the dependence of performance on phenotype, population and model lineage. 3.3. Echocardiography and Point‐of‐Care Ultrasound Echocardiographic and POCUS models generally achieved moderate‐to‐high discrimination, but again with a broad range. Oikonomou et al. reported AI‐Echo AUROC 0.93 and demonstrated divergence of AI probabilities up to 3 years before nuclear diagnosis [24]. Their POCUS‐adapted CNN subsequently achieved AUROC 0.92 at Yale‐New Haven and 0.99 in the external Mount Sinai cohort; among 48 confirmed YNHHS ATTR‐CM cases with POCUS observations, 23 (47.9%) screened positive before confirmatory testing, a median 1.9 years earlier [23]. Garmany et al. reported AUC 0.93, sensitivity 86% and specificity 85% for a previously developed AI‐Echo model in 598 patients referred for scintigraphy [25]. Goto et al. additionally reported ATTR‐specific echocardiographic C‐statistics ranging from 0.94 to 1.00 across five academic‐centre cohorts, including 0.97 (95% CI 0.96–0.98) at BWH, 0.94 (0.89–0.98) at MGH, 1.00 (0.99–1.00) at UCSF, 1.00 (1.00–1.00) at Northwestern and 0.96 (0.91–0.98) at Keio [18]. Importantly, not all echocardiographic models approached these values. The simple GRAAL decision‐tree algorithm achieved AUC 0.90 in derivation but 0.83 in an external multicentre cohort, with 87% sensitivity and 65% specificity [29]. Mori et al. reported external AUC 0.95 for echocardiographic radiomics in a small fourth‐centre test cohort [30]. In external comparative testing in a heart‐failure population, Hourmozdi et al. found AUCs of 0.88 for EchoNet‐LVH and 0.92 for EchoGo Amyloidosis, compared with 0.79 for the Mayo clinical score and 0.49 for a claims‐based random‐forest model [27]. Slivnick et al. reported overall external AUROC 0.93 across 18 sites, but AUROC 0.86 in the clinically relevant PYP‐referral subgroup; subtype sensitivities were 85% for ATTRwt and 86% for ATTRv [28]. In Shiri et al., echocardiography alone produced AUC 0.79, and in Nizam et al., echo alone produced AUC 0.72 [32, 34]. Overall, echocardiography/POCUS AUCs ranged approximately 0.72–1.00. 3.4. Cardiovascular Magnetic Resonance‐Based AI Weberling et al. evaluated 400 participants (116 ATTR, 95 AL, 94 HCM and 95 healthy volunteers) referred from 56 institutions [31]. The clinically most relevant subtype‐discrimination step, ATTR versus AL, achieved an AUC of 0.92. Higher AUCs of 0.99–1.00 applied to broader upstream tasks (HCM versus amyloidosis and healthy versus cardiac disease) and should not be interpreted as ATTR‐specific diagnostic accuracy. When the model was restricted to demographics and imaging data, ATTR‐versus‐AL discrimination fell to 0.88, illustrating the contribution of additional clinical information. 3.5. Computed Tomography, Radiomics and Multimodal AI CT‐based performance was more modest than the highest nuclear‐imaging estimates. In 263 patients with severe AS, Shiri et al. reported AUC 0.70–0.76 for CT‐radiomic features, 0.85 for 4D‐CT strain and 0.84 for the multimodal model; importantly, multimodal integration did not outperform CT strain alone [32]. Antonopoulos et al. reported an AmyloidRS c‐index of 0.88 in training and 0.84 in validation, with very high balanced accuracy after combination with clinical variables [33]. Nizam et al. reported AUROC 0.70 for CT alone and 0.85 for the CT + echo + ECG multimodal model in an independent TAVR workflow test cohort; calibration was explicitly assessed (Brier score 0.081), although the authors cautioned that the affiliated‐centre workflow did not constitute true geographic external validation [34]. The prospective CRONOS‐ATTR study used XGBoost on clinical, ECG and echocardiographic data in 124 patients and achieved AUC 0.84, with SHAP explanations but no independent external test [35]. Thus, multimodal integration improved discrimination in some datasets but not consistently across studies. 3.6. PET and Molecular Imaging The PET evidence remained small and methodologically heterogeneous. Santarelli et al. evaluated early‐frame [18F]‐florbetaben PET in 47 participants and reported ATTR sensitivity 0.935 and specificity 0.897 [36]. Bargagna et al. used neural architecture search on the same 47‐subject source cohort, with patient‐level data splitting; ATTR sensitivity was 93.3%, but specificity was lower at 78.0%, emphasising that the derivative reports should not be counted as independent patient evidence [37]. Hong et al. evaluated 50 patients with paired 11C‐PiB PET/CT and 99mTc‐DPD scintigraphy against biopsy/genetic data; the machine‐learning model achieved AUC 0.94 for subtype differentiation, and a rationale‐based ATTR classifier achieved AUC 0.95 [38]. These studies support feasibility but remain too small for firm conclusions regarding transportability. 