Subtype specific artificial intelligence modelling of cardiac amyloidosis using echocardiography and electrocardiography Authors: Jose K James, Surendra Dasari, Jennifer M Amadio, Christopher G Scott, Eli Muchtar, Morie A Gertz, Shaji Kumar, Francis K Buadi, David Dingli, Taxiarchis V Kourelis, Itzhak Z Attia, Francisco Lopez-Jimenez, Omar AbouEzzeddine, Dennis H Murphree, Paul A Friedman, Martha Grogan, Angela Dispenzieri Published: 2026 ABSTRACT AbstractAimsWe aimed to develop and validate echocardiography-based prediction models for light chain (AL) and transthyretin (ATTR) cardiac amyloidosis and to quantify the incremental value of integrating artificial intelligence–derived electrocardiography (ECG) probabilities.Methods and resultsWe conducted a retrospective, multisite study within a single health system including patients with AL or ATTR cardiac amyloidosis and matched control subjects. AL and ATTR cohorts were handled independently. For each subtype, patients were randomly split into training and test cohorts, with an additional temporally distinct test cohort. Logistic regression models were developed using structured echocardiographic variables with parsimonious core and extended feature sets, as well as models integrating these features with a previously validated AI ECG probability. Model performance was assessed using receiver operating characteristic and precision–recall (PR) analyses. In AL amyloidosis, echocardiography-based models demonstrated moderate discrimination in the primary test cohort (area under the PR curve, AUPRC 0.674) but were inferior to ECG alone (AUPRC 0.824, P < 0.001), with no significant improvement from model combination (AUPRC 0.841 vs. ECG, P = 0.267). In contrast, for ATTR amyloidosis, echocardiography alone (AUPRC 0.687) performed worse than ECG (AUPRC 0.745, P = 0.140), while the combined model showed substantial improvement (AUROC 0.845, P < 0.001). Absolute performance declined in temporally distinct cohorts, but AUPRC comparisons were preserved.ConclusionSubtype-specific echocardiography-based models demonstrate robust discrimination for cardiac amyloidosis, with divergent contributions of echocardiographic features by subtype. These results support tailored modelling strategies and prospective concurrent deployment of AL and ATTR models. Introduction Cardiac amyloidosis (CA) is a life-threatening infiltrative cardiomyopathy comprising two main types: immunoglobulin light chain amyloidosis (AL) and transthyretin amyloidosis (ATTR).1–5 Together, these account for nearly all clinically observed cases of CA.6 Both result from misfolded precursor proteins that deposit as amyloid fibrils within the myocardium and other organs. In AL amyloidosis, monoclonal plasma cells in the bone marrow produce immunoglobulin light chains that serve as the precursor protein, whereas in ATTR amyloidosis, the precursor is transthyretin synthesized in the liver, either in its wild-type form (ATTRwt) or as a mutant variant (ATTRv).4 Diagnosis of CA remains challenging because of its non-specific clinical manifestations and low prevalence. Broader availability of cardiac magnetic resonance imaging and nuclear cardiac scintigraphy has improved recognition of ATTR-related disease, while the emergence of disease-modifying therapies underscores the importance of early identification and treatment.3,5,7 For AL amyloidosis, serum free light-chain measurements interpreted in the appropriate clinical context remain a key screening tool. Several predictive models have recently been developed to screen for CA8,9; however, few integrate data across multiple modalities such as electrocardiography (ECG), transthoracic echocardiography (TTE), and clinical biomarkers. We have developed an artificial intelligence (AI) model using standard 12-lead ECGs to detect CA and facilitate earlier diagnosis.1 This model predicted disease presence in 59% of patients more than 6 months before clinical diagnosis, with a positive predictive value of 0.86. Although AL and ATTR CA both result in myocardial amyloid deposition, they represent biologically and clinically distinct diseases, with differences in pathophysiology, disease progression, and therapeutic approach, motivating separate modelling strategies for each subtype. TTE remains central to the evaluation of cardiac involvement, providing structural and functional information complementary to ECG, with phenotypic patterns that differ between AL and ATTR CA.10 Recently published models integrating AI-ECG with structured echocardiographic and clinical variables have demonstrated promising performance for ATTR