Can AI find heart fat on routine ultrasound? Five cohorts linked the score to metabolic risk—not patient outcomes
A peer-reviewed study trained PanAdipo on 1.1 million echocardiography videos and compared its score with CT-measured fat, biomarkers and later diagnoses. The model found a reproducible signal, but retrospective associations cannot show that screening or treatment would help patients.
The short version
- What happened
- PanAdipo was developed from 1,114,441 videos in 28,797 echocardiography studies and evaluated across five distinct cohorts, including 5,167 people with paired ultrasound and coronary CT.
- Who it affects
- Patients may receive a new metabolic-risk flag from an existing heart scan, but it should be explained as an association rather than a diagnosis or treatment instruction.
- What remains unproven
- Retrospective development and cohort analysis; no randomised screening or treatment intervention and no evidence of improved patient outcomes.
Keep in mind
Retrospective development and cohort analysis; no randomised screening or treatment intervention and no evidence of improved patient outcomes.
The training label came from report text and treated unmentioned prominent fat as absent, favouring specificity but creating label noise and possible missed positives.
External cohorts did not share the report-label endpoint, so they tested transportability through CT, biomarker or diagnosis associations rather than the same classification task.
Clinical cohorts were selected for echocardiography, and later diagnoses can reflect surveillance, healthcare contact, reverse causation and residual confounding.
Asian participants were underrepresented, and threshold-specific error burdens across demographic and clinical groups remain unproven.
Several senior authors disclosed patents, company roles, equity, consulting or industry-linked research interests relevant to imaging and health technology.
Multi-source analysis · 2 linked sources. Primary analysed source date: 7 November 2026. Earliest linked supporting source: 9 October 2026.
Read the full report · 12 min The evidence, methods, implications and limitations
The answer: routine ultrasound contained a signal for fat around the heart, but the study did not prove clinical benefit
Researchers led by Yale developed a video model called PanAdipo to estimate prominent epicardial adipose tissue—the fat depot that lies around the heart—from ordinary two-dimensional echocardiography. The strongest result was not simply that the model reproduced a phrase from an ultrasound report. In 5,167 held-out patients who also had coronary CT angiography within 12 months, the score tracked CT-measured epicardial fat and distinguished people above an age- and sex-specific high-fat threshold with reported areas under the receiver operating characteristic curve between 0.91 and 0.95.
That is evidence that the ultrasound videos contain information related to the CT phenotype. It is not evidence that PanAdipo diagnoses a disease, determines who should receive a medicine or improves health. Epicardial fat is associated with cardiovascular-kidney-metabolic risk, but an imaging association can reflect several pathways at once. The paper is a retrospective model-development and cohort-analysis study. It measured discrimination, correlations, biomarkers and newly documented diagnoses; it did not randomise patients to AI-assisted care, compare treatment strategies or count prevented heart attacks, liver complications or deaths.
The practical finding is therefore narrower and still useful: a score extracted from scans already acquired for cardiac care may offer an additional metabolic-risk signal beyond body mass index. Whether acting on that signal helps a person requires a prospective pathway with defined eligibility, clinician review, confirmatory assessment and evidence that benefits outweigh extra tests, anxiety, cost and inappropriate treatment.[1][2]
How the model was developed and where the denominators came from
The development dataset contained 1,114,441 video clips from 28,797 transthoracic echocardiography studies in 21,479 patients across the Yale-New Haven Health System between 2016 and June 2022. PanAdipo reused the PanEcho video foundation model and fine-tuned it with a binary classification head. Its training label came from a deterministic search for an expert reader's explicit report of a prominent epicardial fat depot. Only about 2.3% of development studies carried that label. Studies without the phrase were treated as negative even though silence in a report cannot prove that prominent fat was absent.
The authors then used several cohorts for different questions. A temporally later Yale set included 4,588 patients contributing 5,130 studies and tested whether the model recovered the report label. An emergency-department point-of-care ultrasound cohort included 10,957 people and tested operation on abbreviated, less standard views. The separate MIMIC-IV cohort contributed 4,549 people from another health system. The Multi-Ethnic Study of Atherosclerosis, or MESA, contributed 2,740 research echocardiograms from six US field centres, with detailed metabolic measurements. Finally, 5,167 Yale patients formed the paired-ultrasound-and-CT comparison.
These are not five interchangeable external validations of one endpoint. Only the temporal Yale cohort had the original report label. CT offered an independent quantitative reference for fat volume in the paired cohort. MESA supported cross-sectional biomarker analysis. The health-system cohorts supported analyses of diagnoses recorded later in clinical data. Keeping those roles separate prevents a large total denominator from creating the false impression that every result was reproduced identically in every population.[1][2]
The distinction that matters
How the evidence moves from ultrasound video to a risk association
Each stage answers a different question; none by itself proves that acting on the score benefits patients.
