Skip to main content
Tessa NewsLink
Paste a health news link, or browse

Source study found

Story checked

Deep learning model using ECGs during sleep studies can predict cardiovascular outcomes (opens in a new tab)

medicalxpress.com · 2026-10-09

Short answerEvidenceSource

Short answer

Mixed

Mixed.

2 claims go further than the study. 2 other points were not covered by the paper.

  • 3 supported
  • 2 overstated
  • 2 not covered

Checked against the study summary. The full text wasn't available, so some details couldn't be settled either way.

Share this check

Follow the evidence trail
1
2

NewsLink checks it

Mixed

Two of seven claims overstate the study. Three of seven check out. Two claims the study doesn't address.

  • 3 supported
  • 2 overstated
  • 2 not covered
Open claim evidence
3
Source paper

Source layer

The 2 papers the story cites

Source study separated from background citations.

The research anchor for the report.

  • The study this story reportsmentioned without context

    Prediction of cardiovascular outcomes using electrocardiography from overnight polysomnography

    SLEEP · 2026

    Publisher (opens in a new tab)
  • The study this story reportsmentioned without context

    10.1093/sleep/zsag229/8885715

    Publisher (opens in a new tab)
Then inspect each claim

Evidence layer

Claim by claim

Each claim gets a verdict. Expand it to see the evidence directly below.

7 claims in this story

Showing all 7 claimsChoose a verdict to focus the list.

Claim 1 of 7OverstatedA research team supported by the NIH found that ECGs recorded during sleep and paired with sleep-stage information can be used to determine the risk of future adverse cardiac events.View evidenceHide evidence

Why this verdict

The abstract supports that a deep-learning model using single-lead ECGs from polysomnography plus sleep-stage data was developed and externally validated for association with 10-year cardiovascular outcomes. However, the story frames the lead claim as a causal-strength finding that ECGs plus sleep-stage information can be used to determine future adverse cardiac-event risk. The paper profile describes retrospective observational prognostic modeling and adjusted associations, not causal evidence or fully established clinical risk determination.

1

Study evidence

NN-output from the fine-tuned deep residual network was strongly associated with incident atrial fibrillation over 10 years in external cohorts, independent of standard risk factors and sleep characteristics.EUH HR 2.03 (1.82–2.27) per 1 SD increase; BIDMC HR 2.72 (2.31–3.21) per 1 SD increase

“We applied a deep residual neural network with attention to single-lead ECGs from PSG, combined with sleep stage data, to predict 10-year risk of atrial fibrillation (AF), stroke, myocardial infarction (MI), heart failure (HF), and death.”
2

Study evidence

Neural-network output strongly associated with 10-year risk of atrial fibrillation after multivariable adjustment.EUH HR 2.03 (95% CI 1.82–2.27); BIDMC HR 2.72 (95% CI 2.31–3.21) per 1-SD increase in NN-output

“External validation used cohorts from Emory University Hospital (EUH) (n = 9810) and Beth Israel Deaconess Medical Center (BIDMC) (n = 12 576), with outcomes derived from electronic health records (ICD-9/10 codes).”
Claim 2 of 7OverstatedThe article says adding the model score improved prediction of cardiac outcomes beyond established risk factors, and that using just one ECG lead could make the approach easier and lower cost than a usual 12-lead setup.View evidenceHide evidence

As statedjust one ECG lead

Why this verdict

The abstract profile supports that the NN-output remained independently associated with outcomes after adjustment for established risk factors and sleep measures. But it also notes that the abstract does not report discrimination, calibration, or decision-curve analyses; independent adjusted association is not the same as demonstrated improvement in prediction beyond established risk factors. The single-lead input is supported, but the easier/lower-cost comparison with a usual 12-lead setup is not established in the supplied profile.