CA11; however, analogous approaches have not been systematically applied to AL amyloidosis. We therefore aim to determine whether integrating echocardiographic variables with the AI-ECG algorithm improves diagnostic performance using distinct models for ATTR and AL CA. Methods Clinical cohort This study was approved by the Mayo Clinic Institutional Review Board and conducted in accordance with the Declaration of Helsinki. The training and test cohorts consisted of patients identified from the Mayo Clinic amyloid database with ECG studies performed between 1 January 2000 and 31 May 2019, as previously described by Grogan et al. (2021).1 In this previously described cohort, TTE reports were also obtained and linked to corresponding ECG studies. The training set included 836 AL, 337 ATTR cases, and 5865 controls; the test set included 849 AL, 323 ATTR, and 5865 controls. For further validation, a temporally distinct validation cohort was independently assembled from the Mayo Clinic echocardiography database and included 14 399 unique patients who underwent studies between 1 January 2019 and 3 September 2025, across all three Mayo Clinic sites (Rochester, MN; Scottsdale, AZ; and Jacksonville, FL). This included 507 AL, 822 ATTR, and 13 084 controls. The temporally distinct validation cohort was carefully reviewed to exclude any patients overlapping with original derivation of the Mayo Clinic amyloid database. Control subjects were selected from patients with clinically obtained ECG and TTE data who had no evidence of systemic or CA before or at the index date, based on amyloid registry membership, diagnostic codes, pathology evidence, and nuclear scintigraphy reports. Because amyloid-specific testing was not systematically performed, undiagnosed amyloidosis cannot be fully excluded. For patients with AL or ATTR, cardiac involvement was defined using criteria adapted from current expert consensus recommendations,12 including one or more of the following: interventricular septal (IVS) thickness > 12 mm, troponin T > 0.035 mcg/L (or high sensitivity troponin >50 ng/L), N-terminal pro-B-type natriuretic peptide (NT-proBNP) > 332 pg/mL, or cardiac symptoms not explained by another cardiac condition. All patients with ATTR were considered to have cardiac involvement. ATTR was diagnosed either by histologic confirmation or by abnormal 99mTc-PYP (or HMDP) scintigraphy in the absence of a monoclonal gammopathy. All participants had a standard 12-lead ECG and a TTE performed within 180 days of each other. When multiple echocardiograms were available during this interval, the study with the most complete set of 12 predefined ‘core’ echocardiographic variables was selected for analysis. Derivative cohort: model training and validation strategy Patients with AL and ATTR amyloidosis were analysed independently to develop subtype-specific predictive models. Within each amyloidosis subtype, patients were randomly divided in a 1:1 ratio into training and test cohorts. Control subjects were age- and sex-matched to cases separately within each subtype and within each data split. This design yielded four primary datasets: AL training, AL test, ATTR training, and ATTR test cohorts. Model development for each amyloidosis subtype was performed using the corresponding training cohort, with hyperparameters optimized through internal five-fold stratified cross-validation with class weights applied to account for outcome imbalance. Final model performance was evaluated in the held-out test cohort and in an independent temporally distinct test cohort that was not used during model development (Figure 1). Study design and cohort assembly for subtype-specific cardiac amyloidosis modelling. Echocardiographic variable selection and model training From an initial set of 38 candidate echocardiographic variables (see Supplementary material online, Appendix), feature selection was performed separately within the AL and ATTR training cohorts to identify subtype-specific predictors of CA. Random forest classifiers were trained within each cohort, and variable importance was quantified using permutation-based importance metrics. Variables were ranked by importance, and those contributing to the inflection point (‘elbow’) of the importance curve, corresponding to ∼20 features per subtype, were retained for further consideration (see Supplementary material online, Figure S1). To enhance model robustness and generalizability, candidate variables were subsequently filtered based on data availability and prevalence in the test cohort. Variables with >30% missingness or low frequency were excluded, resulting