Ensemble Deep Learning Algorithm for Structural Heart Disease Screening Using Electrocardiographic Images

A key challenge of AI-ECG models in detecting specific cardiac disorders using ECGs is low precision driven by the low prevalence of individual disorders.

“A key challenge of AI-ECG models in detecting specific cardiac disorders using ECGs is the low precision driven by the low prevalence of individual disorders.10-12”

• cites [10] Attia, 2019, Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram, Nat Med sourceNot checked — no full text of the source
• cites [11] Sangha, 2023, Detection of Left Ventricular Systolic Dysfunction From Electrocardiographic Images, Circulation sourcePartly supported

  The Results (page unavailable) report a positive predictive value of 0.26 for the AI-ECG model detecting left ventricular systolic dysfunction. The Discussion (page unavailable) notes lower disorder prevalence in a healthier validation cohort, but does not say that low prevalence drove low precision or establish this as a general challenge across individual cardiac disorders. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [12] Cohen-Shelly, 2021, Electrocardiogram screening for aortic valve stenosis using artificial intelligence, Eur Heart J sourceNot checked — no full text of the source

The use of ECG images as input and a flexible strategy that allows for varying formats supports the scalability of the approach across resource settings.

“Moreover, the use of ECG images as the input, and a flexible strategy that allows for varying formats, supports the scalability of the approach across resources settings.38”

• cites [38] Oikonomou, 2024, Tracking the pre-clinical progression of transthyretin amyloid cardiomyopathy using artificial intelligence-enabled electrocardiography and echocardiography, bioRxiv sourceSupported

The low prevalence of individual structural heart diseases limits the potential implementation of models for broad screening because of the low PPVs of the proposed models.

“While these models provide a strong foundation for the role of ECG-based detection of SHDs, the low prevalence of these individual diseases, their potential implementation for broad screening is limited by the low PPVs of the proposed models.10-12,16,32-34”

• cites [10] Attia, 2019, Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram, Nat Med sourceNot checked — no full text of the source
• cites [11] Sangha, 2023, Detection of Left Ventricular Systolic Dysfunction From Electrocardiographic Images, Circulation sourcePartly supported

  The Results (Detection of LV Systolic Dysfunction; page unavailable) report a PPV of 0.26. The Discussion (page unavailable) notes lower disease prevalence in a healthier validation cohort and says screening thresholds may need recalibration. The paper does not conclude that low prevalence and PPV limit broad screening implementation; it calls for prospective assessment in that setting. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [12] Cohen-Shelly, 2021, Electrocardiogram screening for aortic valve stenosis using artificial intelligence, Eur Heart J sourceNot checked — no full text of the source
• cites [16] Sangha, 2023, Identification of hypertrophic cardiomyopathy on electrocardiographic images with deep learning, medRxiv sourcePartly supported

  The Discussion says community-based screening with this HCM model can produce high false-positive rates and that low prevalence limits screening yield. The Results report only 29 HCM cases among 45,389 UK Biobank ECGs. This supports the concern for HCM, but the paper does not establish low PPVs or screening limitations across multiple structural heart diseases and their proposed models. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [32] Ko, 2020, Detection of Hypertrophic Cardiomyopathy Using a Convolutional Neural Network-Enabled Electrocardiogram, J Am Coll Cardiol sourceSupported
• cites [33] Goto, 2022, Multinational federated learning approach to train ECG and echocardiogram models for hypertrophic cardiomyopathy detection, Circulation sourcePartly supported

  The Methods describe an HCM surveillance cohort with 0.5% prevalence, and the Results report a 1% PPV for the ECG model at a high-sensitivity cutoff. The Discussion says PPV affects which follow-up actions are appropriate, but it does not conclude that low PPV precludes broad screening; it proposes systematic ECG screening followed by echocardiography. The paper addresses HCM, not structural heart diseases generally. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [34] Grogan, 2021, Artificial intelligence–enhanced electrocardiogram for the early detection of cardiac amyloidosis, Mayo Clin Proc sourceNot checked — no full text of the source

Of 127,547 individuals at risk in YNHH, 5,346 (4.2%) had new-onset SHD/HF over a median of 4.0 years of follow-up.

“Of the 127,547 individuals at risk in YNHH, 5,346 (4.2%) had new-onset SHD/HF over a median of 4.0 (IQR 1.7-6.4) years of follow-up.”