1

Study evidence

NN-output from the fine-tuned deep residual network was strongly associated with incident atrial fibrillation over 10 years in external cohorts, independent of standard risk factors and sleep characteristics.EUH HR 2.03 (1.82–2.27) per 1 SD increase; BIDMC HR 2.72 (2.31–3.21) per 1 SD increase

“We applied a deep residual neural network with attention to single-lead ECGs from PSG, combined with sleep stage data, to predict 10-year risk of atrial fibrillation (AF), stroke, myocardial infarction (MI), heart failure (HF), and death.”
2

Study evidence

Neural-network output strongly associated with 10-year risk of atrial fibrillation after multivariable adjustment.EUH HR 2.03 (95% CI 1.82–2.27); BIDMC HR 2.72 (95% CI 2.31–3.21) per 1-SD increase in NN-output

“External validation used cohorts from Emory University Hospital (EUH) (n = 9810) and Beth Israel Deaconess Medical Center (BIDMC) (n = 12 576), with outcomes derived from electronic health records (ICD-9/10 codes).”
Claim 3 of 7Not coveredThe study used a deep learning approach and was published in the journal SLEEP.View evidenceHide evidence

Why this verdict

The deep-learning approach is supported: the profile describes a deep residual neural network with attention. The supplied abstract-level profile does not directly verify the publication venue as the journal SLEEP, so the full combined claim is not fully verifiable at this depth.

1

Study evidence

NN-output from the fine-tuned deep residual network was strongly associated with incident atrial fibrillation over 10 years in external cohorts, independent of standard risk factors and sleep characteristics.EUH HR 2.03 (1.82–2.27) per 1 SD increase; BIDMC HR 2.72 (2.31–3.21) per 1 SD increase

“We applied a deep residual neural network with attention to single-lead ECGs from PSG, combined with sleep stage data, to predict 10-year risk of atrial fibrillation (AF), stroke, myocardial infarction (MI), heart failure (HF), and death.”
Claim 4 of 7Not coveredThe scientists said the model could predict atrial fibrillation, heart failure, and death from any cause, but additional optimization is needed to predict myocardial infarction and stroke.View evidenceHide evidence

Why this verdict

The profile supports stronger associations for atrial fibrillation, heart failure, and all-cause mortality, and it shows myocardial infarction was inconsistent across external cohorts. However, the quoted statement that additional optimization is needed for myocardial infarction and stroke is not present in the supplied abstract-level profile. Also, the profile reports statistically significant stroke associations in both external cohorts, so the specific optimization caveat for stroke cannot be verified from this profile alone.

1

Study evidence

Neural-network output strongly associated with 10-year risk of atrial fibrillation after multivariable adjustment.EUH HR 2.03 (95% CI 1.82–2.27); BIDMC HR 2.72 (95% CI 2.31–3.21) per 1-SD increase in NN-output

“External validation used cohorts from Emory University Hospital (EUH) (n = 9810) and Beth Israel Deaconess Medical Center (BIDMC) (n = 12 576), with outcomes derived from electronic health records (ICD-9/10 codes).”
Claim 5 of 7SupportedThe research team aimed to predict the 10-year risk of cardiovascular outcomes including atrial fibrillation, stroke, myocardial infarction, heart failure, and all-cause mortality from sleep-study ECG recordings.View evidenceHide evidence

As stated10-year risk

Why this verdict

The profile states that the model used single-lead PSG ECGs plus sleep-stage data to predict 10-year risk of atrial fibrillation, stroke, myocardial infarction, heart failure, and death.

1

Study evidence

NN-output from the fine-tuned deep residual network was strongly associated with incident atrial fibrillation over 10 years in external cohorts, independent of standard risk factors and sleep characteristics.EUH HR 2.03 (1.82–2.27) per 1 SD increase; BIDMC HR 2.72 (2.31–3.21) per 1 SD increase

“We applied a deep residual neural network with attention to single-lead ECGs from PSG, combined with sleep stage data, to predict 10-year risk of atrial fibrillation (AF), stroke, myocardial infarction (MI), heart failure (HF), and death.”
Claim 6 of 7SupportedThe model was fine-tuned on 15,809 patients at Massachusetts General Hospital and assessed on 9,810 patients from Emory University Hospital and 12,576 patients from Beth Israel Deaconess Medical Center.View evidenceHide evidence

As stated15,809; 9,810; 12,576

Why this verdict

The cohort sizes and institutions match the profile: fine-tuning on 15,809 Massachusetts General Hospital patients, with external validation in 9,810 Emory University Hospital patients and 12,576 Beth Israel Deaconess Medical Center patients.