in a final panel of 10–11 clinical and echocardiographic predictors for each amyloid subtype. Continuous variables were standardized, categorical variables were one-hot encoded, and missing values were imputed within each cross-validation fold to prevent information leakage. Final predictive models, collectively referred to as ECHO models, were developed using logistic regression with L1 regularization to promote sparsity and reduce overfitting. Variables were used to construct de novo echocardiographic models with differing levels of complexity: an extended model incorporating the full set of selected variables and a parsimonious core model restricted to a smaller set of routinely available features. For each amyloid subtype, the core model was limited to six variables: age, sex, weight, relative wall thickness, IVS thickness, and systolic blood pressure. To assess the incremental value of electrocardiographic information, the AI-ECG amyloidosis probability score was incorporated as an additional continuous predictor, yielding combined ECHO + ECG models. Accordingly, four models were developed for each amyloid subtype: ECHO core, ECHO extended, ECHO + ECG core, and ECHO + ECG extended (see Supplementary material online, Table S1). Statistical analysis Baseline population characteristics were summarized separately for the training cohort, the primary test cohort, and the temporally distinct test cohort, with the latter two stratified by case and control status. Given large cohort sizes, between-group differences were quantified using absolute standardized differences (ASDs) rather than hypothesis tests. For each variable, ASD was computed for three pairwise comparisons: AL vs. controls, ATTR vs. controls, and AL vs. ATTR. We finally also compared AL and ATTR cases in the primary test and temporally distinct test cohorts. Equations for ASD for continuous and dichotomous variables are as follows, with values ≥0.10 indicating meaningful imbalance as previously described.13 For continuous variables, let x¯1 and x¯2 denote the group means, and s1 and s2 denote the group sample standard deviations. For dichotomous variables, let p1 and p2 denote the group proportions. In addition to reporting effect sizes using ASD, pairwise hypothesis testing was performed to provide complementary context for baseline comparisons, with categorical variables evaluated using χ2 tests and continuous variables using Wilcoxon rank-sum tests. Model discrimination was evaluated using receiver operating characteristic (ROC) and precision–recall (PR) analyses. For each model, predicted probabilities were compared with the binary reference standard (amyloid diagnosis) separately in the primary test cohort and the temporally distinct test cohort. ROC curves characterize discrimination across all possible decision thresholds by plotting the true positive rate (sensitivity) against the false positive rate (1 − specificity). Discrimination was summarized using the area under the ROC curve (AUROC), a threshold-independent measure that reflects the probability that a randomly selected case is assigned a higher predicted risk than a randomly selected control. An AUROC of 0.5 indicates no discriminative ability beyond chance, whereas a value of 1.0 represents perfect discrimination. PR curves were used to further assess model performance with an emphasis on the positive class, plotting precision (positive predictive value) against recall (sensitivity) across thresholds. The area under the PR curve, reported as average precision (AP) or AUPRC, is particularly informative in class-imbalanced settings because it directly captures the tradeoff between case identification and false-positive burden. In contrast to AUROC, the minimum achievable AUPRC is determined by outcome prevalence, with the baseline equal to the prevalence of the positive class and a maximum of 1.0 indicating perfect performance. Confidence intervals for AUROC and AUPRC were estimated using percentile bootstrap resampling with replacement (300 iterations), excluding bootstrap samples containing only a single outcome class. The AI-ECG amyloidosis probability score was additionally evaluated as a standalone predictor and was computed within each model-specific analytic subset to ensure consistent inclusion criteria across modalities. Differences in AUROC and AUPRC between cohorts, as well as comparisons between multimodal models and the ECG-only comparator, were assessed using two-sided bootstrap tests of the metric difference with independent resampling. Reporting was reviewed, and risk of bias and applicability were evaluated using the