Train
Learn report-labelled prominent fat from 1.1 million videos
PanAdipo score
Compare
Check 5,167 paired coronary CT scans
Agreement with indexed epicardial fat
Associate
Analyse MESA biomarkers and later health-record diagnoses
Metabolic-risk correlations
Test next
Run prospective clinician-reviewed pathways
Evidence about benefit, harm and workload
What the accuracy and CT comparisons actually show
In the temporally held-out Yale test, the model identified reports that mentioned prominent epicardial fat with an AUROC of 0.86 across all views. Combining apical four-chamber and parasternal views raised the reported AUROC to as much as 0.93. AUROC describes how often a randomly chosen positive study is ranked above a randomly chosen negative study across all possible thresholds. It does not select a clinical threshold, reveal the number of false alarms in a screening programme or show whether a particular patient should be treated.
The CT comparison is more informative because it does not depend only on what an echocardiographer happened to write. PanAdipo's score correlated with indexed epicardial fat volume at Spearman ρ=0.76. The corresponding correlation with body mass index was ρ=0.41 in that analysis, and correlations between PanAdipo and BMI ranged from 0.19 to 0.40 across the main cohorts. The authors report that the AI score alone explained 53% of the variation in indexed epicardial fat volume, compared with 23% for age, sex and BMI together.
The high-fat discrimination used an age- and sex-specific threshold at or above the 95th percentile of indexed epicardial fat. That makes the comparison suitable for finding unusually high values within the study population, but it is not a universal diagnostic boundary. A percentile identifies a tail of a distribution by construction. Before a health service could use it, researchers would need to show what absolute risk follows, which action is justified and how thresholds behave across ancestry, age, sex, body composition, scanners and clinical settings.[1][2]
The metabolic associations were specific—and remain associations
In MESA, each standard-deviation increase in the log-transformed PanAdipo score was associated with higher insulin resistance measured by HOMA-IR: the reported standardised coefficient was 0.17 after adjustment for age, sex and body mass index, with a 95% confidence interval from 0.11 to 0.22. Higher scores were also associated with higher triglycerides and lower HDL cholesterol. The pattern persisted in several sensitivity models, including adjustment for waist-to-hip ratio and additional demographic and medication variables. Little or no adjusted association was reported for LDL cholesterol, total cholesterol or systolic blood pressure.
That specificity supports the interpretation that the score is not merely reproducing a generic illness signal. It still does not establish that epicardial fat caused the biomarker differences or that reducing the score would improve them. MESA's analysis is cross-sectional: ultrasound, body composition and biomarkers describe people at a point in time. Unmeasured behaviour, medicines, genetics, socioeconomic conditions and other fat depots may contribute to both the image pattern and metabolic measurements.
The longitudinal health-record analyses began after a 90-day blanking period and excluded people with the condition already documented. After adjustment for body mass index, higher PanAdipo scores remained associated with newly documented hyperlipidaemia, with a pooled hazard ratio of 1.11 per standard deviation, and metabolic dysfunction-associated steatotic liver disease or steatohepatitis, with a pooled hazard ratio of 1.31. Associations with type 2 diabetes and hypertension did not persist after the same adjustment. A new code in a record can reflect surveillance and contact with care as well as true disease onset, and fatty-liver disease is often underdiagnosed.[1][2]
What this could mean for patients and clinical teams
If prospective evidence supports it, the model could turn an existing heart ultrasound into an opportunistic prompt to review metabolic risk. That may be helpful when body mass index understates visceral or ectopic fat. A clinician could explain that the scan contains a risk-associated pattern and then use established history, laboratory tests and shared decision-making to decide whether further assessment is warranted. The score should not be presented as a diagnosis of diabetes, liver disease or future cardiovascular events.
The human consequences depend on workflow. A low-friction flag can become a high-volume alert when applied to millions of scans. False positives may lead to repeat imaging, laboratory tests, referrals, anxiety or treatment without proven benefit. False negatives may reassure people whose risk comes through other pathways. Teams would need a pre-specified threshold, a named professional responsible for review, a route for patients to understand and challenge the result, and monitoring of who receives follow-up.
The study also creates implementation work for sonographers, cardiologists, primary-care clinicians and information teams. The model has to tolerate different vendors, views and acquisition quality, yet staff must know when an image falls outside validated conditions. A risk flag discovered during an emergency scan may not belong to the team managing the acute problem, so hand-off and responsibility matter. The paper did not measure alert burden, consultation time, patient understanding, downstream testing, cost or access inequalities.[1][2]
Transportability, selection and fairness still need direct testing
Using MIMIC-IV and MESA improves the evidence beyond a single-hospital split. MIMIC-IV brings a different health system, patient mix and scanner environment; MESA brings a community cohort recruited without requiring cardiac symptoms. But the external cohorts did not have the same expert report label, so they tested associations rather than identical classification performance. The emergency and clinical echocardiography populations were selected because someone already needed imaging, which can introduce referral, selection and collider biases.