3.7. Bone Scintigraphy and SPECT/CT Bone scintigraphy and SPECT/CT studies included both direct ATTR‐CM classification and automated detection/quantification of the high‐grade cardiac uptake phenotype used in ATTR‐CM diagnostic pathways. Halme et al. reported AUC > 0.88 for discrimination of Perugini grades < 2 versus ≥ 2 and AUC 0.94 for grade 3 uptake, but only 47 of 1334 scans were grade ≥ 2 [39]. Delbarre et al. used a two‐centre design (3048 development/internal cross‐validation; 1633 independent external testing) and reported external AUC 0.999, sensitivity 96.1% and specificity 99.5% for abnormal cardiac uptake; the same paper demonstrated that positive predictive value would decrease markedly when the model was applied at low disease prevalence [40]. Spielvogel et al. provided the strongest external‐validation evidence for automated bone‐scan interpretation, including 16,241 patients and 19,401 scans from nine international centres. AUC was 1.000 in Austrian cross‐validation and 0.925–1.000 across independent UK, Chinese and Italian cohorts, with robustness assessed across tracers/scanners and AI predictions independently associated with mortality (adjusted HR 1.44) [41]. Salimi et al. similarly demonstrated multitracer/multiscanner external testing, with external detection AUCs approximately 0.956–0.999 and external severity‐scoring AUCs 0.889–0.992 [42]. Miller et al. used deep‐learning segmentation to automate PYP SPECT/CT quantification in 299 patients with 83 confirmed ATTR‐CA cases; cardiac pyrophosphate activity and volume of involvement achieved AUCs 0.989 and 0.988, respectively, and were associated with cardiovascular death or heart‐failure hospitalisation [43]. Bhattaru et al. reported 100% diagnostic accuracy in a 77‐patient primary cohort and 98% in a 93‐patient external cohort for an automated segmentation/quantification workflow [44]. Newer work has broadened the methodological focus. Haberl et al. used generative AI to augment a small training set and tested downstream uptake classifiers across four external sites; for the cardiac‐amyloidosis‐indicative task, reported AUCs reached 0.968 and 0.978 in external cohorts containing positive cases, while synthetic augmentation improved AUC by a mean 5 ± 4 percentage points [45]. Bati et al. combined PYP images with clinical metadata in 109 patients; EfficientNet CMF‐Net achieved 90.9% accuracy and 91.4% F1‐score, although the cohort was small, single‐centre and enriched for Perugini 2–3 positivity [46]. These studies reinforce the high technical potential of nuclear‐image AI as well as illustrate why high AUC/accuracy values based on uptake labels cannot be directly equated with end‐to‐end ATTR‐CM diagnosis. 3.8. Clinical Implementation and Prediagnostic Detection Several reports moved beyond static discrimination. Oikonomou et al. demonstrated longitudinal divergence in ECG/echo model probabilities up to 3 years before nuclear diagnosis, while the POCUS study identified nearly half of confirmed cases with a positive screen a median 1.9 years before definitive diagnosis [23, 24]. Grogan et al. also reported AI‐ECG positivity more than 6 months before clinical diagnosis in 59% of detected cases [21]. Jain et al. evaluated a multimodal model (ATTRACTnet) using ECG waveforms, echocardiographic measures, demographics and orthopaedic diagnosis codes, followed by a nonrandomised clinical implementation programme. Model AUC was 0.85 internally and 0.82 externally; among 50 AI‐positive patients who underwent confirmatory testing, 24 (48%) had ATTR‐CM and 21/24 (88%) initiated treatment, with diagnostic yield 2.8‐fold higher than the historical comparator [47]. These findings are clinically important, but the verification of only a subset of eligible AI‐positive patients creates potential flow/verification bias and means implementation yield should not be interpreted as conventional test accuracy. 