PROBAST + AI framework.14 A completed checklist and domain-level assessment are provided in the Supplementary material online, Appendix. Results Baseline characteristics Baseline clinical, electrocardiographic, and echocardiographic characteristics are shown for the test (Table 1) and temporally distinct test (Table 2) cohorts. To account for biological and phenotypic differences between amyloid subtypes, analyses were performed separately for ATTR and AL CA, with subtype-specific models evaluated within their respective test cohorts; non-amyloid control groups were shared across subtype analyses. Table 1Baseline characteristics of ATTR and AL cohorts in the test setVariableControlsALASDaATTRASDbASDc n 5865849323Age (years)65.32 ± 11.1463.22 ± 10.780.19**72.57 ± 10.300.68**0.89**Male (%)4187 (71.4%)553 (65.1%)0.14**292 (90.4%)0.50**0.64**BMI (kg/m2)29.67 ± 8.2726.78 ± 5.270.42**27.11 ± 4.310.39**0.07*Weight (kg)88.60 ± 20.5479.70 ± 17.960.46**83.08 ± 15.340.30**0.20**Systolic blood pressure (mmHg)126.70 ± 19.73111.19 ± 19.270.80**117.30 ± 18.170.50**0.33**Cardiac index (L/min/m2)3.10 ± 0.752.93 ± 0.820.22**2.53 ± 0.640.82**0.55**LV ejection fraction (%)58.79 ± 12.8557.00 ± 13.750.13**51.96 ± 14.150.51**0.36**LV end-diastolic dimension (mm)50.07 ± 6.7244.37 ± 6.140.89**45.99 ± 6.240.63**0.26**LV end-systolic dimension (mm)32.93 ± 7.6029.73 ± 5.520.48**32.74 ± 6.690.030.49**LV posterior wall thickness (mm)10.60 ± 1.7913.43 ± 2.871.18**14.82 ± 2.941.73**0.48**Interventricular septal thickness (mm)11.18 ± 2.0914.31 ± 3.151.17**16.58 ± 3.491.88**0.68**Relative wall thickness0.43 ± 0.090.62 ± 0.191.33**0.66 ± 0.181.65**0.21**Aortic valve peak velocity (m/s)1.95 ± 1.571.40 ± 0.440.47**1.40 ± 0.650.46**0.00*E velocity (m/s)0.79 ± 0.260.86 ± 0.240.25**0.83 ± 0.240.16**0.09E/A1.11 ± 0.581.71 ± 1.150.66**2.15 ± 1.311.03**0.35** e′ (medial, m/s)0.07 ± 0.100.05 ± 0.020.30**0.04 ± 0.020.39**0.29** E/e′ (medial)13.14 ± 7.2420.62 ± 10.780.82**21.72 ± 9.151.04**0.11*RV systolic pressure (mmHg)34.40 ± 11.3739.77 ± 11.420.47**41.43 ± 11.110.63**0.15*Atrial arrhythmias (%)309 (5.3%)60 (7.1%)0.08*92 (28.5%)0.65**0.58**ATTR amyloidosis (%)0 (0.0%)0 (0.0%)NA323 (100.0%)NANAAL amyloidosis (%)0 (0.0%)849 (100.0%)NA0 (0.0%)NANAAortic stenosis (%)1509 (25.7%)27 (3.2%)0.68**26 (8.0%)0.49**0.21**Mitral regurgitation (%)5390 (91.9%)798 (94.0%)0.08*316 (97.8%)0.27**0.20*Tricuspid regurgitation (%)5717 (97.5%)813 (95.8%)0.10*318 (98.5%)0.070.16*LV thrombus (%)7 (0.1%)0 (0.0%)0.050 (0.0%)0.05NAContinuous variables are presented as mean ± standard deviation, and dichotomous variables are presented as counts and percentages.ASD, absolute standardized difference; ASD values ≥0.1 indicate a meaningful imbalance.13 aAL cases vs. controls. bATTR cases vs. controls. cAL cases vs. ATTR cases.*P-value < 0.05, **P-value < 0.001. Baseline characteristics of ATTR and AL cohorts in the test set Continuous variables are presented as mean ± standard deviation, and dichotomous variables are presented as counts and percentages. ASD, absolute standardized difference; ASD values ≥0.1 indicate a meaningful imbalance.13 aAL cases vs. controls. bATTR cases vs. controls. cAL cases vs. ATTR cases. *P-value < 0.05, **P-value < 0.001. Table 2Baseline characteristics of ATTR and AL cohorts in the temporally distinct test setVariableControlsALASDaATTRASDbASDcASDdASDe n 13 084507822Age (years)72.22 ± 10.2767.09 ± 9.350.52**75.59 ± 9.390.34**0.91**0.38**0.31**Male (%)10 476 (80.1%)325 (64.1%)0.36**742 (90.3%)0.29**0.66**0.020.01BMI (kg/m2)29.56 ± 7.5027.26 ± 5.410.35**29.14 ± 5.130.070.36**0.090.43**Weight (kg)89.29 ± 20.9182.20 ± 19.120.35**89.72 ± 17.880.020.41**0.13*0.40**Systolic blood pressure (mmHg)133.63 ± 20.49118.17 ± 19.990.76**128.13 ± 19.840.27**0.50**0.36**0.57**Cardiac index (L/min/m2)3.02 ± 0.652.89 ± 0.710.20**2.59 ± 0.610.69**0.45**0.060.09LV ejection fraction (%)61.06 ± 5.0558.34 ± 11.030.32*51.88 ± 13.140.92**0.53**0.110.01LV end-diastolic dimension (mm)49.65 ± 5.5345.86 ± 6.370.64**48.25 ± 6.220.24**0.38**0.24**0.36**LV end-systolic dimension (mm)32.39 ± 4.4830.59 ± 5.580.36**34.33 ± 6.760.34**0.60**0.16*0.24**LV posterior wall thickness (mm)9.92 ± 1.7312.13 ± 2.740.97**13.50 ± 3.131.42**0.47**0.46**0.43**Interventricular septal thickness (mm)10.50 ± 2.0013.04 ± 3.230.95**14.91 ± 3.701.48**0.54**0.40**0.46**Relative wall thickness0.40 ± 0.080.55 ± 0.171.06**0.58 ± 0.181.25**0.17**0.44**0.49**Aortic valve peak velocity (m/s)1.70 ± 0.761.41 ± 0.490.45**1.52 ± 0.630.25**0.200.020.19**E velocity (m/s)0.71 ± 0.230.79 ± 0.230.32**0.80 ± 0.250.38**0.070.29**0.13*E/A1.00 ± 0.461.39 ± 0.960.52**1.74 ± 