CoChat
Sign Up
Cited by 50Health Promotion and Cardiovascular PreventionCardiac Imaging and DiagnosticsLipoproteins and Cardiovascular Health

Deep Learning to Estimate Cardiovascular Risk From Chest Radiographs

Open access
Weiss, Jakob, Raghu, Vineet K., Paruchuri, Kaavya, Zinzuwadia, Aniket, Natarajan, Pradeep, Aerts, Hugo J.W.L., Lu, Michael T.·2024·Annals of Internal Medicine
DOI 10.7326/m23-1898PMC PMC11613935
Open this paper in CoChat — free account

Abstract

BACKGROUND: Guidelines for primary prevention of atherosclerotic cardiovascular disease (ASCVD) recommend a risk calculator (ASCVD risk score) to estimate 10-year risk for major adverse cardiovascular events (MACE). Because the necessary inputs are often missing, complementary approaches for opportunistic risk assessment are desirable. OBJECTIVE: To develop and test a deep-learning model (CXR CVD-Risk) that estimates 10-year risk for MACE from a routine chest radiograph (CXR) and compare its performance with that of the traditional ASCVD risk score for implications for statin eligibility. DESIGN: Risk prediction study. SETTING: Outpatients potentially eligible for primary cardiovascular prevention. PARTICIPANTS: The CXR CVD-Risk model was developed using data from a cancer screening trial. It was externally validated in 8869 outpatients with unknown ASCVD risk because of missing inputs to calculate the ASCVD risk score and in 2132 outpatients with known risk whose ASCVD risk score could be calculated. MEASUREMENTS: 10-year MACE predicted by CXR CVD-Risk versus the ASCVD risk score. RESULTS: Among 8869 outpatients with unknown ASCVD risk, those with a risk of 7.5% or higher as predicted by CXR CVD-Risk had higher 10-year risk for MACE after adjustment for risk factors (adjusted hazard ratio [HR], 1.73 [95% CI, 1.47 to 2.03]). In the additional 2132 outpatients with known ASCVD risk, CXR CVD-Risk predicted MACE beyond the traditional ASCVD risk score (adjusted HR, 1.88 [CI, 1.24 to 2.85]). LIMITATION: Retrospective study design using electronic medical records. CONCLUSION: On the basis of a single CXR, CXR CVD-Risk predicts 10-year MACE beyond the clinical standard and may help identify individuals at high risk whose ASCVD risk score cannot be calculated because of missing data. PRIMARY FUNDING SOURCE: None.

AI snapshot

Written from the paper’s full text.

TL;DR

A deep-learning model (CXR CVD-Risk) estimates 10-year risk of major adverse cardiovascular events (MACE) from a routine chest radiograph. In external validation, it predicted MACE in outpatients with unknown and known ASCVD risk and added prognostic value beyond the traditional ASCVD risk score.

Key findings

  • Among 8,869 outpatients with unknown ASCVD risk because of missing inputs, CXR CVD-Risk ≥7.5% classified 47.9% (4,245/8,869) as statin eligible; 11.8% (500/4,245) of those developed incident MACE versus 6.2% (285/4,624) of those classified as ineligible, with adjusted HR 1.73 (95% CI 1.47–2.03).
  • Among 2,132 outpatients with known ASCVD risk, CXR CVD-Risk predicted 10-year MACE beyond the traditional ASCVD risk score: adjusted HR 1.88 (95% CI 1.24–2.85) for statin-eligible versus ineligible patients independent of the ASCVD risk score.
  • Only 19.4% (2,132/11,001) of patients potentially eligible for primary prevention had all inputs available to calculate the ASCVD risk score; the remaining 81% (8,869/11,001) had unknown ASCVD risk.
  • CXR CVD-Risk and the ASCVD risk score were concordant for statin eligibility in 69.7% (1,485/2,132); among discordant cases, CXR CVD-Risk upclassified 50.4% (326/647) as statin eligible, with a MACE rate of 4.6% (15/326).