PRESENT-SHD generalized to the external validation cohort at Lawrence + Memorial with an AUROC of 0.871 (95% CI, 0.864-0.878).

“Further, PRESENT-SHD generalized well to the external validation cohorts at Bridgeport (AUROC, 0.854 [95% CI, 0.847-0.862]), Greenwich (AUROC, 0.900 [95% CI, 0.888-0.913]), Lawrence + Memorial (AUROC, 0.871 [95% CI, 0.864-0.878]), and Westerly (AUROC, 0.887 [95% CI, 0.874-0.902]) Hospitals, with sensitivities and specificities ranging 88-96% and 51-66%, respectively.”

The model for LVSD had an AUROC of 0.914 (95% CI, 0.904-0.924).

“The models for LVSD, moderate or severe valvular disease, and sLVH had AUROCs of 0.914 (95% CI, 0.904-0.924), 0.805 (95% CI, 0.793-0.817), and 0.903 (95% CI, 0.850-0.946; Figure 2), respectively.”

A positive PRESENT-SHD screen portended a 4-fold higher risk of new-onset SHD/HF in YNHH (age- and sex-adjusted HR, 4.28; 95% CI, 3.95-4.64; Harrell's C-statistic, 0.823).

“A positive PRESENT-SHD screen portended a 4-fold higher risk of new-onset SHD/HF in YNHH (age- and sex-adjusted HR [aHR], 4.28 [95% CI, 3.95-4.64], Harrell’s C-statistic, 0.823 [95% CI, 0.817-0.828]) and every 10% increment in model probability was progressively associated with a 36% higher hazard for incident SHD/HF (aHR, 1.36 [1.35-1.38]).”

Across the external validation hospitals, sensitivities and specificities ranged 88-96% and 51-66%, respectively.

“Further, PRESENT-SHD generalized well to the external validation cohorts at Bridgeport (AUROC, 0.854 [95% CI, 0.847-0.862]), Greenwich (AUROC, 0.900 [95% CI, 0.888-0.913]), Lawrence + Memorial (AUROC, 0.871 [95% CI, 0.864-0.878]), and Westerly (AUROC, 0.887 [95% CI, 0.874-0.902]) Hospitals, with sensitivities and specificities ranging 88-96% and 51-66%, respectively.”

Across validation sites, PRESENT-SHD performance was consistent across demographic subgroups.

“Across validation sites, model performance was consistent across demographic subgroups (Supplementary Tables 12-16).”

At the probability threshold for sensitivity above 90% in the internal validation set, PRESENT-SHD had a PPV of 57.4% (95% CI, 56.1-58.6) for detecting SHD in the held-out test set in YNHH.

“At the probability threshold for sensitivity above 90% in the internal validation set, the model had a sensitivity of 89.8% (95% CI, 89.0-90.5), specificity of 66.2% (95% CI, 65.0-67.4), PPV of 57.4% (95% CI, 56.1-58.6), and NPV of 92.8% (95% CI, 92.1-93.4) for detecting SHD in the held-out test set in YNHH (Table 2; Supplementary Figure 7).”

Across ECG screenshots and smartphone photographs of monitors and printouts, PRESENT-SHD demonstrated high agreement with plotted images (Pearson correlation coefficients, 0.959-0.977).

“Across ECG screenshots and smartphone photographs of monitors and printouts, the model demonstrated high agreement with plotted images (Pearson correlation coefficients, 0.959-0.977; Supplementary Figure 9) and consistent performance across all image types (AUROCs: Plotted images, 0.939 [95% CI, 0.887-0.976]; ECG screenshots, 0.934 [95% CI, 0.885-0.973]; Smartphone photographs of computer monitors”

The CNN model for AS had an AUROC of 0.804 (95% CI, 0.784-0.824).

“The performance of CNN models for individual valvular heart diseases varied, with an AUROC of 0.722 (95% CI, 0.784-0.824) for moderate or severe AR, 0.804 (95% CI, 0.784-0.824) for AS, and 0.792 (95% CI, 0.776-0.807) for MR.”

Of 15,105 participants in ELSA-Brasil, 3,014 who underwent ECG and transthoracic echocardiogram during their baseline visit were included.