3.9. Methodological Quality, Reporting and Cohort Overlap The adapted QUADAS‐2 assessment identified substantial methodological heterogeneity (Table 3). Recurrent high‐risk features were case‐control or deliberately enriched populations, very small positive‐event counts relative to model complexity, internal‐only testing, optimisation of thresholds or architectures within the available dataset, and use of imaging surrogates rather than complete etiologic ATTR‐CM confirmation. The index‐test/model domain was the most frequent source of concern. Stronger studies used patient‐level splitting, independent external sites, prespecified thresholds, explicit calibration or implementation‐level evaluation. Consequently, near‐perfect discrimination in selected datasets was not interpreted as direct evidence of equivalent performance in unselected clinical populations. TABLE 3Risk‐of‐bias assessment using an adapted QUADAS‐2 framework with AI‐specific considerations. StudyPatient selectionIndex test/AIReference standardFlow/timingKey rationaleGrogan et al. [21]HighHighLowLowMatched case‐control design and internally optimised threshold; no geographic external testingHarmon et al. [22]UnclearLowLowLowTemporal post‐development testing improves transport assessment but remains within the same health system and uses matched samplingGoto et al. [18]HighUnclearUnclearLowLarge multisite testing, but case‐control enrichment and mixed cardiac‐amyloidosis target/reference standards limit applicability to ATTR‐CMOikonomou et al. (POCUS) [23]UnclearUnclearLowLowLarge real‐world external cohort; rare‐event label ascertainment and retrospective sampling remain potential sources of biasOikonomou et al. (ECG/TTE) [24]UnclearUnclearLowLowExternal health‐system testing and longitudinal design, but the cohort was enriched through referral for nuclear testingGarmany et al. [25]UnclearLowLowLowPrespecified thresholds and absence of retraining are strengths; the high‐risk referral population limits generalisabilityGonzalez‐Lopez et al. [26]HighHighLowLowSingle‐centre enriched cohort, internal tuning and no external‐site testing despite held‐out patient testingHourmozdi et al. [27]UnclearLowLowLowIndependent heart‐failure testing and fairness assessment; prevalence was intentionally constructed through matchingSlivnick et al. [28]HighUnclearUnclearHighStrong international external testing, but case‐control enrichment and uncertain/abstained outputs affecting approximately 13% of cases raise flow and applicability concernsFraix et al. [29]UnclearLowUnclearLowExternal multicentre testing and an interpretable decision tree are strengths; retrospective suspected‐ATTR cohorts and uncertain grade‐1 classification remain concernsMori et al. [30]HighUnclearLowLowExternal‐centre testing was performed, but the study used case‐control enrichment and a small external sampleWeberling et al. [31]HighUnclearLowLowMulticentre recruitment was used, but selected phenotype groups and a modest sample for complex multistage modelling limit representativenessShiri et al. [32]LowHighLowLowProspective consecutive cohort, but only 27 events were available and model selection was performed internally across a large feature spaceAntonopoulos et al. [33]UnclearHighLowUnclearRadiomics development across study arms involved limited event counts and a highly selected severe‐aortic‐stenosis populationNizam et al. [34]UnclearUnclearLowLowCalibration was assessed and an independent workflow test was used, but affiliated centres and a small TAVR test cohort limit true external validationRamos‐Polo et al. [35]HighHighLowLowVery small, high‐prevalence single‐centre cohort with internal‐only model development and evaluationSantarelli et al. [36]HighHighLowUnclearVery small cohort and internal image‐model evaluation; subtype counts were inconsistently reportedBargagna et al. [37]HighHighLowUnclearDerivative small PET cohort with internal architecture search and cross‐validation only, despite patient‐level splittingHong et al. [38]HighHighLowLowOnly 50 patients were included; evaluation relied on repeated internal cross‐validation without independent external testingHalme et al. [39]HighHighUnclearLowOnly 47 positive scans were available; evaluation used internal cross‐validation and a visual uptake reference rather than definitive ATTR‐CM diagnosisDelbarre et al. [40]UnclearLowUnclearLowStrong independent second‐centre testing; the reference standard was expert Perugini grading rather than complete clinical ATTR‐CM confirmationSpielvogel et al. [41]UnclearLowUnclearLowLarge multinational external testing and audit; the target was high‐grade cardiac uptake and two external cohorts were deliberately enrichedSalimi et al. [42]UnclearUnclearUnclearLowMulticentre external testing and explainability were strengths, but labelled development/testing datasets were relatively small and the target was tracer uptakeMiller et al. [43]UnclearHighLowLowClinically confirmed ATTR endpoint, but evaluation was retrospective, single‐centre and lacked independent external testingBhattaru et al. [44]HighUnclearUnclearLowTwo‐centre testing was performed, but sample sizes were small and the principal endpoint was uptake/segmentation rather than