CoChat
Sign Up
Cited by 95ECG Monitoring and AnalysisNon-Invasive Vital Sign MonitoringCardiac electrophysiology and arrhythmias

Detection of left ventricular systolic dysfunction from single-lead electrocardiography adapted for portable and wearable devices

Open access
Khunte, Akshay, Sangha, Veer, Oikonomou, Evangelos K., Dhingra, Lovedeep S., Aminorroaya, Arya, Mortazavi, Bobak J., Coppi, Andreas, Brandt, Cynthia A. +2 more·2023·npj Digital Medicine
DOI 10.1038/s41746-023-00869-wPMC PMC10336107

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Artificial intelligence (AI) can detect left ventricular systolic dysfunction (LVSD) from electrocardiograms (ECGs). Wearable devices could allow for broad AI-based screening but frequently obtain noisy ECGs. We report a novel strategy that automates the detection of hidden cardiovascular diseases, such as LVSD, adapted for noisy single-lead ECGs obtained on wearable and portable devices. We use 385,601 ECGs for development of a standard and noise-adapted model. For the noise-adapted model, ECGs are augmented during training with random gaussian noise within four distinct frequency ranges, each emulating real-world noise sources. Both models perform comparably on standard ECGs with an AUROC of 0.90. The noise-adapted model performs significantly better on the same test set augmented with four distinct real-world noise recordings at multiple signal-to-noise ratios (SNRs), including noise isolated from a portable device ECG. The standard and noise-adapted models have an AUROC of 0.72 and 0.87, respectively, when evaluated on ECGs augmented with portable ECG device noise at an SNR of 0.5. This approach represents a novel strategy for the development of wearable-adapted tools from clinical ECG repositories.

AI snapshot

Written from the paper’s full text.

TL;DR

A noise-adapted deep-learning model detects left ventricular systolic dysfunction (LVSD) from single-lead ECG and remains more robust to real-world noise than a standard model. It was developed on 385,601 ECG–echocardiogram pairs and tested on noise-augmented ECGs, including noise isolated from a portable ECG device.

Key findings

  • On the clean held-out test set, the noise-adapted and standard models performed similarly: AUROC 0.90 (95% CI 0.89–0.91) vs 0.90 (95% CI 0.88–0.91), p = 0.60; AUPRC 0.46 vs 0.48.
  • With portable ECG device noise at an SNR of 0.5, the noise-adapted model had AUROC 0.87 (95% CI 0.86–0.88) vs the standard model’s 0.72 (95% CI 0.71–0.74), p < 0.001.
  • Across MIT-BIH noises at SNRs 0.5–2.0, the noise-adapted model had AUROC ranges of 0.86–0.89 for electrode motion, 0.87–0.89 for muscle artifact, and 0.88–0.89 for baseline wander; the standard model was lower at 0.79–0.86, 0.81–0.86, and 0.80–0.86, respectively.
  • Last-layer prediction shifts were smaller for the noise-adapted model: for portable ECG noise, scaled Euclidean distance was 0.41 (95% CI 0.40–0.42) vs 0.50 (95% CI 0.49–0.51); for baseline wander, 0.36 (95% CI 0.36–0.37) vs 0.52 (95% CI 0.51–0.53), both p < 0.001.

Methods

  • Retrospective cohort of 116,210 patients aged 18 or older with clinically indicated ECG and paired transthoracic echocardiogram within 15 days at Yale-New Haven Hospital, 2015–2021; 385,601 ECG–TTE pairs were used, with Lead I isolated from 12-lead ECGs and patient-level splits of 85% training, 5% validation, and 10% test.
  • The noise-adapted model was trained with random Gaussian noise in four frequency bands (3–12, 12–50, 50–100, and 100–150 Hz) at SNRs of 0.50, 0.75, 1.00, and 1.25; the standard model was trained on original ECGs. Both used a CNN with seven 2D convolutional layers and dense layers, targeting LVEF < 40%.
  • Evaluation used the held-out test set without noise and augmented with MIT-BIH electrode motion, muscle artifact, and baseline wander noise, plus noise extracted from a KardiaMobile 6L portable ECG device, across multiple SNRs; analyses included AUROC, AUPRC, sensitivity, specificity, PPV, NPV, DeLong’s test, UMAP, and scaled Euclidean distances.