“Of the 15,105 participants in ELSA-Brasil, 3,014 who underwent ECG and TTE during their baseline visit were included.”

PRESENT-SHD had high discrimination for detecting the severe SHD phenotype (LVSD, severe left-sided valve disease, or sLVH).

“Moreover, the model had consistent performance across subsets where TTEs were performed before, on the same day as, or after the ECG (Supplementary Table 6) and had high discrimination for detecting the severe SHD phenotype (LVSD, severe left-sided valve disease, or sLVH; Supplementary Figure 8).”

Every 10% increment in PRESENT-SHD model probability was associated with a 36% higher hazard for incident SHD/HF (aHR, 1.36; 95% CI, 1.35-1.38).

“A positive PRESENT-SHD screen portended a 4-fold higher risk of new-onset SHD/HF in YNHH (age- and sex-adjusted HR [aHR], 4.28 [95% CI, 3.95-4.64], Harrell’s C-statistic, 0.823 [95% CI, 0.817-0.828]) and every 10% increment in model probability was progressively associated with a 36% higher hazard for incident SHD/HF (aHR, 1.36 [1.35-1.38]).”

In UKB, 413 (1.0%) of 41,800 individuals developed SHD/HF over 3.0 years of follow-up.

“In UKB, 413 (1.0%) of 41,800 individuals developed SHD/HF over 3.0 (IQR 2.1-4.5) years of follow-up.”

At the probability threshold for sensitivity above 90% in the internal validation set, PRESENT-SHD had an NPV of 92.8% (95% CI, 92.1-93.4) for detecting SHD in the held-out test set in YNHH.

“At the probability threshold for sensitivity above 90% in the internal validation set, the model had a sensitivity of 89.8% (95% CI, 89.0-90.5), specificity of 66.2% (95% CI, 65.0-67.4), PPV of 57.4% (95% CI, 56.1-58.6), and NPV of 92.8% (95% CI, 92.1-93.4) for detecting SHD in the held-out test set in YNHH (Table 2; Supplementary Figure 7).”

The model for sLVH had an AUROC of 0.903 (95% CI, 0.850-0.946).

“The models for LVSD, moderate or severe valvular disease, and sLVH had AUROCs of 0.914 (95% CI, 0.904-0.924), 0.805 (95% CI, 0.793-0.817), and 0.903 (95% CI, 0.850-0.946; Figure 2), respectively.”

Across all sites, higher PRESENT-SHD model probabilities were associated with progressively higher risk of new-onset SHD/HF.

“Across all sites, higher model probabilities were associated with progressively higher risk of new-onset SHD/HF (Supplementary Table 28; Figure 3).”

PRESENT-SHD had consistent performance across subsets where TTEs were performed before, on the same day as, or after the ECG.

“Moreover, the model had consistent performance across subsets where TTEs were performed before, on the same day as, or after the ECG (Supplementary Table 6) and had high discrimination for detecting the severe SHD phenotype (LVSD, severe left-sided valve disease, or sLVH; Supplementary Figure 8).”

The model for moderate or severe valvular disease had an AUROC of 0.805 (95% CI, 0.793-0.817).

“The models for LVSD, moderate or severe valvular disease, and sLVH had AUROCs of 0.914 (95% CI, 0.904-0.924), 0.805 (95% CI, 0.793-0.817), and 0.903 (95% CI, 0.850-0.946; Figure 2), respectively.”

The CNN model AUPRCs varied with individual disease prevalence.

“The CNN model AUPRCs varied with individual disease prevalence (Supplementary Tables 18-23).”

PRESENT-SHD achieved an AUROC of 0.886 (95% CI, 0.877-0.894) for detecting the composite SHD outcome in the held-out test set.

“The ensemble XGBoost model, PRESENT-SHD, demonstrated an AUROC of 0.886 (95% CI, 0.877-0.894) and an AUPRC of 0.807 (95% CI, 0.791-0.823) for the detection of the composite SHD outcome in the held-out test set (Table 1).”

PRESENT-SHD generalized to the external validation cohort at Bridgeport with an AUROC of 0.854 (95% CI, 0.847-0.862).