end‐to‐end ATTR‐CM diagnosisHaberl et al. [45]UnclearUnclearUnclearLowCross‐centre testing was a strength, but the task targeted abnormal tracer uptake and shared data lineage with previous scintigraphy studiesBati et al. [46]HighHighUnclearLowSmall single‐centre enriched cohort with internally selected architecture/thresholds and a scan‐positivity surrogate endpointJain et al. [47]UnclearUnclearLowHighExternal model testing and prospective implementation were strengths; only a subset of eligible AI‐positive patients underwent confirmatory testing, creating potential verification and flow bias Note: Judgements reflect the diagnostic question used in this review. Risk of bias was assessed using an adapted QUADAS‐2 framework [14], with AI‐specific considerations informed by QUADAS‐AI development work [15]. Nuclear‐imaging studies targeting high‐grade cardiac tracer uptake were assessed against their stated index‐test target; applicability to end‐to‐end ATTR‐CM diagnosis is considered separately in the narrative synthesis. Risk‐of‐bias assessment using an adapted QUADAS‐2 framework with AI‐specific considerations. Note: Judgements reflect the diagnostic question used in this review. Risk of bias was assessed using an adapted QUADAS‐2 framework [14], with AI‐specific considerations informed by QUADAS‐AI development work [15]. Nuclear‐imaging studies targeting high‐grade cardiac tracer uptake were assessed against their stated index‐test target; applicability to end‐to‐end ATTR‐CM diagnosis is considered separately in the narrative synthesis. The TRIPOD + AI/CLAIM mapping (Table 4) showed that external testing and explainability have become more common, but calibration and public model/code availability remain inconsistent. Explicit calibration was reported in only a minority of recent studies, notably Nizam et al. and Slivnick et al. Several imaging papers used Grad‐CAM, saliency maps, attention maps or SHAP; decision‐tree and radiomic models offered more intrinsically interpretable feature structures. Public source code was uncommon, although some studies, including Shiri and Bati, reported GitHub availability and Santarelli/Bargagna offered code on request. TABLE 4AI development, testing and reporting characteristics selected domains were mapped descriptively against TRIPOD + AI [16] and, where applicable, CLAIM 2024 [17].StudyArchitecture/rationalePartitioningInternal testingExternal testingCalibrationExplainabilityCode/model availabilityGrogan et al. [21]Deep neural network; architecture described, justification limitedPatient‐level 60/20/20Yes (internal holdout)No geographic externalNRNRNot public/proprietaryHarmon et al. [22]Evaluation of pretrained Grogan AI‐ECG; new architecture N/APost‐development temporal cohort; fixed prior thresholdN/A (evaluation study)Temporal, same institutionNRSubgroup performance auditProprietaryGoto et al. [18]CNN‐based ECG and echo pipelines; architecture describedPatient‐level derivation/validation/testYesYes, multiple academic centresNRLimited; emphasis on downstream confirmatory pathwayInference code reported; weights/data restrictedOikonomou et al. (POCUS) [23]View‐agnostic video CNN; design tailored to POCUS qualityPatient‐level training/validation/testYesYes (Mount Sinai)NRGrad‐CAM saliency mapsNROikonomou et al. (ECG/TTE) [24]EfficientNet‐based ECG and video echo models; prior validated architecturesPatient‐level development and held‐out testingYesYes (Houston Methodist)Probability trajectories modelled; formal calibration NRLongitudinal probability trajectoriesExecutable apps available on reasonable requestGarmany et al. [25]Previously developed AI‐Echo/AI‐ECG; architecture N/A for evaluationNo retraining; prespecified thresholdsN/AClinical evaluation in independent referred cohortNot recalibratedClinical utility analysisModels proprietaryGonzalez‐Lopez et al. [26]Time‐slicing CNN; rationale linked to long ECG recordingsPatient‐level 5‐fold CV + testYesNoNRNo formal explainability; acknowledged limitationWillem platform/proprietaryHourmozdi et al. [27]External evaluation of four existing modelsIndependent HF cohort; no retrainingN/AYesNRFairness/bias auditMixed; some proprietarySlivnick et al. [28]Ensemble of five 3D CNNsTraining + tuning + external cohorts; five‐fold trainingYesYes, 18 sitesYes; calibrated probabilities/calibration curvesUncertainty/abstention analysis; failure‐mode subgroup analysesNR/proprietaryFraix et al. [29]Pruned decision tree (GRAAL); intentionally simple/interpretableRepeated train/validation splitting + pruningYesYes, multicentreNRIntrinsic decision‐tree interpretabilityNRMori et al. [30]Logistic regression