1.080.89**0.34**0.31**0.35** e′ (medial, m/s)0.07 ± 0.070.06 ± 0.020.31**0.05 ± 0.020.43**0.31**0.27**0.36** E/e′ (medial)10.88 ± 4.6716.45 ± 9.710.73**18.71 ± 9.641.03**0.23**0.41**0.32**RV systolic pressure (mmHg)34.39 ± 11.4838.12 ± 11.420.33**40.67 ± 11.750.54**0.22**0.15*0.07Atrial arrhythmias (%)1439 (11.0%)48 (9.5%)0.05223 (27.1%)0.42**0.47**0.090.03ATTR amyloidosis (%)0 (0.0%)0 (0.0%)NA822 (100.0%)NANANANAAL amyloidosis (%)0 (0.0%)507 (100.0%)NA0 (0.0%)NANANANAAortic stenosis (%)1528 (11.7%)19 (3.7%)0.30**102 (12.4%)0.020.32**0.030.14*Mitral regurgitation (%)12 685 (97.0%)491 (96.8%)0.01813 (98.9%)0.14*0.14*0.14*0.09Tricuspid regurgitation (%)12 857 (98.3%)497 (98.0%)0.02817 (99.4%)0.11*0.12*0.13*0.09LV thrombus (%)2 (0.0%)0 (0.0%)0.020 (0.0%)0.02NANANAASD, absolute standardized difference; ASD values ≥0.1 indicate a meaningful imbalance.13 aAL temp. distinct cases vs. controls. bATTR temp. distinct cases vs. controls. cAL temp. distinct cases vs. ATTR temp. distinct cases. dAL initial test cases vs. AL temp. distinct cases. eATTR initial test cases vs. ATTR temp. distinct cases.*P-value < 0.05, **P-value < 0.001. Baseline characteristics of ATTR and AL cohorts in the temporally distinct test set ASD, absolute standardized difference; ASD values ≥0.1 indicate a meaningful imbalance.13 aAL temp. distinct cases vs. controls. bATTR temp. distinct cases vs. controls. cAL temp. distinct cases vs. ATTR temp. distinct cases. dAL initial test cases vs. AL temp. distinct cases. eATTR initial test cases vs. ATTR temp. distinct cases. *P-value < 0.05, **P-value < 0.001. In the test set (Table 1), patients with ATTR were older and more frequently male compared with controls, whereas patients with AL were younger and included a lower proportion of men. Both ATTR and AL were associated with increased left ventricular wall thickness, including IVS thickness, posterior wall thickness, and relative wall thickness, with greater wall thickness observed in ATTR compared with AL. Markers of diastolic dysfunction differed between amyloid subtypes and controls, with both ATTR and AL CA associated with lower early diastolic mitral annular velocities and higher transmitral filling indices, and more pronounced abnormalities observed in ATTR. Structural valve disease also differed between groups, with a lower prevalence of aortic stenosis in both amyloid subtypes compared with controls. Mitral regurgitation was modestly more prevalent in both amyloid subtypes, whereas tricuspid regurgitation was highly prevalent across groups with only small differences. In the temporally distinct test set (Table 2), the overall patterns observed in the primary test cohort were largely preserved. Patients with ATTR remained older and more frequently male than controls, whereas patients with AL were younger and included a lower proportion of men. However, controls in the temporally distinct cohort were older and more male-predominant than those in the primary test set, narrowing age- and sex-based separation for ATTR vs. controls while increasing separation for AL vs. controls. Both amyloid subtypes continued to demonstrate increased left ventricular wall thickness compared with controls, although differences were reduced relative to the primary test set. Markers of diastolic dysfunction remained more abnormal in both ATTR and AL than in controls, with greater abnormalities again observed in ATTR. Differences in structural valve disease were less uniform than in the primary test set: aortic stenosis remained less prevalent in AL compared with controls, whereas its prevalence in ATTR was similar to that of controls. Mitral and tricuspid regurgitation were common across all groups, with only modest differences between amyloid subtypes and controls. Atrial arrhythmias remained more prevalent in ATTR than in AL and controls, although the magnitude of difference was reduced relative to the primary test cohort. Core model performance All core models demonstrated strong discriminative performance for both AL and ATTR CA in the primary test cohort, as reflected by high AUROC values (Figure 2). Across both amyloid subtypes, all core models exhibited statistically significant reductions in AUROC when evaluated in the temporally distinct test cohort, reflecting reduced model generalizability across temporally separated populations. Discrimination of AL and ATTR cardiac amyloidosis prediction models across test cohorts. (Top panel) ROC curves for AL and ATTR cardiac amyloidosis prediction models are shown. Curves correspond to the AI ECG model (black), ECHO-only model (blue), and combined ECHO and ECG model (green). Model performance is displayed for the primary test cohort (solid lines) and the temporally distinct test cohort (dashed lines). The diagonal dotted line denotes the performance of a non-informative classifier. (Middle panel) A table summarizes AUROC values with 95% CI for each model in the primary test and temporally distinct test cohorts. (Bottom panel) A table presents P-values for pairwise comparisons of AUROC between cohorts and for comparisons with the AI ECG-only model. For ATTR amyloidosis, the core ECHO model demonstrated excellent discrimination in the primary test cohort [AUROC 0.930 (95% CI, 0.904–0.954)], which decreased in the temporally distinct test cohort [AUROC 0.852 (95% CI, 0.836–0.869); P < 0.001]. The combined ECHO + ECG core model yielded the highest discrimination for ATTR in both cohorts, achieving an AUROC of 0.972 (95% CI, 0.961–0.982) in the primary test cohort and 0.903 (95% CI, 0.887–0.917) in the temporally distinct test cohort. In the primary test cohort, discrimination of the combined model was not statistically different from the ECG-only model [AUROC 0.956 (95% CI, 0.944–0.967); P = 0.060]. However, in the temporally distinct test cohort, the combined ECHO + ECG model significantly outperformed the ECG-only model (P < 0.001). In contrast, for AL amyloidosis, although the core ECHO model achieved an AUROC of 0.868 (95% CI, 0.846–0.890) in the primary test cohort, declining to 0.816 (95% CI, 0.792–0.838) in the temporally distinct test cohort (P < 0.001), in neither cohort did the combination of the AL ECHO model perform better than the AI ECG alone using the AUROC analysis. Precision–recall analysis demonstrated broadly consistent performance patterns across AL and ATTR CA, reflecting the influence of class imbalance and differences in cohort composition (Figure 3). For ATTR amyloidosis, the core ECHO model achieved an AUPRC of 0.687 (95% CI, 0.611–0.754) in the primary test cohort, which decreased in the temporally distinct test cohort (0.515, 95% CI 0.482–0.551; P < 0.001). The combined ECHO + ECG core model achieved the highest AP for ATTR in both cohorts, with values of 0.845 (95% CI, 0.807–0.879) in the primary test cohort and 0.629 (95% CI, 0.590–0.676) in the temporally distinct test cohort. In contrast to AUROC-based comparisons, the combined ECHO + ECG model significantly outperformed the ECG-only model for ATTR in both the primary test cohort (P < 0.001) and the temporally distinct test cohort (P < 0.001). In contrast for AL amyloidosis, although the core ECHO model achieved an AUPRC of 0.674 (95% CI, 0.628–0.716) in the primary test cohort, declining to 0.368 (95% CI, 0.325–0.421) in the temporally distinct test cohort (P < 0.001), in neither cohort did the combination of the AL ECHO model perform better than the AI ECG alone using the AUPRC analysis. Precision–recall performance of AL and ATTR cardiac amyloidosis prediction models across test cohorts. (Top panel) Precision–recall curves for AL and ATTR cardiac amyloidosis prediction models are shown. (Middle panel) A table summarizes AUPRC, reported as average precision (AP), with 95% CI for each model in the primary test and temporally distinct test cohorts. (Bottom panel) A table presents P-values for pairwise comparisons of AP between cohorts and for comparisons with the AI ECG-only model. Core vs. extended model performance Across both AL and ATTR amyloidosis, extended models demonstrated modest numerical improvements in discrimination compared with the core models, with limited statistical significance by AUROC (see Supplementary material online, Figure S2). In the primary test cohort, the extended AL ECHO model achieved higher discrimination than the core ECHO model (AUROC 0.910, 95% CI 0.889–0.929 vs. 0.868, 95% CI 0.846–0.890; P = 0.007). However, no statistically significant differences were observed between core and extended models for the combined ECHO + ECG architecture in AL amyloidosis, nor for either ECHO-only or ECHO + ECG models in ATTR amyloidosis. In the temporally distinct test cohort, extended models consistently yielded numerically higher AUROC values across all model architectures and amyloid subtypes, though none of these differences reached statistical significance. In contrast, PR analysis revealed more pronounced and consistent benefits of model extension, particularly in the temporally distinct test cohort (see Supplementary material online, Figure S3). In the primary test cohort, the extended AL ECHO model demonstrated significantly higher AUPRC than the core ECHO model (0.750, 95% CI 0.711–0.790 vs. 0.674, 95% CI 0.628–0.716; P = 