Skip to main content
NCBI home page

This is a preprint.

It has not yet been peer reviewed by a journal.

The National Library of Medicine is running a pilot to include preprints that result from research funded by NIH in PMC and PubMed.

medRxiv logoLink to medRxiv
[Preprint]. 2026 Jul 24:2026.07.22.26358690. [Version 1] doi: 10.64898/2026.07.22.26358690

AI-Enabled Echocardiography Identifies an Adverse Epicardial Adiposity Phenotype Associated with Cardiometabolic Dysfunction

Arya Aminorroaya a,b, Andreas Coppi a, Robert L McNamara a, Joao AC Lima c, Harlan M Krumholz a,d,e, Charalambos Antoniades f, Rohan Khera a,b,e,g,h, Evangelos K Oikonomou a,b,*
PMCID: PMC13419655  PMID: 42539104

Abstract

Background:

Excess epicardial adipose tissue (EAT) is associated with cardiovascular-kidney-metabolic (CKM) dysfunction, but its assessment has traditionally required advanced imaging. We tested whether AI-enhanced echocardiography could enable scalable phenotyping of adverse epicardial adiposity and identify individuals at increased cardiometabolic risk.

Methods:

We developed PanAdipo, a video-based deep learning model in 1,114,441 videos from 28,797 studies across the Yale-New Haven Health System (YNHHS; 2016–2022), using expert reader annotations of prominent EAT (2.5% of studies). PanAdipo was evaluated in four cohorts: a temporally distinct YNHHS TTE set (n=4,588), an emergency-department point-of-care ultrasound cohort across YNHHS (n=10,957), the geographically distinct MIMIC-IV cohort (n=4,549), and the community-based Multi-Ethnic Study of Atherosclerosis (MESA Exam 6, n=2,740). Analyses examined (i) discrimination of prominent EAT; (ii) independence from conventional echocardiographic outputs; (iii) spatial explainability using gradient-weighted class activation mapping; (iv) correspondence with paired CT-derived body composition phenotypes (n=5,594); and (v) age-, sex-, and BMI-independent associations with cardiometabolic biomarkers and incident metabolic disease.

Results:

In the held-out health system test set, PanAdipo discriminated prominent EAT with an AUROC of 0.91 (95% CI, 0.88–0.94), exceeding conventional measures of cardiac function and structure. In explainability analyses, the model’s attention localized to the epicardial area across views and throughout the cardiac cycle. On paired cardiac CT imaging, the PanAdipo score correlated most strongly with epicardial adiposity (Spearman ρ=0.75; P<10−300) with weaker correlations with other adipose and non-adipose compartments and only modest correlations with BMI across cohorts (ρ=0.19–0.40). In MESA, greater PanAdipo scores were independently associated with higher HOMA-IR and triglycerides, associations that persisted among normoglycemic participants. Higher PanAdipo scores were also associated with newly documented metabolic disease, including MASLD/MASH, after adjustment for BMI (HRpooled 1.25 [1.08–1.44] per 1-SD increment in log[PanAdipo]).

Conclusions:

AI-enabled echocardiography provides a scalable, view-agnostic biomarker that characterizes adverse epicardial adiposity and is associated with cardiometabolic dysfunction, highlighting a new role for echocardiography in CKM risk stratification.

Keywords: artificial intelligence, echocardiography, epicardial adipose tissue, cardiovascular-kidney-metabolic syndrome

INTRODUCTION

Cardiovascular-kidney-metabolic (CKM) syndrome, as recently defined by the American Heart Association (AHA), frames the pathophysiologic continuum that links excess and dysfunctional adiposity to type 2 diabetes (T2D), chronic kidney disease, atherosclerotic cardiovascular disease, and heart failure.1 Within this construct, dysfunctional adipose tissue is the proximate driver of cardiometabolic risk,2 and CKM staging explicitly identifies adiposity as a target for early intervention.1–3 However, body mass index (BMI), the dominant clinical descriptor of adiposity, only partially reflects the type, location, and biological behavior of the fat that determines risk.2,4 Visceral and ectopic depots, in particular epicardial adipose tissue (EAT), are more closely linked to insulin resistance and dyslipidemia than total body weight, even across the normal BMI range.5–8