radiomics; feature selection tied to ROI‐dependence questionLeave‐one‐out CVYesYes, fourth centreMachine‐specific image intensity calibration; probability calibration NRRegression coefficients/feature rankingNRWeberling et al. [31]36 candidate ML algorithms with grid search; 3‐stage cascade70/30 split + 10‐fold CV model selectionYesSensitivity analysis using internal‐only training/external referrals for testingNRSelected‐feature reportingNRShiri et al. [32]Multiple tabular ML classifiers and feature‐selection strategies70/30 repeated over 100 seedsYesNoNRFeature‐selection outputsCode/models publicly available on GitHubAntonopoulos et al. [33]Radiomics score + clinical modelTraining and validation cohorts by study armYesYes/held‐out validation cohortsNRRadiomic score/feature‐level interpretabilityNRNizam et al. [34]Modality‐specific encoders with multimodal fusionTraining cohort + separate TAVR workflow testYesNo true external site (authors explicitly caution)Yes; Brier score 0.081 and reliability diagramShapley values + Grad‐CAMNRRamos‐Polo et al. [35]XGBoost selected for multimodal tabular prediction; grid‐search tuningSingle cohort internal developmentYesNoNRSHAP patient‐ and feature‐level explanationsNR; commercial component algorithmsSantarelli et al. [36]Custom CNN (CAclassNet); architecture describedCross‐validation/internal testYesNoNRNRCustom code and anonymised images available on requestBargagna et al. [37]Evolutionary neural architecture search; explicit aim to automate architecture designPatient‐level split; repeated internal splitsYesNoNRArchitecture visualisationCode available on requestHong et al. [38]Logistic regression one‐vs‐rest + random survival forest100‐fold stratified Monte Carlo CVYesNoNRSHAP feature importanceData available on reasonable request; code NRHalme et al. [39]Custom Linear/Residual CNNs compared with standard architecturesInternal CV with augmentation and validation subsetYesNoNRMaximum activation mapsNRDelbarre et al. [40]CNN classifier; model selected by F2 in 5‐fold CV5‐fold CV then independent site testYesYes (Lille)Predictive values modelled across prevalence; formal calibration NRActivation heat mapsNRSpielvogel et al. [41]Deep‐learning cardiac‐uptake detector with medical algorithmic auditAustrian CV then external sitesYesYes, cross‐tracer/cross‐scanner multinationalNRAttention maps + algorithmic auditModel availability not clearly publicSalimi et al. [42]Segmentation + ensemble classification networksInternal development plus 3 labelled external datasetsYesYes, multi‐tracer/multi‐scannerNRGrad‐CAM and saliency mapsModel/code sharing described; source data unavailableMiller et al. [43]Foundational DL CT segmentation used for quantitative SPECT metricsSingle‐cohort evaluationYesNoNRInterpretable quantitative CPA/VOI/TBR outputsNRBhattaru et al. [44]U‐net myocardial segmentationPrimary + independent second‐centre cohortYesYesNRSegmentation masks/quantitative HBP ratioNRHaberl et al. [45]Generative model + downstream classifier; architecture motivated by data scarcityLocal training, internal + 4 external sitesYesYes, cross‐tracer/cross‐scannerNRReader study + downstream performance analysisRepresentative synthetic data available; full pipeline availability limitedBati et al. [46]CNN image encoders + MLP; Late Fusion/CMF‐Net with attentionPatient‐level train/validation/test; LOO and 100‐iteration MCCVYesNoNRAttention‐based fusion; architecture ablationsSource code/model class publicly available on GitHubJain et al. [47]ATTRACTnet multimodal deep learning; designed for clinical case enrichment5‐fold internal CV + external model test + implementation cohortYesYesNRClinical‐pathway outputs; feature contribution reporting limitedNR Note: These are reporting descriptors, not a numerical quality score. NR = not reported/unclear in the publication; N/A = not applicable. AI development, testing and reporting characteristics selected domains were mapped descriptively against TRIPOD + AI [16] and, where applicable, CLAIM 2024 [17]. Note: These are reporting descriptors, not a numerical quality score. NR = not reported/unclear in the publication; N/A = not applicable. Potential participant or model‐lineage overlap was also clinically relevant (Table 2). Santarelli and Bargagna used the same 47‐subject florbetaben source cohort. Spielvogel and Haberl shared Vienna and related external scintigraphy infrastructure. Grogan and Harmon represent development and postdevelopment testing of the same Mayo AI‐ECG; Garmany subsequently evaluated previously developed AI models in a Mayo referral cohort. Related Ultromics/EchoGo/EchoNet model families appear across Slivnick, Garmany and Hourmozdi. These reports provide complementary evidence at different development stages but should not be treated as statistically independent studies or summed as unique participants. 