1

Study evidence

NN-output from the fine-tuned deep residual network was strongly associated with incident atrial fibrillation over 10 years in external cohorts, independent of standard risk factors and sleep characteristics.EUH HR 2.03 (1.82–2.27) per 1 SD increase; BIDMC HR 2.72 (2.31–3.21) per 1 SD increase

“We applied a deep residual neural network with attention to single-lead ECGs from PSG, combined with sleep stage data, to predict 10-year risk of atrial fibrillation (AF), stroke, myocardial infarction (MI), heart failure (HF), and death.”
2

Study evidence

Neural-network output strongly associated with 10-year risk of atrial fibrillation after multivariable adjustment.EUH HR 2.03 (95% CI 1.82–2.27); BIDMC HR 2.72 (95% CI 2.31–3.21) per 1-SD increase in NN-output

“External validation used cohorts from Emory University Hospital (EUH) (n = 9810) and Beth Israel Deaconess Medical Center (BIDMC) (n = 12 576), with outcomes derived from electronic health records (ICD-9/10 codes).”
Claim 7 of 7SupportedThe model could sort people into groups with different levels of long-term cardiovascular risk and retained predictive value after adjustment for age, sex, BMI, diabetes, hypertension, sleep apnea severity, arousals, and sleep efficiency.View evidenceHide evidence

Why this verdict

The profile supports risk stratification in the sense that NN-output was associated with long-term cardiovascular risk and remained significant after multivariable adjustment for demographics, clinical risk factors, and sleep characteristics. The story’s covariate list is somewhat abbreviated relative to the profile, which also mentions smoking, periodic limb movement index, and time in sleep stages, but the core adjusted-association claim is supported.

1

Study evidence

Neural-network output strongly associated with 10-year risk of atrial fibrillation after multivariable adjustment.EUH HR 2.03 (95% CI 1.82–2.27); BIDMC HR 2.72 (95% CI 2.31–3.21) per 1-SD increase in NN-output

“External validation used cohorts from Emory University Hospital (EUH) (n = 9810) and Beth Israel Deaconess Medical Center (BIDMC) (n = 12 576), with outcomes derived from electronic health records (ICD-9/10 codes).”
Then look for missing context

Context layer

What the story left out

Important study details the story did not include.

  • Outcome ascertainment was based on electronic health record ICD-9/10 codes.

    The story does not mention ICD-code-based outcome labeling, an important caveat for interpreting prognostic outcome validity.

    From Retrospective model development and fine-tuning; retrospective multi-cohort external validation using time-to-event mode

  • The evidence is observational and retrospective, so it supports prognostic association rather than causal inference or proven clinical utility.

    The story includes predictive language but does not clearly identify the retrospective observational nature of the evidence. This matters because the lead claim is framed at causal strength.

    From Retrospective model development and fine-tuning; retrospective multi-cohort external validation using time-to-event mode

6 things the story did carry across
  • The paper developed a deep residual neural network with attention using single-lead nocturnal PSG ECG combined with sleep-stage data to predict 10-year cardiovascular outcomes.
  • The target outcomes were atrial fibrillation, stroke, myocardial infarction, heart failure, and all-cause mortality over a 10-year horizon.
  • The model was fine-tuned on an MGH cohort and externally validated in EUH and BIDMC cohorts with the reported sample sizes.
  • External validation relied on retrospective hospital cohorts and Cox proportional hazards models estimating adjusted associations between NN-output and time-to-event outcomes.
  • NN-output remained associated with outcomes after adjustment for conventional risk factors and sleep/PSG-derived covariates.
  • Myocardial infarction prediction was less consistent, with the BIDMC confidence interval including 1.0; stroke associations were smaller but reported as significant in both external cohorts.
Then read the study layer

Study layer

Study at a glance

Scan the study first. Expand only the parts you want to inspect.

Pieces of work

2

Evidence read

study summary

Lead result

secondary data

1Lead resultsecondary dataDevelop and validate a deep-learning model that uses single-lead nocturnal ECG from polysomnography (plus sleep stage data) to predict long-term (10-year) cardiovascular outcomes (AF, stroke, MI, HF, death).Retrospective model development and fine-tuningExpand

In plain English

The authors developed and externally validated a deep residual neural network with attention that inputs single-lead electrocardiograms recorded during overnight polysomnography together with sleep-stage information to predict 10-year risk of atrial fibrillation (AF), stroke, myocardial infarction (MI), heart failure (HF), and all-cause mortality. The network was pretrained/trained for arrhythmia detection and then fine-tuned on a retrospective Massachusetts General Hospital (MGH) cohort (n = 15,809). External validation cohorts were from Emory University Hospital (EUH, n = 9,810) and Beth Israel Deaconess Medical Center (BIDMC, n = 12,576). Outcomes were ascertained from electronic health records using ICD-9/10 codes. In Cox proportional hazards models, the model output (NN-output) was strongly associated with long-term risk for all outcomes (p < .0001) and remained significant after adjustment for demographic and clinical risk factors and sleep characteristics.