Methods

  • CXR CVD-Risk was developed in the Prostate, Lung, Colorectal, Ovarian (PLCO) Cancer Screening Trial using a random 80% subset (n=40,718) and chest radiographs from all four timepoints (n=147,801); the only input was a CXR image, and the output was probability of cardiovascular mortality.
  • External validation without retraining used outpatients aged 50–75 years from Mass General Brigham with routine posterior-anterior CXRs obtained in 2010–2011, not on statin, without prevalent diabetes or LDL-C ≤70–>190 mg/dl if available, and without prior MACE, coronary artery disease, or peripheral artery disease; two cohorts were assessed: 8,869 with unknown ASCVD risk and 2,132 with known ASCVD risk.
  • Analyses included Cox proportional hazards regression, cumulative incidence curves, Harrell’s c-statistic, calibration graphs and Integrated Calibration Index, decision curves, and reclassification tables; MACE was defined using ICD-9/10 codes for acute stroke and myocardial infarction, with manual event confirmation in the known-ASCVD-risk cohort.

Limitations

  • Stated: retrospective study design across two hospitals using risk factors and outcomes extracted from the electronic medical record; the cohort likely had higher risk than the general population; ICD-based phenotyping may have low specificity, with manual validation only in the 2,132 patients with ASCVD-risk inputs available, possibly overestimating event rates in the broader cohort.
  • Stated: the study included individuals aged 50–75 years although primary prevention guidelines apply to ages 40–75; the cohort primarily consisted of non-Hispanic whites, so generalizability to racially and ethnically diverse cohorts and other care settings is uncertain; convolutional neural networks are “black boxes,” making individual CXR CVD-Risk estimates difficult to explain.
  • Evident from the text: analyses used complete case analysis with no data imputation, and 232 (11%) individuals were censored because available follow-up was below 10 years.

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

Integrity signals

Registry

Crossref · 10.7326/m23-1898

Cites

24 of 24 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

3 of 6 cited sources checked · 3 supported · 0 flagged

  • Premisecitation

    Deep learning can capture imaging features beyond traditional findings (e.g. enlarged heart silhouette, lung nodules) with valuable prognostic and potentially actionable information.

    “the results of this and previously published studies(10–12) suggest that deep learning can capture imaging features beyond traditional findings (e.g. enlarged heart silhouette, lung nodules) with valuable prognostic and potentially actionable information.”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [10] Lu MT, Ivanov A, Mayrhofer T, Hosny A, Aerts HJWL, Hoffmann U. Deep Learning to Assess Long-term Mortality From Chest Radiographs. JAMA Netw Open. 2019 Jul 3; 2(7): e197416. 31322692 PMC6646994 source
    • cites [11] Lu MT, Raghu VK, Mayrhofer T, Aerts HJWL, Hoffmann U Deep Learning Using Chest Radiographs to Identify High-Risk Smokers for Lung Cancer Screening Computed Tomography: Development and Validation of a Prediction Model. Ann Intern Med. 2020 Nov 3; 173(9): 704– 13. 32866413 PMC9200444 source
    • cites [12] Raghu Vineet K, Weiss Jakob, Hoffmann Udo, Aerts Hugo JWL, Lu Michael T Deep Learning to Estimate Biological Age From Chest Radiographs. JACC Cardiovasc Imaging [Internet]. 0(0). Available from: 10.1016/j.jcmg.2021.01.008 PMC11381093 33744131 source
  • Findingown data

    In the cohort of 8,869 outpatients with unknown ASCVD risk, CXR CVD-Risk had a higher net benefit than baseline strategies of treating everyone and treating no one across risk thresholds close to the 7.5% standard for statin eligibility.

    “In the cohort of 8,869 outpatients with unknown ASCVD risk, CXR CVD-Risk had a higher net benefit than baseline strategies of treating everyone and treating no one (Figure 2a) across risk thresholds close to the 7.5% standard for statin eligibility.”
    Quote in the paperSubstantial claim
  • Findingown data

    Of those identified as statin eligible by CXR CVD-Risk, 7.6% (60/787) developed incident MACE vs. 3.1% (42/1,345) of those identified as ineligible (HR 2.77 [95% CI 1.87-4.11]).