4. Discussion This systematic review demonstrates a broad and increasingly mature evidence base for AI‐assisted detection of ATTR‐CM. The available studies show a clear gradient of evidential maturity, with large externally validated cohorts supporting automated recognition of cardiac‐amyloidosis‐suggestive bone‐tracer uptake and echocardiographic phenotypes, while subtype‐classification, PET, CMR and smaller multimodal studies remain predominantly developmental. Overall, the evidence supports cautious optimism for AI‐assisted case enrichment, screening and diagnostic interpretation, but does not justify a general claim of uniformly high ATTR‐CM diagnostic accuracy across all modalities. The largest nuclear‐imaging studies demonstrate impressive discrimination across scanners, tracers and countries. Spielvogel et al. is particularly important because its 16,241‐patient multinational dataset, independent cross‐tracer testing, reader comparison and prognostic analyses address several limitations of earlier single‐centre studies. However, the task was automated detection of high‐grade cardiac uptake suggestive of amyloidosis rather than the entire clinical algorithm required to distinguish ATTR from AL and other causes of uptake. This distinction matters: technical accuracy at reading a bone scan cannot replace the necessary assessment for a monoclonal protein disorder, SPECT confirmation where appropriate and clinical integration. A consistent finding across modalities was that diagnostic performance depended strongly on the population and stage of model evaluation. Performance frequently decreased when models progressed from selected derivation cohorts to more clinically representative testing populations. For example, the Mayo AI‐ECG demonstrated an AUC of 0.91 in the development study by Grogan et al., compared with 0.84 during subsequent postdevelopment testing by Harmon et al. Similarly, the GRAAL algorithm achieved an AUC of 0.90 in derivation and 0.83 in external multicentre testing. In the comparative evaluation by Hourmozdi et al., two echocardiographic AI models retained good discrimination, whereas the claims‐based model performed close to chance. Nizam et al. similarly reported marked variation by input modality, with AUROC 0.55 for ECG alone compared with 0.85 for the multimodal model. Collectively, these findings emphasise the importance of validation setting and intended‐use population when interpreting reported discrimination and caution against relying on best‐performing estimates in isolation. Prevalence is equally important. AUC describes ranking discrimination but does not determine the number of false positives that will arise in low‐prevalence screening. Several studies deliberately enriched cases or constructed case‐control datasets, while large consecutive scintigraphy cohorts contained only about 2% high‐grade uptake. Delbarre explicitly showed that positive predictive value could fall substantially when prevalence approached that expected in general populations. For deployment, calibration, PPV, decision‐curve/net‐benefit analysis, workflow capacity and downstream confirmatory testing are therefore at least as important as AUC. Echocardiography and ECG are attractive because they are routinely acquired and can provide opportunistic signals before definitive testing. The longitudinal Oikonomou studies and real‐world POCUS data support the possibility that AI probabilities may become abnormal before conventional diagnosis. The Gonzalez‐Lopez data additionally show that dedicated ATTR‐CM AI‐ECG retains useful discrimination in early presentations, although sensitivity is lower among asymptomatic or NYHA I patients. These models should be conceptualised as triage or safety‐net tools that trigger confirmatory evaluation rather than stand‐alone etiologic diagnoses. Multimodal integration is promising but not automatically superior. In Shiri, the multimodal model did not outperform CT strain alone. Nizam found meaningful improvement from integration, but the test cohort remained within affiliated academic workflows. CRONOS‐ATTR achieved interpretable multimodal prediction in a prospective setting but remains small and internally evaluated. Future multimodal research should therefore prioritise prospective external testing and calibration over simply increasing the number of input modalities. The findings should also be considered in the