The authors explicitly note that Asian participants were underrepresented even though cardiometabolic risk can occur at lower body mass index in Asian populations. More broadly, age- and sex-specific CT percentiles do not settle performance across race, ethnicity, disability, socioeconomic position or different ultrasound access. A model can appear accurate overall while producing different false-alert burdens across groups or missing people whose images, body composition or clinical pathways differ from the training data.
Prospective evaluation should publish calibration and threshold-specific sensitivity, specificity, positive predictive value and negative predictive value for each intended setting. It should include people not already referred for cardiac imaging, report missing scans and failed inference, and compare the model with simpler measurements such as waist circumference, established risk scores and clinician assessment. External teams should be able to reproduce the analysis; the authors say analytical code is available and model weights and an inference pipeline are intended for release, while the underlying clinical images remain access-controlled.[1][2]
Funding, commercial interests and the evidence that would change the assessment
The work received support from the American Heart Association, several US National Institutes of Health programmes, the British Heart Foundation, an NHS AI award, Innovate UK and European research funding. Several senior authors disclosed patents, company roles, equity, consulting, honoraria or industry-supported institutional research connected to imaging, health technology or cardiometabolic medicine. One author is editor-in-chief of JACC. These disclosures do not invalidate the results, but they make independent replication, transparent code and prospective evaluation especially important.
The assessment would strengthen if independent systems reproduced CT agreement and calibration, including in Asian and other underrepresented populations, then ran a preregistered prospective study. A useful trial would compare ordinary care with clinician-reviewed PanAdipo alerts and measure confirmed metabolic disease, appropriate treatment, unnecessary testing, patient-reported understanding, clinician workload, cost and hard clinical outcomes. It should also test whether the score changes in a meaningful way after weight loss or evidence-based treatment rather than merely tracking noise or imaging conditions.
The assessment would weaken if threshold performance collapses outside Yale, if the score mostly duplicates simpler risk measures, or if acting on it increases tests without improving decisions. For now, the peer-reviewed paper supports a promising imaging biomarker and a clear research programme. It does not support population screening, autonomous diagnosis or a claim that analysing routine echocardiograms improves health. That is why the report belongs in Latest, not Breaking.[1][2]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring medical imaging: compare the finding and its evidence limits.
Can AI spot liposarcoma on ultrasound?
A peer-reviewed Chinese study reports strong results in a 95-patient internal test, including 0.97 accuracy. But all 317 patients came from one hospital and there was no external validation, so this is a proof of concept—not a clinically cleared diagnostic system.
Read the next reportExplore the full source and evidence record
What this means for people
- Patients may receive a new metabolic-risk flag from an existing heart scan, but it should be explained as an association rather than a diagnosis or treatment instruction.
- Clinicians need threshold-specific performance, clear responsibility and a confirmatory pathway so an opportunistic finding does not become an unreviewed high-volume alert.
- Health systems must measure downstream testing, workload, cost and unequal effects before treating the score as a routine screening service.
Global context
The evidence comes from US health-system and community cohorts, with an Oxford-led CT-analysis contribution. Echocardiography access, metabolic-disease prevalence, body-composition patterns, referral pathways and treatment thresholds vary internationally. Any deployment would need local calibration, comparison with affordable non-imaging risk assessment and a pathway that does not widen access gaps between people who do and do not receive cardiac ultrasound.
What the evidence does not yet show
- Retrospective development and cohort analysis; no randomised screening or treatment intervention and no evidence of improved patient outcomes.
- The training label came from report text and treated unmentioned prominent fat as absent, favouring specificity but creating label noise and possible missed positives.
- External cohorts did not share the report-label endpoint, so they tested transportability through CT, biomarker or diagnosis associations rather than the same classification task.
- Clinical cohorts were selected for echocardiography, and later diagnoses can reflect surveillance, healthcare contact, reverse causation and residual confounding.
- Asian participants were underrepresented, and threshold-specific error burdens across demographic and clinical groups remain unproven.
- Several senior authors disclosed patents, company roles, equity, consulting or industry-linked research interests relevant to imaging and health technology.
What to watch next
- Independent external validation with threshold-specific calibration and subgroup error rates
- Prospective studies comparing clinician-reviewed alerts with ordinary care
- Confirmed diagnoses, appropriate treatment, unnecessary testing and patient-reported understanding
- Effects on clinician workload, cost and access across health systems
- Whether the score responds reliably to weight loss or evidence-based metabolic treatment
Living evidence record
Impact record IAI-18LL97E
Evidence stage
Studied
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
8 November 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 8 November 2026
Research question, topics and regions
Whether a video deep-learning score derived from routine heart ultrasound can identify prominent epicardial adiposity and retain associations with metabolic risk beyond body mass index
Topics: Cardiometabolic risk · Echocardiography · Medical imaging · Deep learning · Clinical validation
Regions: United States · North America · Global
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.