0.013), whereas differences between core and extended models were not statistically significant for the combined ECHO + ECG architecture or for ATTR amyloidosis. In the temporally distinct test cohort, extended models demonstrated significantly higher AUPRC across all model architectures and both amyloid subtypes, including ECHO-only and combined ECHO + ECG models, indicating improved positive predictive performance in lower-prevalence populations. Discussion In this study, we developed and tested subtype-specific prediction models for AL and ATTR CA using routinely available echocardiographic variables and quantified the incremental value of incorporating an AI-derived ECG amyloidosis probability score. Across both subtypes, echocardiography alone provided strong discrimination; however, the combined ECHO + ECG approach consistently improved performance for ATTR amyloidosis. In AL amyloidosis, the combined model provided modest improvements that did not reach statistical significance. For both subtypes, the improvements by the combined model were apparent on PR analysis, which evaluates how positive predictive value changes with sensitivity and more directly reflects the downstream false-positive burden of screening in low-prevalence conditions. Together, these findings support a multimodal strategy in which echocardiographic features refine ECG-based AI algorithms, yielding more selective identification of patients at high risk for CA. Across both subtypes, the observed reduction in model performance in the temporally distinct cohort likely reflects clinically meaningful distribution shift rather than model overfitting. The temporally distinct test cohort was assembled from a later time period (2019–2025), and its baseline composition differs from the original test cohort in ways that directly impact discrimination. First, controls in the temporally distinct cohort are systematically older and more male-predominant, narrowing separation for demographic predictors that are explicitly incorporated in the core models and that serve as proxies for comorbidity burden and myocardial remodelling. Second, for several canonical amyloidosis features, case–control separation is attenuated in the temporally distinct cohort as reflected by smaller ASDs for wall thickness and relative wall thickness, reducing signal-to-noise along feature axes that drive echocardiography-based discrimination. As a result, the temporally distinct controls function as ‘harder’ negatives. They reflect contemporary referral and surveillance patterns and include a larger proportion of patients with structural heart disease and HFpEF phenocopies that overlap with amyloid-associated remodelling. This increased feature overlap reduces case–control separation and degrades model discrimination, even when the underlying disease biology is unchanged. Together, these factors represent expected temporal drift in real-world practice and provide a plausible explanation for the consistent decline in AUROC and AUPRC observed for all model architectures in the temporally distinct test cohort. A complementary insight from these analyses is that the relative contribution of ECG- and echocardiography-based models differs by amyloid subtype. In AL amyloidosis, much of the discriminatory signal appears to be captured by ECG-based features, with echocardiographic measures providing limited incremental value. In contrast, in ATTR amyloidosis, echocardiographic assessments of cardiac structure and function consistently contributed information beyond ECG alone. This subtype-specific pattern aligns with prior work, which has shown that ATTR is more readily identified using echocardiographic features.15 Because ATTR amyloidosis is characterized by more pronounced and homogeneous increases in ventricular wall thickness and myocardial mass, echocardiography-based models are more likely to capture disease-specific structural remodelling that complements ECG-derived risk estimates. Nonetheless, the incremental contribution of structured clinical variables to the baseline ECG model was modest, in particular for the amyloid subtype, but aligns with recent reports demonstrating limited added value when combining electronic health record-derived features with ECG-based deep learning models for structural heart disease detection.16 This likely reflects the fact that the ECG model captures not only changes in cardiac structures17 but also extra-cardiac disease18 and demographic signatures, including age- and sex19-related variation, that are embedded within the raw waveform. As a result, through learned latent representation, the