context of the existing evidence syntheses. Previous systematic reviews have evaluated AI applications in cardiac amyloidosis broadly, while an ATTR‐CM‐specific systematic review and an AI‐ECG meta‐analysis have also been published [6–8, 10]. A 2026 meta‐analysis of machine‐learning diagnostic studies reported pooled ATTR‐CA sensitivity of 0.84, specificity of 0.85 and an SROC AUC of 0.91 [9]. The present review addresses a complementary question by examining AI across the multimodality ATTR‐CM diagnostic pathway, distinguishing models predicting definitive disease from those predicting imaging‐surrogate endpoints, evaluating model development and reporting characteristics and explicitly considering potentially overlapping cohorts and model lineages. Given the substantial clinical, methodological and target‐label heterogeneity identified, generation of a single additional pooled diagnostic estimate was considered unlikely to provide a clinically meaningful summary of the evidence. Future studies should prospectively evaluate locked models in geographically, demographically and technically independent populations; report event prevalence and complete 2 × 2 performance at prespecified thresholds; assess calibration and clinical utility at realistic screening prevalence; use patient‐level data partitioning; publish model/version details and code where feasible and report subgroup performance by age, sex, race/ethnicity, ATTR genotype and important phenocopies. For multimodal AI, incremental value should be compared formally against simpler single‐modality models and established clinical scores. Implementation studies should measure diagnostic yield, false‐positive burden, confirmatory‐test utilisation, time to diagnosis, treatment initiation, patient outcomes and cost‐effectiveness. 4.1. Limitations Several limitations should be acknowledged. First, the evidence base was dominated by retrospective studies and selected referral or case‐control populations, with relatively few prospective or implementation studies. Second, target definitions varied across studies: some models predicted aetiologically confirmed ATTR‐CM, others evaluated mixed cardiac amyloidosis with ATTR‐specific subgroup performance and several nuclear‐imaging studies predicted high‐grade cardiac tracer uptake rather than definitive ATTR‐CM. Third, several publications shared source cohorts, institutions or model lineages, preventing calculation of a defensible unique‐patient denominator and limiting statistical independence across reports. Fourth, calibration, use of prespecified thresholds, public availability of code or models and complete reporting of test‐set event counts were inconsistent. Finally, AI in ATTR‐CM is a rapidly evolving field, and continuing model development and external validation may alter the strength and applicability of the available evidence over time. 5. Conclusion AI can identify ATTR‐CM‐relevant patterns across ECG, echocardiography/POCUS, CT, CMR, PET and nuclear scintigraphy, and several models have now demonstrated meaningful external or clinical‐workflow performance. Nevertheless, the evidence is not uniform: performance varies by task, population and stage of testing, while many studies remain susceptible to spectrum bias, low event counts, internal optimisation and incomplete calibration. AI is therefore best positioned at present as an adjunct for case enrichment, opportunistic screening and standardised interpretation, followed by established confirmatory pathways. Prospective multicentre validation and transparent reporting are required before routine adoption. Author Contributions Eyad Jamileh and Ahmed T. Elmewafy collected and analysed data, contributed to drafting and revised the article. Muhammad Abdullah Qadeer, Zuhaib Zulfiqar and Kiran Kang collected and analysed data and contributed to drafting. Kaveh Hosseini and Ibrahim Antoun contributed to the design of the study and final approval. Funding The authors received no financial support for the research, authorship and/or publication of this article. Ethics Statement A statement of ethics is not applicable because this study is based exclusively on published literature. Conflicts of Interest The authors declare no conflicts of interest. Supporting Information Additional supporting information can be found online in the Supporting Information section. Supporting information Supporting Information The supporting information accompanying this manuscript includes Supporting File 1, which contains the PRISMA 2020 checklist, and Supporting File 2, which details the comprehensive search strategy used across the databases.