EAT is uniquely positioned over the myocardium and coronary arteries, shares its microcirculation with the heart, and exchanges paracrine and vasocrine signals with the adjacent myocardium.5,9,10 Beyond its volume, the composition of EAT is also informative; radiomic phenotyping of EAT on cardiac computed tomography (CT) encodes early myocardial remodeling and predicts future heart failure before clinical onset.11,12 EAT is also a clinically actionable target, responsive to weight loss, glucagon-like peptide-1 receptor agonists (GLP-1 RAs), and sodium-glucose cotransporter-2 (SGLT2) inhibitors, each of which carries CKM benefits beyond glycemic control.13,14 Despite this biological and therapeutic relevance, scalable phenotyping of adverse EAT remains confined to dedicated cardiac CT or magnetic resonance imaging,7,13 which restricts its use for population-level prevention and trial design.

Echocardiography is one of the most widely performed cardiac imaging modalities and directly captures the epicardial fat compartment.15,16 Prior work has established the feasibility of manually measuring EAT thickness from the parasternal long-axis (PLAX) view;15,17 yet manual quantification remains operator-dependent, view-restricted, and difficult to deploy at scale. Recent advances in foundation models for echocardiography enable view-agnostic, multi-task interpretation from both protocolized transthoracic echocardiography (TTE) and point-of-care ultrasound (POCUS) acquisitions.18,19 However, existing tools have focused on ventricular function, chamber geometry, and valvular phenotypes, and do not capture adverse adiposity.

Here, we present PanAdipo, a video-based deep learning model that derives a phenotype of adverse epicardial adiposity from standard and portable echocardiographic acquisitions. Across temporally and geographically distinct health system and community-based cohorts, we evaluated whether this AI-derived phenotype captures information beyond existing echocardiographic measurements, corresponds to distinct adiposity phenotypes on paired cross-sectional imaging, and is associated with both cardiometabolic biomarkers and subsequently documented metabolic disease. This work seeks to open a new avenue for echocardiography, extending its role beyond the characterization of cardiac structure and function to metabolic risk stratification using images already acquired as part of routine clinical care.

METHODS

Study overview and ethics

This was a multi-cohort, retrospective study of AI-enabled echocardiographic phenotyping of adverse epicardial adiposity. The study was approved by the Yale Institutional Review Board with a waiver of informed consent (Protocol #2000040942). Analyses of de-identified data from Medical Information Mart for Intensive Care, version IV (MIMIC-IV), accessed through PhysioNet-credentialed access,20,21 and the Multi-Ethnic Study of Atherosclerosis (MESA), accessed through NHLBI BioLINCC and BioData Catalyst, were performed under approved data use agreements.22 PanAdipo model weights and inference code will be released upon publication; the underlying clinical data are not publicly redistributable but can be made available from the respective providers under appropriate access procedures.

Development cohort and label definition

We developed PanAdipo using 1,114,441 videos from 28,797 random TTEs in 21,479 unique patients performed across Yale-New Haven Health System (YNHHS) hospitals and outpatient clinics between January 1, 2016, and June 30, 2022, building on the curated multi-view dataset assembled for PanEcho, as described below.18 The training label was an expert reader-assigned label for a prominent epicardial fat depot. Board-certified echocardiographers recorded this label during routine clinical reporting in the structured findings and conclusions fields of the final report. We extracted the label at scale using a deterministic, expert-curated parser that required at least one explicit mention of “prominent epicardial adipose tissue noted” or “prominent epicardial fat [depot]”. This label was intentionally designed to favor specificity over sensitivity, capturing overt clinical recognition of marked epicardial adiposity rather than a quantitative threshold susceptible to view-, angle-, and operator-dependent variation in linear measurements. Patient-level training, validation, and held-out testing splits mirrored the PanEcho development pipeline to prevent data leakage.18