Limitations

  • The paper states the model was developed in patients with clinically indicated ECG and echocardiogram, not the intended broader screening population, so prospective assessment is warranted; performance may vary by LVSD severity and near the LVEF 40% cutoff, and echocardiographic LVEF is less precise than MRI.
  • The paper also states it cannot ascertain whether the model maintains performance on every possible wearable noise type, magnitude, or device-specific noise signature.
  • Evident from the text: there was no external validation cohort, the data cannot be shared publicly, and training/evaluation relied on synthetic Gaussian noise bands and selected real-world noise recordings rather than actual wearable ECG datasets.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41746-023-00869-w

Cites

36 of 36 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

20 of 42 cited sources checked · 4 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    Models trained on clinical ECGs have been applied to wearable device ECGs, but have traditionally shown significantly lower performance on wearable device ECGs than on held-out clinical ECG test sets.

    “Models trained on clinical ECGs have been applied to wearable device ECGs, but have traditionally shown significantly lower performance on wearable device ECGs than on held-out clinical ECG test sets14,15.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [14] Bachtiger, P. et al. Point-of-care screening for heart failure with reduced ejection fraction using artificial intelligence during ECG-enabled stethoscope examination in London, UK: a prospective, observational, multicentre study. Lancet Digit. Health 4, e117–e125 (2022). source

      The Methods (AI-ECG algorithm; page unavailable) describe a model trained on clinical 12-lead ECG data and tested using single-lead recordings from an ECG-enabled stethoscope. The Discussion (page unavailable) reports an internal 12-lead validation AUROC of 0.93, but says the study’s single-lead findings compare favourably with an external 12-lead validation AUROC of 0.82. The paper does not establish a traditional, statistically significant performance drop on wearable-device ECGs relative to held-out clinical ECG test sets. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [15] Attia, Z. I. et al. Automated detection of low ejection fraction from a one-lead electrocardiogram: application of an AI algorithm to an electrocardiogram-enabled Digital Stethoscope. Eur. Heart J. Digit. Health https://doi.org/10.1093/ehjdh/ztac030 (2022). source

      The Methods (Construction of neural network for use on single-lead electrocardiogram; page unavailable) confirm that a model trained on clinical 12-lead ECGs was adapted for single-lead use and report an AUC of 0.90 on the original clinical test set. The Results (Accuracy of single-lead digital stethoscope-enabled electrocardiogram; page unavailable) report an AUC of 0.91 for the optimal stethoscope ECG at the same EF threshold—not significantly lower performance. This study does not establish a broader historical trend. (LLM, conf=0.86)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
  • Premisecitation

    Models which have successfully generalized to wearable device ECGs have relied on automated algorithms to select high-quality subsets of available ECG data, with reduced performance when evaluated on all collected ECGs.

    “Additionally, models which have successfully generalized to wearable device ECGs have relied on automated algorithms to select high-quality subsets of available ECG data, with reduced performance when evaluated on all collected ECGs15,26.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [15] Attia, Z. I. et al. Automated detection of low ejection fraction from a one-lead electrocardiogram: application of an AI algorithm to an electrocardiogram-enabled Digital Stethoscope. Eur. Heart J. Digit. Health https://doi.org/10.1093/ehjdh/ztac030 (2022). source

      The Methods section describes an automated algorithm that selects the highest-quality ECG recording from those acquired with a digital stethoscope. The Results section reports higher AUCs for that selection than for scores averaged across all leads. This supports the comparison for this device, but not the broader claim that wearable-device ECG models generally rely on such selection or that performance was evaluated on all collected ECGs without selection. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [26] Attia, Z. I. et al. Prospective evaluation of smartwatch-enabled detection of left ventricular dysfunction. Nat. Med. 28, 2497–2503 (2022). source

      The Discussion (page not available) reports that an automated filter selected clean smartwatch ECG tracings and improved AUC to 0.88, compared with 0.85 for the ECG closest in time to the echocardiogram. That supports the benefit of quality selection, but the lower-performance comparison was not an evaluation on all collected ECGs. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • Premisecitation

    The reliability of AI-based screening tools for LVSD is limited by the presence of noise in data collected from wearable and handheld devices.