“Further, PRESENT-SHD generalized well to the external validation cohorts at Bridgeport (AUROC, 0.854 [95% CI, 0.847-0.862]), Greenwich (AUROC, 0.900 [95% CI, 0.888-0.913]), Lawrence + Memorial (AUROC, 0.871 [95% CI, 0.864-0.878]), and Westerly (AUROC, 0.887 [95% CI, 0.874-0.902]) Hospitals, with sensitivities and specificities ranging 88-96% and 51-66%, respectively.”

At the probability threshold for sensitivity above 90% in the internal validation set, PRESENT-SHD had a specificity of 66.2% (95% CI, 65.0-67.4) for detecting SHD in the held-out test set in YNHH.

“At the probability threshold for sensitivity above 90% in the internal validation set, the model had a sensitivity of 89.8% (95% CI, 89.0-90.5), specificity of 66.2% (95% CI, 65.0-67.4), PPV of 57.4% (95% CI, 56.1-58.6), and NPV of 92.8% (95% CI, 92.1-93.4) for detecting SHD in the held-out test set in YNHH (Table 2; Supplementary Figure 7).”

PRESENT-SHD achieved an AUPRC of 0.807 (95% CI, 0.791-0.823) for detecting the composite SHD outcome in the held-out test set.

“The ensemble XGBoost model, PRESENT-SHD, demonstrated an AUROC of 0.886 (95% CI, 0.877-0.894) and an AUPRC of 0.807 (95% CI, 0.791-0.823) for the detection of the composite SHD outcome in the held-out test set (Table 1).”

PRESENT-SHD generalized to the population-based ELSA-Brasil cohort with an AUROC of 0.853 (95% CI, 0.811-0.897).

“PRESENT-SHD also generalized well to the population-based ELSA-Brasil, with an AUROC of 0.853 (95% CI: 0.811-0.897) and a sensitivity and specificity of 87.5% and 61.9%, respectively (Table 2; Supplementary Table 11).”

PRESENT-SHD performance was higher than the CNN models trained to directly detect SHD and other XGBoost ensemble strategies.

“Notably, the performance of PRESENT-SHD was higher than the CNN models trained to directly detect SHD and other XGBoost ensemble strategies (Supplementary Tables 7 and 8).”

PRESENT-SHD generalized to the external validation cohort at Westerly with an AUROC of 0.887 (95% CI, 0.874-0.902).

“Further, PRESENT-SHD generalized well to the external validation cohorts at Bridgeport (AUROC, 0.854 [95% CI, 0.847-0.862]), Greenwich (AUROC, 0.900 [95% CI, 0.888-0.913]), Lawrence + Memorial (AUROC, 0.871 [95% CI, 0.864-0.878]), and Westerly (AUROC, 0.887 [95% CI, 0.874-0.902]) Hospitals, with sensitivities and specificities ranging 88-96% and 51-66%, respectively.”

PRESENT-SHD performed consistently across subgroups of age, sex, race, and ethnicity.

“PRESENT-SHD performed consistently across subgroups of age, sex, race, and ethnicity (Table 1), and generalized well to novel ECG formats not encountered during training (Supplementary Table 5).”

• Also stated in: § DISCUSSION · plain

  “Model performance was consistent across demographic subgroups and similar to the corresponding signal-based models.”

PRESENT-SHD had an AUROC of 0.853 (0.811-0.897), 88% sensitivity, and 62% specificity in ELSA-Brasil.

“The model generalized well to ELSA-Brasil (AUROC, 0.853 [0.811-0.897], 88% sensitivity, 62% specificity).”

PRESENT-SHD performance was similar to the corresponding signal-based model for detecting SHD.

“PRESENT-SHD performance was similar to the corresponding signal-based model for detecting SHD (Supplementary Table 9).”

• Also stated in: § DISCUSSION · plain

  “Model performance was consistent across demographic subgroups and similar to the corresponding signal-based models.”

The performance for individual disease CNNs was consistent across external validation cohorts and real-world ECG image modalities.

“The performance for individual disease CNNs was consistent across external validation cohorts (Supplementary Figure 10; Supplementary Tables 18-23) and real-world ECG image modalities (Supplementary Table 24).”

At the probability threshold for sensitivity above 90% in the internal validation set, PRESENT-SHD had a sensitivity of 89.8% (95% CI, 89.0-90.5) for detecting SHD in the held-out test set in YNHH.