    “Of those identified as statin eligible by CXR CVD-Risk, 7.6% (60/787) developed incident MACE vs. 3.1% (42/1,345) of those identified as ineligible (HR 2.77 [95% CI 1.87-4.11]).”
    Quote in the paperSubstantial claim
  • Findingown data

    Cumulative incidence curves showed a graded association between CXR CVD-Risk categories and 10-year incident MACE.

    “Cumulative incidence curves showed a graded association between CXR CVD-Risk categories and 10-year incident MACE (Figure 3).”
    Quote in the paperSubstantial claim
  • Findingown data

    The CXR CVD-risk model revealed a clear risk stratification among the 8,869 patients with unknown ASCVD risk.

    “In addition, the CXR CVD-risk model also revealed a clear risk stratification among the 8,869 patients with unknown ASCVD risk but potentially eligible for primary cardiovascular prevention (Table 1).”
    Quote in the paperSubstantial claim
  • May combine two claims. 15
  • Classified as probably not a substantial claim. 9

  • May combine two claims.

    Findingown data

    Of those classified as statin eligible by CXR CVD-Risk, 11.8% (500/4,245) developed incident MACE compared to 6.2% (285/4,624) of those classified as ineligible.

    “Of those classified as statin eligible by CXR CVD-Risk, 11.8% (500/4,245) developed incident MACE compared to 6.2% (285/4,624) of those classified as ineligible.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    CXR CVD-Risk also stratified among Low, Borderline and Intermediate ASCVD risk score groups (risk of 10-year MACE for Low: 1.3%-4.2%; Borderline: 2.5%-6.3%; Intermediate: 4.3%-13.3%) whereas this effect was not seen for the High ASCVD risk groups.

    “Moreover, CXR CVD-Risk also stratified among Low, Borderline and Intermediate ASCVD risk score groups (risk of 10-year MACE for Low: 1.3%-4.2%; Borderline: 2.5%-6.3%; Intermediate: 4.3%-13.3%) whereas this effect was not seen for the High ASCVD risk groups (Table 1).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The association between CXR CVD-Risk as well as the traditional ASCVD risk score and time to incident MACE was estimated using cumulative incidence curves by the four ordinal risk categories, which show a strong separation for both the CXR CVD-Risk and the traditional ASCVD risk score.

    “The association between CXR CVD-Risk as well as the traditional ASCVD risk score and time to incident MACE was estimated using cumulative incidence curves by the four ordinal risk categories (Figure 3), which show a strong separation for both the CXR CVD-Risk and the traditional ASCVD risk score.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The mean estimates for 10-year MACE risk were 8.6% (SD 6.9) for CXR CVD-Risk and 7.2% (SD 6.0) for the traditional ASCVD risk score.

    “The mean estimates for 10-year MACE risk were 8.6% (SD 6.9) for CXR CVD-Risk and 7.2% (SD 6.0) for the traditional ASCVD risk score.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    The study followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) reporting guideline.

    “This study followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) reporting guideline.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductreasoning

    All presented results are based on actual observed 10-year incident MACE and are reported for the MGB testing dataset only, which was not used during any part of the model development process.

    “All presented results are based on actual observed 10-year incident MACE and are reported for the MGB testing dataset only, which was not used during any part of the model development process.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Chest radiographs are the most commonly performed imaging test and are acquired during the course of many patients' clinical workup at low cost and radiation dose.

    “Chest radiographs are the most commonly performed imaging test(9) and are acquired during the course of many patients’ clinical workup at low cost(23) and radiation dose(24).”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [9] Ron E Cancer risks from medical radiation. Health Phys. 2003 Jul; 85(1): 47– 59. 12852471 source
    • cites [23] Website [Internet]. [cited 2021 Mar 18]. Available from: American College of Radiology. Medicare physician fee sched- ule impacts for 70,000-series CPT codes. Accessed at www.acr.org/-/media/ACR/Files/Advocacy/MPFS/2019-Impact-Table-70000

      Not checked in the free run.