    “but the reliability of such tools is limited by the presence of noise in data collected from wearable and handheld devices12,13.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [12] Dörr, M. et al. The WATCH AF Trial: SmartWATCHes for detection of atrial fibrillation. JACC Clin. Electrophysiol. 5, 199–208 (2019). source

      The WATCH AF discussion reports that insufficient wrist-signal quality prevented the smartwatch algorithm from producing an AF diagnosis in 21.8% of subjects, and the Study Limitations section identifies a worsening signal-to-noise ratio as a potential problem. This supports a signal-quality limitation for wearable AF screening, but the paper does not establish the claim for AI-based LVSD screening. Page numbers are not available in the supplied passages. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [13] Mannhart, D. et al. Clinical validation of 5 direct-to-consumer wearable smart devices to detect atrial fibrillation: BASEL wearable study. JACC Clin. Electrophysiol. 9, 232–242 (2023). source

      The Discussion says technical artefacts, including noise, affected the devices’ automated interpretation of single-lead ECGs. The study evaluates atrial fibrillation detection, however, not AI-based screening for left ventricular systolic dysfunction (LVSD), so it does not establish that limitation for LVSD tools. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • Premisecitation

    Models demonstrate poorer performance when tested on all available wearable ECG data as opposed to selected high-quality subsets.

    “as models demonstrate poorer performance when tested on all available wearable ECG data as opposed to selected high-quality subsets15.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [15] Attia, Z. I. et al. Automated detection of low ejection fraction from a one-lead electrocardiogram: application of an AI algorithm to an electrocardiogram-enabled Digital Stethoscope. Eur. Heart J. Digit. Health https://doi.org/10.1093/ehjdh/ztac030 (2022). source
  • Premisecitation

    Early diagnosis can effectively mitigate the risk of LVSD.

    “Early diagnosis can effectively mitigate this risk2–4”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [2] Srivastava, P. K. et al. Heart failure hospitalization and guideline-directed prescribing patterns among heart failure with reduced ejection fraction patients. JACC Heart Fail 9, 28–38 (2021). source

      The Discussion (page unavailable) examines prescribing changes after heart-failure hospitalization in patients who already have reduced ejection fraction. Searches for early diagnosis, detection, and prevention found no passage saying that early diagnosis mitigates the risk of LVSD. (LLM, conf=0.73)

      Experimental — accuracy not yet measured
    • cites [3] Wolfe, N. K., Mitchell, J. D. & Brown, D. L. The independent reduction in mortality associated with guideline-directed medical therapy in patients with coronary artery disease and heart failure with reduced ejection fraction. Eur. Heart J. Qual Care Clin Outcomes. 7, 416–421 (2021). source

      The Introduction describes a study of guideline-directed medical therapy and mortality in patients who already had coronary artery disease and reduced ejection fraction. The Clinical implications section discusses early initiation of therapy, not early diagnosis or a reduced risk of developing LVSD. Searches for early diagnosis and LVSD prevention found no passage addressing the claimed effect. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
    • cites [4] Heidenreich, P. A. et al. 2022 AHA/ACC/HFSA guideline for the management of heart failure: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 145, e895–e1032 (2022). source
  • Classified as probably not a substantial claim. 21
  • May combine two claims. 18

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    LVEF < 40% is a threshold present in most heart failure diagnosis guidelines and consistent with prior work.

    “The cutoff for low LVEF was set as LVEF < 40%, a threshold present in most heart failure diagnosis guidelines4, and consistent with prior work in this space8,22.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [4] Heidenreich, P. A. et al. 2022 AHA/ACC/HFSA guideline for the management of heart failure: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 145, e895–e1032 (2022). source
    • cites [8] Attia, Z. I. et al. Screening for cardiac contractile dysfunction using an artificial intelligence–enabled electrocardiogram. Nat. Med. 25, 70–74 (2019). source
    • cites [22] Sangha, V. et al. Detection of left ventricular systolic dysfunction from electrocardiographic images. Preprint at https://www.medrxiv.org/content/10.1101/2022.06.04.22276000v2 (2022). source

  • May combine two claims.

    Premisecitation

    Studies have shown high rates of poor signal quality among data recorded by wearable devices, and significant variation in quality across different manufacturers.