ECG model alone may already account for clinically relevant comorbidities. Our findings are consistent with reports that CA screening models can show modest attenuation in discrimination when evaluated post-development or in operationally distinct samples. In a post-development evaluation of an AI-ECG model for CA, Harmon et al. observed lower performance in the post-development cohort compared with the derivation cohort, while month-to-month performance remained stable without evidence of a consistent temporal drift signal.20 This highlights how changes in case mix and subgroup composition can drive apparent performance differences even in the absence of overt model drift. Related multimodal work has also combined deep learning ECG waveforms with structured echocardiographic and demographic features, and in some implementations incorporated diagnostic codes. Jain et al.11 demonstrated that AI-augmented screening can yield a high diagnostic positivity rate among those who complete confirmatory testing, but that real-world impact is strongly shaped by downstream implementation factors, including clinician response and patient contact or uptake, which can lead to substantial drop-off between eligibility and completed scintigraphy. Nonetheless, our results reinforce the value of parsimonious models. We intentionally evaluated a core model composed of routinely available variables and an extended model incorporating additional selected predictors, enabling a realistic tradeoff between portability and maximal performance. While AUROC differences between core and extended models were generally modest, PR analyses suggested more consistent gains from model extension in lower prevalence settings, especially in the temporally distinct cohort, which may be relevant for implementation sites where data completeness supports richer feature sets. This study has several limitations. Although the validation cohort was multisite, it was drawn from a single integrated health system, which may limit transportability to external institutions with different referral patterns, imaging protocols, and reporting practices. In addition, the study population was predominantly male, which may restrict generalizability to women and could affect calibration in settings with a different sex distribution, despite this being directionally consistent with the epidemiology of ATTR cardiomyopathy in contemporary referral cohorts. A limitation of the control cohort is that systematic amyloid-specific testing was not performed; therefore, undiagnosed CA cannot be fully excluded. However, CA remains uncommon among patients undergoing echocardiography, even in tertiary-care echocardiography laboratories.21 Any resulting control misclassification would therefore likely be limited and would be expected to attenuate, rather than inflate, model discrimination. Finally, the echocardiographic models rely on structured variables derived from routine clinical interpretation rather than fully automated image-based feature extraction and therefore inherit real-world variability in measurement, reporting, and data completeness across laboratories. A further limitation relates to real-world implementation of subtype-specific models. AL and ATTR were modelled independently in this study, whereas clinical deployment will require concurrent use of both models. In that setting, performance depends not only on discrimination of each subtype from controls but also on how each model behaves in patients with the alternate subtype and how outputs are integrated into a single workflow. Cross-subtype discrimination and optimal decision rules for joint deployment were not fully addressed here, and operational thresholds are expected to be sensitive to disease prevalence and downstream testing strategy. Future work should focus on prospective validation, calibration, threshold selection, and workflow integration of AI-based screening for CA. The AI-ECG algorithm used in this study has been licensed and has received regulatory clearance; however, commercial availability and end-user deployment of that algorithm are outside the scope of the present manuscript. The PREDICT-AMY trial (Artificial Intelligence Enhanced ECG to Detect Cardiac Amyloidosis: Protocol for a Pragmatic Cluster Randomized Clinical Trial; ClinicalTrials.gov ID NCT05557162) prospectively evaluated the subtype-specific AL and ATTR ECHO and ECHO + ECG models described in this study, including core and extended model configurations, in a real-world clinical workflow. The trial has been completed, and the manuscript is in preparation. Supplementary Material