Model architecture and fine-tuning

We built PanAdipo on the publicly released PanEcho video foundation model.18 As previously described, the PanEcho backbone integrates standard 2D B-mode echocardiographic views (PLAX, parasternal short-axis [PSAX], apical 2-, 3-, 4-, and 5-chamber (A5C), subcostal, and Doppler) within a shared spatial encoder and a temporal transformer, producing a 768-dimensional pooled embedding from a 16-frame clip at 224×224 pixel resolution. To this embedding, we attached a lightweight binary classification head consisting of a dropout layer (rate 0.3) followed by a single linear projection, and we fine-tuned the full network end-to-end.

We trained the network in distributed data-parallel mode across four NVIDIA H100 GPUs, with a per-GPU mini-batch of 48 video clips, yielding an effective global batch size of 192. We set the initial learning rate to 10−4 for both the head and the backbone and used the AdamW optimizer and a patience of five epochs. Recognizing that PanEcho was originally trained to detect cardiac and not adiposity phenotypes, the backbone was unfrozen. To address the expected low prevalence of the target label (2.5% of studies), we applied a positive-class weight of 15 directly within the binary cross-entropy loss. We trained for up to 30 epochs with early stopping on the validation area under the precision-recall curve (AUPRC), terminating training when validation AUPRC failed to improve for five consecutive epochs, and retained the best-AUPRC checkpoint for downstream evaluation. At inference time, we generated a single per-study PanAdipo score by averaging the predicted probabilities across all eligible 2D B-mode clips for a given study, providing a view-pooled score that does not require a specific acquisition view.18 The resulting output probability (range: 0–1) was operationalized as a continuous biomarker (score) of the prominent epicardial fat depot phenotype, with higher values denoting greater model confidence in the presence of prominent epicardial adiposity by echocardiography.

Internal and external testing cohorts

We evaluated the PanAdipo model in four independent cohorts whose composition was chosen to stress test the model along four orthogonal axes of generalizability: temporal, modality, geographic, and freedom from confounding by indication (Figure 1). For internal testing, we used a temporally distinct YNHHS TTE set acquired in the second half of 2022 (n=4,588 individuals contributing 5,130 unique studies), thus probing temporal generalization within the same institution and acquisition protocol. To evaluate the cross-cohort deployment and external transportability of the biomarker, we deployed PanAdipo to three additional cohorts. The first was cardiac-focused POCUS performed in YNHHS emergency departments (n=10,957 studies from unique individuals), which captures abbreviated, off-axis acquisitions by frontline clinicians and tests modality generalization. The second was a temporally and geographically distinct sample of adult TTE studies from MIMIC-IV (n=4,549 studies from unique individuals),21 a publicly available critical care and hospital-based dataset from Beth Israel Deaconess Medical Center (Boston, MA, USA). MIMIC-IV represents an entirely separate health system, vendor mix, and patient population, and tests both geographic and institutional generalization.20,21 The third was research-grade adult TTE from Exam 6 of MESA (n=2,740 studies from unique individuals), a longitudinal, NHLBI-sponsored, multi-center, prospective community-based cohort enrolled across six United States field centers without regard to cardiovascular symptoms; Exam 6 was conducted between 2016 and 2018 in middle-aged to older adults with detailed cardiometabolic phenotyping.22 MESA is therefore relatively free of the confounding-by-indication that affects clinically referred cohorts. Of note, because expert reader-defined EAT labels were available only within the YNHHS TTE cohort, the held-out (temporal) YNHHS TTE set was used to assess discrimination of prominent EAT, whereas the external cohorts, which lacked such labels, were used to evaluate the clinical and prognostic implications of the PanAdipo score.

Figure 1 |. Study overview.