    “Studies have shown high rates of poor signal quality among data recorded by wearable devices, and significant variation in quality across different manufacturers13.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [13] Mannhart, D. et al. Clinical validation of 5 direct-to-consumer wearable smart devices to detect atrial fibrillation: BASEL wearable study. JACC Clin. Electrophysiol. 9, 232–242 (2023). source

      The Discussion reports that 17%–26% of recordings were labeled inconclusive by device algorithms and says these rates differed significantly across devices. But inconclusive does not necessarily mean poor signal quality: the Results say cardiologists could interpret 95% of those tracings, and only 1% of 825 recordings were noninterpretable because of low quality or noise. The paper therefore supports variation in algorithm outcomes, not high rates of poor-quality signals across manufacturers. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Echocardiography LVEF measurement is less precise relative to other approaches such as magnetic resonance imaging.

    “This might also be attributable to a lack of precision in LVEF measurement by echocardiography, which has shown to be less precise relative to other approaches, such as magnetic resonance imaging27,28.”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [27] Pellikka, P. A. et al. Variability in ejection fraction measured by echocardiography, gated single-photon emission computed tomography, and cardiac magnetic resonance in patients with coronary artery disease and left ventricular dysfunction. JAMA Netw. Open 1, e181456 (2018). source

      The Results (Comparison of Imaging Modalities; page unavailable) show greater disagreement between echocardiography and CMR than between CMR and SPECT, supporting a concern about variability in echocardiographic LVEF measurements. But the Discussion and Limitations say there is no gold standard and do not establish that echocardiography itself is less precise than CMR; the study primarily compares agreement between modalities. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [28] Jenkins, C. et al. Left ventricular volume measurement with echocardiography: a comparison of left ventricular opacification, three-dimensional echocardiography, or both with magnetic resonance imaging. Eur. Heart J. 30, 98–106 (2009). source

      The Introduction and ‘Limitations of each technique’ sections (page not specified) describe poor reproducibility and observer variability for non-contrast 2D echocardiography, supporting the concern about its precision. But the Introduction also says contrast enhancement can bring inter-observer reproducibility to a level comparable with MRI, so the claim is too broad for echocardiography generally. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    LVEF below 40% is a threshold present in most heart failure diagnosis guidelines.

    “Both models were trained to detect LVEF below 40%, a threshold present in most heart failure diagnosis guidelines4, and consistent with prior work8,22.”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites [4] Heidenreich, P. A. et al. 2022 AHA/ACC/HFSA guideline for the management of heart failure: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 145, e895–e1032 (2022). source
    • cites [8] Attia, Z. I. et al. Screening for cardiac contractile dysfunction using an artificial intelligence–enabled electrocardiogram. Nat. Med. 25, 70–74 (2019). source
    • cites [22] Sangha, V. et al. Detection of left ventricular systolic dysfunction from electrocardiographic images. Preprint at https://www.medrxiv.org/content/10.1101/2022.06.04.22276000v2 (2022). source

  • May combine two claims.

    Findingown data

    The noise-adapted model had specificity and sensitivity of 0.68 and 0.92, respectively, and a PPV and NPV of 0.20 and 0.99, respectively, in the held-out test set.

    “the noise-adapted model had specificity and sensitivity of 0.68 and 0.92, respectively, and a PPV and NPV of 0.20 and 0.99, respectively, in the held-out test set.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The standard model had lower performance across all SNRs for every noise combination, with AUROC between 0.77-0.85, 0.77-0.85, and 0.78-0.86 for portable ECG device noise combined with electrode motion, muscle artifact, and baseline wander noise, respectively.

    “The standard model had lower performance across all SNRs for every noise combination, with AUROC between 0.77–0.85, 0.77–0.85, and 0.78–0.86 for portable ECG device noise combined with electrode motion, muscle artifact, and baseline wander noise, respectively (Table S3).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For portable ECG noise at SNR 0.5, the average scaled Euclidean distance was 0.50 for the standard model and 0.41 for the noise-adapted model.

    “In the ECG with noise derived from portable ECGs, for instance, the average scaled Euclidean distance for the standard model was 0.50 (95% CI 0.49–0.51) and for the noise-adapted model was 0.41 (95% CI 0.40–0.42).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    The noise-adapted approach resulted in selective removal of noise from signal, even for noises the model had not encountered before, while preserving the model's robustness in discerning complex hidden labels.