    • cites [24] Website [Internet]. Available from: Radiological Society of North America, American College of Radiology. Radiation dose to adults from common imaging examinations. Accessed at www.acr.org/-/media/ACR/Files/Radiology-Safety/Radiation-Safety/Dose-Reference-Card

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

  • May combine two claims.

    Interpretationown data

    Opportunistic screening of existing chest radiographs could complement the traditional ASCVD risk score to identify patients at high cardiovascular risk who may benefit from primary prevention, and would be of particular impact in the majority where inputs to calculate the 10-year ASCVD risk score were not available.

    “Opportunistic screening of existing chest radiographs could complement the traditional ASCVD risk score to identify patients at high cardiovascular risk who may benefit from primary prevention, and would be of particular impact in the majority (81%, 8,869/11,001) of our cohort where the inputs to calculate the 10-year ASCVD risk score were not available in the EMR.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    In patients with known ASCVD risk, CXR CVD-Risk had similar performance and additive value to the traditional ASCVD risk score.

    “in these patients with known ASCVD risk, CXR CVD-Risk had similar performance and additive value to the traditional ASCVD risk score”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    The funding organizations had no role in the design or conduct of the study.

    “These organizations had no role in the design or conduct of the study.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Opportunistic screening of chest radiographs may help identify individuals at high risk of cardiovascular disease, prompting risk factor assessment and targeted prevention.

    “Opportunistic screening of chest radiographs may help identify individuals at high risk of cardiovascular disease, prompting risk factor assessment and targeted prevention.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    In the group of patients with known ASCVD risk, MACE events were manually confirmed using discharge summaries, cardiology reports and radiology reports.

    “In the group of patients with known ASCVD risk, MACE events were manually confirmed using discharge summaries, cardiology reports and radiology reports.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    Convolutional neural networks are inherently "black boxes," and CXR CVD-Risk estimates are difficult to explain on an individual basis as they quantify the underlying 10-year risk rather than a specific disease or imaging feature that directly leads to MACE.

    “Finally, convolutional neural networks are inherently “black boxes,” and CXR CVD-Risk estimates are difficult to explain on an individual basis as they quantify the underlying 10-year risk rather than a specific disease or imaging feature that directly leads to MACE.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The cohort included patients who had a routine chest radiograph, lipid panel, and blood pressure measurements within a 2-year period, and thus are likely at higher risk compared to the general population.

    “Our cohort included patients who had a routine chest radiograph, lipid panel, and blood pressure measurements within a 2-year period, and thus are likely at higher risk compared to the general population.”
    Quote in the paperNot a substantial claim

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

    Interpretationown data

    The study cohort primarily consisted of non-Hispanic whites, so further studies are necessary to assess generalizability in racially and ethnically diverse cohorts and in other care settings.

    “Moreover, as our study cohort primarily consisted of non-Hispanic whites, further studies are necessary to assess generalizability in racially and ethnically diverse cohorts and in other care settings.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The proposed model is a new risk score/factor that, instead of traditional risk factors, takes a CXR as input to estimate the underlying risk of an individual.

    “we consider the proposed model as a new risk score/factor that, instead of traditional risk factors (e.g. diabetes or smoking), takes a CXR as input to estimate the underlying risk of an individual”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The proportional hazards assumption was verified using scaled Schoenfeld residuals.

    “The proportional hazards assumption was verified using scaled Schoenfeld residuals.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    ICD-based phenotyping may have low specificity for outcome determination; the paper manually validated events in the 2,132 patients with inputs available to calculate ASCVD risk but may be overestimating the event rate in the broader cohort.

    “ICD-based phenotyping may have low specificity for outcome determination; we manually validated events in the 2,132 patients with inputs available to calculate ASCVD risk but may be overestimating the event rate in the broader cohort.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    232 (11%) individuals were censored as available follow-up was below 10 years.

    “232 (11%) individuals were censored as available follow-up was below 10 years.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductreasoning

    Guidelines for primary prevention apply to individuals 40-75 years of age; however, the study only included individuals 50-75 years of age, as the CXR CVD-Risk model was developed in people 55-74 years of age.