“At the probability threshold for sensitivity above 90% in the internal validation set, the model had a sensitivity of 89.8% (95% CI, 89.0-90.5), specificity of 66.2% (95% CI, 65.0-67.4), PPV of 57.4% (95% CI, 56.1-58.6), and NPV of 92.8% (95% CI, 92.1-93.4) for detecting SHD in the held-out test set in YNHH (Table 2; Supplementary Figure 7).”

The association between PRESENT-SHD and incident SHD/HF remained consistent after adjusting for comorbidities at baseline and the competing risk of death.

“This association remained consistent after adjusting for comorbidities at baseline and the competing risk of death (Supplementary Table 26).”

PRESENT-SHD generalized to the external validation cohort at Greenwich with an AUROC of 0.900 (95% CI, 0.888-0.913).

“Further, PRESENT-SHD generalized well to the external validation cohorts at Bridgeport (AUROC, 0.854 [95% CI, 0.847-0.862]), Greenwich (AUROC, 0.900 [95% CI, 0.888-0.913]), Lawrence + Memorial (AUROC, 0.871 [95% CI, 0.864-0.878]), and Westerly (AUROC, 0.887 [95% CI, 0.874-0.902]) Hospitals, with sensitivities and specificities ranging 88-96% and 51-66%, respectively.”

A similar pattern of association between PRESENT-SHD and incident SHD/HF was observed across all external validation hospital sites.

“A similar pattern was observed across all external validation hospital sites (Supplementary Tables 26 and 27).”

In UKB, a positive versus negative PRESENT-SHD screen was associated with twice the hazard of developing SHD/HF (aHR, 2.39; 95% CI, 1.87-3.04; Harrell's C-statistic, 0.754).

“In the UKB, a positive vs. negative PRESENT-SHD screen was associated with twice the hazard of developing SHD/HF (aHR, 2.39 [95% CI, 1.87-3.04], Harrell’s C-statistic, 0.754 [95% CI, 0.728-0.780]).”

PRESENT-SHD generalized well to novel ECG formats not encountered during training.

“PRESENT-SHD performed consistently across subgroups of age, sex, race, and ethnicity (Table 1), and generalized well to novel ECG formats not encountered during training (Supplementary Table 5).”

• Also stated in: § DISCUSSION · plain

  “Additionally, PRESENT-SHD maintained high performance when tested on novel ECG formats, screenshots of ECGs in the EHR, as well as smartphone photographs of ECGs captured from laptop monitors and printouts.”

PRESENT-SHD model performance was lower in individuals aged 65 and older.

“Fifth, model performance was lower in individuals aged 65 and older, potentially limiting reliability as a standalone tool to rule out the need for cardiac imaging.”

The CNN model for MR had an AUROC of 0.792 (95% CI, 0.776-0.807).

“The performance of CNN models for individual valvular heart diseases varied, with an AUROC of 0.722 (95% CI, 0.784-0.824) for moderate or severe AR, 0.804 (95% CI, 0.784-0.824) for AS, and 0.792 (95% CI, 0.776-0.807) for MR.”

The CNN model for moderate or severe AR had an AUROC of 0.722 (95% CI, 0.784-0.824).

“The performance of CNN models for individual valvular heart diseases varied, with an AUROC of 0.722 (95% CI, 0.784-0.824) for moderate or severe AR, 0.804 (95% CI, 0.784-0.824) for AS, and 0.792 (95% CI, 0.776-0.807) for MR.”

All structural heart disease conditions were ascertained based on interpretation of transthoracic echocardiograms by board-certified cardiologists according to American Society of Echocardiography guidelines.

“All conditions were ascertained based on the interpretation of TTE by board-certified cardiologists according to the American Society of Echocardiography guidelines.20”

• cites [20]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Any processing of ECG images is not required for the real-world application of PRESENT-SHD.

“Thus, any processing of ECG images is not required for the real-world application of PRESENT-SHD, as also demonstrated in the publicly accessible application of the model.”

• AssumptionRests on an assumption: Signal preprocessing before image plotting represents standard steps used in ECG machines before ECG images are generated or printed.

By focusing on detecting any clinically relevant SHDs that require TTE for confirmation, PRESENT-SHD enables efficient screening by limiting false discovery.

“Through a focus on detecting any of the clinically relevant SHDs that require TTE for confirmation, PRESENT-SHD enables efficient screening by limiting false discovery.”

None of the ECGs on the self-supervised pretraining task represented individuals in the structural heart disease model development.