    “Compared with the current approach of developing models, our noise-adapted approach resulted in selective removal of noise from signal, even for noises the model had not encountered before, while preserving the model’s robustness in discerning complex hidden labels.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For the standard model, predictions for noised ECGs were visually distinct from those of standard ECGs in UMAP projections, whereas for the noise-adapted model there was no visual separation.

    “For the standard model, the predictions for each of the noised ECGs were visually distinct from those of the standard ECGs, despite being for the same set of 1000 ECGs, and differing only on the added noise. However, for the noise-adapted model, there was no visual separation in the model’s predictions between standard and noised ECGs”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For baseline wander at SNR 0.5, the average scaled Euclidean distance was 0.52 for the standard model and 0.36 for the noise-adapted model.

    “Similarly for the baseline wander it was 0.52 (0.51–0.53) and 0.36 (0.36–0.37), respectively (Table 3).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    AUPRC on the held-out test set without added noise was 0.46 for the noise-adapted model and 0.48 for the standard model.

    “Area under the precision-recall curve (AUPRC) on this held-out test set was 0.46 and 0.48, respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For the noise-adapted model, performance was comparable across all SNRs for combinations of portable ECG device noise and electrode motion, muscle artifact, and baseline wander noise with AUROC between 0.86-0.89, 0.87-0.89, and 0.88-0.89, respectively, for SNRs from 0.5 to 2.0.

    “For the noise-adapted model, model performance was comparable across all SNRs for the combinations of portable ECG device noise and electrode motion, muscle artifact, and baseline wander noise with AUROC between 0.86–0.89, 0.87–0.89, and 0.88–0.89, respectively, for SNRs from 0.5 to 2.0.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The standard model had lower performance across all SNRs for every MIT-BIH noise, with AUROC between 0.79-0.86, 0.81-0.86, and 0.80-0.86 for electrode motion, muscle artifact, and baseline wander noise, respectively.

    “The standard model had lower performance across all SNRs for every noise, with AUROC between 0.79–0.86, 0.81–0.86, and 0.80–0.86 for electrode motion, muscle artifact, and baseline wander noise, respectively (Table 2 and Fig. 1).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The standard model had sensitivity and specificity of 0.69 and 0.91, respectively, and a PPV and NPV of 0.21 and 0.99, respectively.

    “The standard model had sensitivity and specificity of 0.69 and 0.91, respectively, and a PPV and NPV of 0.21 and 0.99, respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For the noise-adapted model, performance was comparable across all SNRs for each MIT-BIH noise (electrode motion, muscle artifact, baseline wander) with AUROC between 0.86-0.89, 0.87-0.89, and 0.88-0.89 respectively for SNRs from 0.5 to 2.0.

    “For the noise-adapted model, model performance was comparable across all SNRs for each MIT-BIH noise with AUROC between 0.86–0.89, 0.87–0.89, and 0.88–0.89 for electrode motion, muscle artifact, and baseline wander noise, respectively, for SNRs from 0.5 to 2.0.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The selected architecture yielded the highest AUROC on the validation set with the fewest number of parameters for the standard model.

    “We selected the architecture yielding the highest area under the receiver operator characteristic (AUROC) curve on the validation set with the fewest number of parameters for the standard model.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    The approach suggests a novel approach for the development of wearable-adapted tools from clinical ECG repositories.

    “suggesting a novel approach for the development of wearable-adapted tools from clinical ECG repositories.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Threshold values were chosen based on cutoffs that achieved sensitivity of 0.90 in validation data.

    “chose threshold values based on cutoffs that achieved sensitivity of 0.90 in validation data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The noise-adapted deep-learning algorithm accurately identifies LV systolic dysfunction from single-lead ECG data and is resilient to significant noisy artifacts, despite not encountering the specific noises in the model development process.

    “we developed a noise-adapted deep-learning algorithm that accurately identifies LV systolic dysfunction from single-lead ECG data and was resilient to significant noisy artifacts, despite not encountering the specific noises in the model development process.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    The noise-adapted approach defines a novel paradigm on how to build robust, wearable-ready, single-lead ECG cardiovascular screening models from clinical ECG repositories.