    “Guidelines for primary prevention apply to individuals 40-75 years of age; however, we only included individuals 50-75 years of age, as the CXR CVD-Risk model was developed in people 55-74 years of age.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Based on a single CXR, CXR CVD-Risk predicts 10-year MACE beyond the clinical standard and may help identify high-risk individuals with missing data to calculate the ASCVD risk score.

    “Based on a single CXR, CXR CVD-Risk predicts 10-year MACE beyond the clinical standard and may help identify high-risk individuals with missing data to calculate the ASCVD risk score.”
    Quote in the paperSubstantial claim
21 of 21

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.

  • Discussion understates the adjusted risk increase

    Minor

    The discussion describes a 1.5-fold risk for patients classified as statin eligible, independent of available risk factors. The corresponding adjusted hazard ratio reported in the results is 1.73, not 1.5.

  • LDL-C is described as a required calculator input but is absent from the input list

    Minor

    The paper gives LDL-C as an example of an input necessary to calculate the ASCVD risk score, but its own list of calculator inputs does not include LDL-C.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Nalichowski R, 2006, AMIA Annu Symp Proc, 2006, 1044

Cited by

Sorted by citations

Related

Authors
Weiss, J., Raghu, V. K., Paruchuri, K., Zinzuwadia, A., Natarajan, P., Aerts, H. J. W. L., & Lu, M. T.
Affiliations
  • Harvard University (US)
  • Broad Institute (US)
  • Brigham and Women's Hospital (US)
Corresponding author
  • Weiss, Jakob
  • Raghu, Vineet K.
ORCID
Venue
Annals of Internal Medicine
Publisher
American College of Physicians
Published
2024-04
Volume
177
Issue
4
Pages
409-417
ISSN
  • 0003-4819 (ISSN-L)
  • 1539-3704
DOI
10.7326/m23-1898
PMC
PMC11613935
OpenAlex
W4393139437
Cited by
50
References
23
Citations by year
2026: 20 · 2025: 22 · 2024: 8
FWCI
18.60
Type
journal article
Language
en
Open access
Licence
Locations
  • Annals of Internal Medicine (publishedVersion)
  • PubMed (publishedVersion)
  • Research Publications (Maastricht University) (publishedVersion, other-oa)
  • PubMed Central (submittedVersion)
  • Europe PMC (PubMed Central) (submittedVersion)
Funders
  • American Heart Association (935176)
  • Massachusetts General Hospital
  • Dana-Farber Cancer Institute
  • Broad Institute
  • Deutsche Forschungsgemeinschaft (525002713)
  • Fondation Leducq (R01HL148050, TNE‐18CVD04, R01HL148565, T32HL007208, 18CVD04)
  • Albert-Ludwigs-Universität Freiburg
  • National Institutes of Health (5r01hl148050-04, 5r01hl148565-07)
  • Medizinische Fakultät der Albert-Ludwigs-Universität Freiburg
  • National Heart, Lung, and Blood Institute (T32 HL 007208, R01‐HL148565, TNE-18CVD04, R01 HL151283, 5-T32HL007208-43, K01 HL168231, R01‐HL148050, T32HL007208-43, R01HL142711)
  • National Cancer Institute
Topics
  • Health Sciences › Medicine › Public Health, Environmental and Occupational Health › Health Promotion and Cardiovascular Prevention
  • Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › Cardiac Imaging and Diagnostics
  • Health Sciences › Medicine › Surgery › Lipoproteins and Cardiovascular Health
Keywords
cardiovascular risk prediction, chest radiograph, major adverse cardiovascular events, atherosclerotic cardiovascular disease, ASCVD risk score, deep learning, statin eligibility, primary prevention, 10-year risk
MeSH terms
Deep Learning, Heart Disease Risk Factors, Cardiovascular Diseases, Humans, Retrospective Studies, Risk Factors, Risk Assessment, Atherosclerosis

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 2026-10-27.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-27.

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