Key findings

  • NN-output from the fine-tuned deep residual network was strongly associated with incident atrial fibrillation over 10 years in external cohorts, independent of standard risk factors and sleep characteristics.EUH HR 2.03 (1.82–2.27) per 1 SD increase; BIDMC HR 2.72 (2.31–3.21) per 1 SD increase
  • NN-output was associated with incident ischemic stroke over 10 years after adjustment for covariates.EUH HR 1.19 (1.09–1.31) per 1 SD increase; BIDMC HR 1.40 (1.12–1.74) per 1 SD increase
“We applied a deep residual neural network with attention to single-lead ECGs from PSG, combined with sleep stage data, to predict 10-year risk of atrial fibrillation (AF), stroke, myocardial infarction (MI), heart failure (HF), and death.”
What this piece can’t prove
  • Retrospective development and fine-tuning on MGH registry data; potential for selection biases inherent to retrospective EHR cohorts.
  • Outcome labels were derived from ICD-9/10 diagnosis codes in EHRs; abstract does not report validation of coding-based outcome ascertainment.
  • Abstract does not report discrimination metrics (e.g., AUROC, C-index), calibration, or decision-curve analyses for the predictive model.
  • Abstract does not provide detailed model training hyperparameters, sample splits (train/validation/test) for the MGH cohort, or handling of missing data.

1 further detail could not be confirmed from the summary.

2secondary dataDemonstrate that the model’s output is independently associated with long-term cardiovascular risk after adjustment for conventional risk factors and sleep/PSG-derived covariates using time-to-event modeling, and quantify hazard ratios across external cohorts.retrospective multi-cohort external validation using time-to-event modelingExpand

In plain English

External validation in two independent hospital cohorts showed that a deep-learning neural-network output derived from single-lead nocturnal ECGs (with sleep-stage input) was independently associated with 10-year risk of multiple cardiovascular outcomes when evaluated with multivariable Cox proportional hazards models; hazard ratios per 1 standard-deviation increase in NN-output are reported for AF, stroke, MI, HF, and all-cause mortality in Emory University Hospital (EUH, n=9,810) and Beth Israel Deaconess Medical Center (BIDMC, n=12,576).

Key findings

  • Neural-network output strongly associated with 10-year risk of atrial fibrillation after multivariable adjustment.EUH HR 2.03 (95% CI 1.82–2.27); BIDMC HR 2.72 (95% CI 2.31–3.21) per 1-SD increase in NN-output
  • Neural-network output associated with increased 10-year risk of stroke after multivariable adjustment.EUH HR 1.19 (95% CI 1.09–1.31); BIDMC HR 1.40 (95% CI 1.12–1.74) per 1-SD increase in NN-output
“External validation used cohorts from Emory University Hospital (EUH) (n = 9810) and Beth Israel Deaconess Medical Center (BIDMC) (n = 12 576), with outcomes derived from electronic health records (ICD-9/10 codes).”
Finally, the search trail

Method layer

NewsLink found the paper. Tessa takes you deeper.

NewsLink checks the story. Tessa is where you inspect the paper, authors, evidence, and research context.

Open the paper in Tessa

Prediction of cardiovascular outcomes using electrocardiography from overnight polysomnography

SLEEP · 2026

Source record (opens in a new tab)

Why this one

Near certain

NewsLink found the paper. Tessa is where you inspect it deeply.

Papers considered

The selected paper, plus nearby candidates.

Crossref, PubMed, Europe PMC · 16 candidate papers

Selected

Prediction of cardiovascular outcomes using electrocardiography from overnight polysomnography

SLEEP · 2026 · Crossref

Source (opens in a new tab)

And 10 more candidates considered.