    “The noise-adapted approach defines a novel paradigm on how to build robust, wearable-ready, single-lead ECG cardiovascular screening models from clinical ECG repositories”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The MIT-BIH dataset noises represent three types of noises frequently encountered in ECGs: baseline wander, muscle artifact, and electrode motion artifact.

    “The MIT-BIH dataset noises, each obtained at a sampling frequency of 360 Hz, represent three types of noises frequently encountered in ECGs: baseline wander noise, a low-frequency noise produced by lead or subject movement17, muscle artifact, caused by muscle contractions21, and electrode motion artifact, which is caused by irregular movement of the electrodes during ECG recordings17,21.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [17] Moody, G. B., Muldrow, W. & Mark, R. A noise stress test for arrhythmia detectors. Comput. Cardiol. 11, 381–384 (1984).

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

    • cites [21] Friesen, G. M. et al. A comparison of the noise sensitivity of nine QRS detection algorithms. IEEE Trans. Biomed. Eng. 37, 85–98 (1990). source

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Echocardiograms were evaluated by expert cardiologists.

    “These echocardiograms had been evaluated by expert cardiologists, and the LVEF defined in their interpretation was identified.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    The current 12-lead ECG-based models are limited to investments by health systems to incorporate tools into digital ECG repositories, and thereby limited to individuals who already seek care in those systems.

    “The current 12-lead ECG-based models are limited to investments by health systems to incorporate tools into digital ECG repositories, and thereby limited to individuals who already seek care in those systems.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    100-150 Hz contains higher-frequency muscle noises.

    “the 12–50 Hz and 100–150 Hz ranges contain the lower and higher-frequency muscle noises, respectively17,21.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [17] Moody, G. B., Muldrow, W. & Mark, R. A noise stress test for arrhythmia detectors. Comput. Cardiol. 11, 381–384 (1984).

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

    • cites [21] Friesen, G. M. et al. A comparison of the noise sensitivity of nine QRS detection algorithms. IEEE Trans. Biomed. Eng. 37, 85–98 (1990). source

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Even this technology may not be available or cost-effective for smaller hospitals and clinics with limited access to digital ECGs.

    “In addition to the clinically indicated ECGs limiting the scope of screening, even this technology may not be available or cost-effective for smaller hospitals and clinics with limited access to digital ECGs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    The study was reviewed and approved by the Yale Institutional Review Board, which waived the need for informed consent.

    “The study was reviewed by the Yale Institutional Review Board, which approved the study protocol and waived the need for informed consent as the study represents secondary analysis of existing data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Model performance may vary by the severity of LV systolic dysfunction.

    “Second, the model performance may vary by the severity of LV systolic dysfunction.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Model performance among patients with an LVEF near 40% may differ compared with those with LVEF significantly higher or lower than 40%.

    “it is possible that the model performance among patients with an LVEF near to this cutoff differed compared with those individuals with LVEF significantly higher or lower than 40%.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    A model which generalizes to ECGs with varying types and levels of noise may be necessary for implementation in community-wide screening settings.

    “A model which generalizes to ECGs with varying types and levels of noise may be necessary for implementation in such settings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    50-100 Hz reflects consistent electrode contact noise.

    “The 50–100 Hz domain reflects consistent electrode contact noise21”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [21] Friesen, G. M. et al. A comparison of the noise sensitivity of nine QRS detection algorithms. IEEE Trans. Biomed. Eng. 37, 85–98 (1990). source

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    Model weights were calculated using the effective number of samples class re-weighting scheme so that learning was not affected by the lower frequency of LVEF < 40%.

    “Model weights were calculated for the loss function such that learning was not affected by the lower frequency of LVEF < 40% compared to the incidence of LVEF ≥ 40% using the effective number of samples class re-weighting scheme33 (Fig. S2).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [33] Cui, Y., Jia, M., Lin, T.-Y., Song, Y. & Belongie, S. Class-balanced loss based on effective number of samples. https://arxiv.org/abs/1901.05555 (2019). source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    12-50 Hz contains lower-frequency muscle noises.

    “the 12–50 Hz and 100–150 Hz ranges contain the lower and higher-frequency muscle noises, respectively17,21.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [17] Moody, G. B., Muldrow, W. & Mark, R. A noise stress test for arrhythmia detectors. Comput. Cardiol. 11, 381–384 (1984).