“None of the ECGs on the self-supervised pretraining task represented individuals in the SHD model development.”

A focus on developing PRESENT-SHD in diverse populations enabled its consistent performance in demographic subgroups across validation sites.

“A focus on developing PRESENT-SHD in diverse populations enabled its consistent performance in demographic subgroups across validation sites.”

PRESENT-SHD is an ensemble deep learning model that uses an ECG image as the input to detect a range of SHDs.

“We developed and validated PRESENT-SHD, an ensemble deep learning model that uses an ECG image as the input to detect a range of SHDs.”

The use of raw ECG voltage data precludes widespread use of AI-ECG approaches for broad cardiovascular screening because data integrations are not commonly available.

“This precludes the widespread use of AI-ECG approaches for broad cardiovascular screening, as these data integrations are not commonly available.”

PRESENT-SHD is an AI-ECG tool that identifies a range of structural heart diseases using images of 12-lead ECGs.

“We developed and validated PRESENT-SHD, an AI-ECG tool identifying a range of SHD using images of 12-lead ECGs, representing a robust, scalable, and accessible modality for automated SHD screening and risk stratification.”

Given that the individual components of SHD share a common diagnostic test, TTE, screening with PRESENT-SHD can help triage the use of TTE testing.

“Given that the individual components of SHD share a common diagnostic test, a TTE, screening with PRESENT-SHD can help triage the use of TTE testing.”

None of the patients in the external validation sets were in the model development population.

“Of note, none of the patients in the external validation sets were in the model development population.”

The focus on a composite model that detects a broad range of SHDs simultaneously reduces the burden of false positive screens and downstream testing, which is a major concern for AI-ECG models developed for individual cardiovascular conditions.

“The focus on a composite model that detects a broad range of SHDs simultaneously reduces the burden of false positive screens and downstream testing, which is a major concern for AI-ECG models developed for individual cardiovascular conditions.”

An AI-ECG-based approach that leverages ECG images and photographs can enable opportunistic screening through automated deployment across clinical settings where ECGs are obtained.

“Since early disease detection and intervention can alter the trajectory and outcomes of patients with SHDs, an AI-ECG-based approach that leverages ECG images and photographs can enable opportunistic screening through automated deployment across clinical settings where ECGs are obtained.42,43”

• cites [42] Levin, 1996, Prevalent left ventricular hypertrophy in the predialysis population: identifying opportunities for intervention, Am J Kidney Dis sourceNot checked — no full text of the source
• cites [43] Topol, 2024, for medical scansNot checked yet

  Not checked in the free run.

In individuals without SHD at baseline, PRESENT-SHD identified those with an up to 4-fold higher risk of developing new-onset SHD/HF across both US health system-centered cohorts and a UK community-based cohort.

“Further, in individuals without SHD at baseline, PRESENT-SHD identified those with an up to 4-fold higher risk of developing new-onset SHD/HF, across both health system-centered cohorts in the US and in a community-based cohort in the UK.”

PRESENT-SHD was well calibrated to the risk of new-onset disease, such that higher PRESENT-SHD probabilities were associated with progressively higher risk of developing SHD/HF.

“The model was well calibrated to the risk of new-onset disease, such that higher PRESENT-SHD probabilities were associated with progressively higher risk of developing SHD/HF.”

PRESENT-SHD generalized well to ELSA-Brasil.

“The model generalized well to ELSA-Brasil (AUROC, 0.853 [0.811-0.897], 88% sensitivity, 62% specificity).”

Lower PRESENT-SHD performance in individuals aged 65 and older potentially limits its reliability as a standalone tool to rule out the need for cardiac imaging.

“Fifth, model performance was lower in individuals aged 65 and older, potentially limiting reliability as a standalone tool to rule out the need for cardiac imaging.”

The development population represented a selected set of patients with a clinical indication for an ECG and a TTE.

“First, the development population represented a selected set of patients with a clinical indication for an ECG and a TTE.”

The consistent validation of PRESENT-SHD across populations with a broad range of clinical subpopulations seen in community and referral hospitals suggests that the model learned generalizable signatures of the SHDs.

“The consistent validation of the model across populations with a broad range of clinical subpopulations seen in community as well as referral hospitals suggests that the model learned generalizable signatures of the SHDs.”