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

    • cites [21] Friesen, G. M. et al. A comparison of the noise sensitivity of nine QRS detection algorithms. IEEE Trans. Biomed. Eng. 37, 85–98 (1990). source

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    The approach has significant potential to expand the screening of LV structural cardiac disorders to low-resource settings with limited access to hospital-grade equipment.

    “with significant potential to expand the screening of LV structural cardiac disorders to low-resource settings with limited access to hospital-grade equipment.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    Wearable devices allow obtaining ECGs that are more accessible and allow for community-wide screening.

    “Wearable devices allow obtaining ECGs that are more accessible and allow for community-wide screening, an important next step in the early detection of common and rare cardiomyopathies.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The approach avoids the unnecessary exclusion of collected data, which increases its generalizability across different device platforms.

    “Additionally, this approach avoids the unnecessary exclusion of collected data, which increases its generalizability across different device platforms”
    Quote in the paperSubstantial claim
35 of 35

  • The paper relies on this without support.

    There is a lack of publicly available wearable device ECG datasets, making training models using wearable ECG data directly challenging.

    “Due to the lack of publicly available wearable device ECG datasets, training models using wearable ECG data directly is challenging.”

  • The paper relies on this without support.

    In the absence of large, labelled datasets of wearable ECGs, the development of algorithms that can detect underlying structural heart disease on wearable devices relies on single-lead information specifically adapted from 12-lead ECGs extracted from clinical ECG libraries.

    “In the absence of large, labelled datasets of wearable ECGs, the development of algorithms that can detect underlying structural heart disease on wearable devices relies on single-lead information specifically adapted from 12-lead ECGs extracted from clinical ECG libraries.”

  • The paper relies on this without support.

    AI-ECG algorithms have been designed in clinically obtained 12-lead ECGs.

    “Even though AI-ECG is a promising screening tool for detecting LVSD, the algorithms have been designed in clinically obtained 12-lead ECGs.”

  • The paper relies on this without support.

    The intended broader real-world use of this algorithm is as a screening method for LV systolic dysfunction among individuals with no clinical disease.

    “this population differs from the intended broader real-world use of this algorithm as a screening method for LV systolic dysfunction among individuals with no clinical disease”
4 of 4

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Standard-model sensitivity and specificity are reversed in the text

    Minor

    The text reports sensitivity of 0.69 and specificity of 0.91, while Table 2 reports sensitivity of 0.910 and specificity of 0.688 for the standard model without noise.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Moody, G. B., Muldrow, W. & Mark, R. A noise stress test for arrhythmia detectors. Comput. Cardiol. 11, 381–384 (1984).

Cited by

Sorted by citations

Related

Authors
Khunte, A., Sangha, V., Oikonomou, E. K., Dhingra, L. S., Aminorroaya, A., Mortazavi, B. J., Coppi, A., Brandt, C. A., Krumholz, H. M., & Khera, R.
Affiliations
  • Yale University (US)
  • Yale New Haven Hospital (US)
Corresponding author
Khera, Rohan
ORCID
Venue
npj Digital Medicine
Publisher
Springer Science and Business Media LLC
Published
2023-07-11
Volume
6
Issue
1
Pages
124-124
ISSN
2398-6352 (ISSN-L)
DOI
10.1038/s41746-023-00869-w
PMC
PMC10336107
OpenAlex
W4383911793
Cited by
95
References
33
Citations by year
2026: 26 · 2025: 38 · 2024: 24 · 2023: 7
FWCI
18.86
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by)
  • medRxiv (acceptedVersion, cc-by-nd)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Doris Duke Charitable Foundation (2022060)
  • U.S. Department of Health & Human Services | NIH | National Heart, Lung, and Blood Institute (K23HL153775)
Topics
  • Health Sciences › Medicine › Cardiology and Cardiovascular Medicine › ECG Monitoring and Analysis
  • Physical Sciences › Engineering › Biomedical Engineering › Non-Invasive Vital Sign Monitoring
  • Health Sciences › Medicine › Cardiology and Cardiovascular Medicine › Cardiac electrophysiology and arrhythmias
Keywords
left ventricular systolic dysfunction, wearable devices, portable ECG, ECG noise, signal augmentation, artificial intelligence, cardiovascular screening, signal-to-noise ratio

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-28.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-28.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review