Figure 1 |

Open in a new tab

(A) PanAdipo is a video-based deep learning model developed to characterize adverse epicardial adiposity (continuous score 0 to 1) from any standard 2D B-mode echocardiographic acquisition. The model was developed at the YNHHS and tested across four cohorts: a temporally distinct YNHHS TTE set, cardiac-focused emergency-department POCUS, MIMIC-IV, and MESA. (B) Downstream evaluations included benchmarking against paired structural and functional echocardiographic outputs, cross-modal radiomic correlations against automated body and organ analysis on CCTA, cross-sectional associations with cardiometabolic biomarkers, and prediction of incident metabolic disease. Abbreviations: AxC, apical x-chamber; CCTA, coronary computed tomography angiography; ED, emergency department; MESA: Multi-Ethnic Study of Atherosclerosis; MIMIC: Medical Information Mart for Intensive Care; PLAX, parasternal long-axis; POCUS: point-of-care ultrasound; PSAX, parasternal short-axis; YNHHS, Yale-New Haven Health System.

Benchmarking against existing PanEcho outputs

To test whether the prominent epicardial fat depot phenotype could be inferred from existing deep learning-derived echocardiographic measurements, we benchmarked PanAdipo against the panel of PanEcho structural and functional outputs in the held-out YNHHS TTE testing set. For each PanEcho output, we computed its AUROC for the binary expert reader label of prominent EAT and compared it with the AUROC of PanAdipo for the same label. We further quantified, for each PanEcho output, the proportion of its variance explained by PanAdipo and, separately, by deep learning-derived left ventricular ejection fraction (LVEF) as a reference for a globally reported function metric. Together, these analyses test whether a dedicated adiposity-focused video model captures information about cardiac structure and function that is not already encoded by existing deep learning-derived phenotypes.

Explainability and cross-modal radiomic analyses

Saliency mapping:

Gradient-weighted class activation maps (Grad-CAM++) were generated from the final convolutional block of the PanEcho ConvNeXt image encoder for high-confidence true-positive predictions.23 Per-frame maps were computed independently across each 16-frame clip, sharpened with an unsharp mask, and thresholded at a normalized activation of 0.25 for visualization.

Cross-sectional radiomic body and organ analysis:

Furthermore, to examine cross-modal correspondence between PanAdipo and body- and organ-level adiposity phenotypes derived from coronary CT angiography (CCTA), we identified, within YNHHS, all individuals who had undergone a CCTA (between 2013 and 2025) within one year of a TTE, and who are included in the multicenter, multisite Oxford Risk Factors and Non-invasive Imaging (ORFAN) study.11 We then identified wide field-of-view (FoV) series that extended sufficiently to allow cross-sectional assessment of subcutaneous, visceral, intermuscular, pericardial, skeletal muscle, and bone adiposity in addition to the cardiac chambers. We processed each CCTA volume with the open-source Body and Organ Analysis (BOA) pipeline,24,25 performing fully automated multi-organ segmentation and extracting 833 size, density, texture, and shape features across anatomical compartments (Table S1, Data S1).

Cross-sectional cardiometabolic associations in MESA

Whereas cardiometabolic biomarker measurements in the hospital-based cohorts are indication-driven and heterogeneous in timing relative to the index echocardiogram, MESA provided access to protocolized phenotypes regardless of cardiovascular symptoms. Leveraging this protocolized phenotyping, we evaluated the association between key cardiometabolic biomarkers (triglycerides, high-density lipoprotein [HDL] and low-density lipoprotein [LDL] cholesterol, total cholesterol, fasting glucose, HbA1c, log[homeostatic model assessment for insulin resistance (HOMA-IR)], systolic and diastolic blood pressure, and estimated glomerular filtration rate using the CKD-EPI equation) and the standardized log(PanAdipo) using regression models with sequential adjustment. We repeated all models in a pre-specified normoglycemic subgroup defined as HbA1c <5.7% and fasting glucose <100 mg/dL.

Longitudinal cardiometabolic outcomes

In YNHHS (TTE and POCUS) and MIMIC-IV, where longitudinal follow-up was available, we evaluated the prognostic value of PanAdipo for four pre-specified incident cardiometabolic endpoints (T2D, hypertension, hyperlipidemia, and metabolic dysfunction-associated steatotic liver disease and steatohepatitis [MASLD/MASH]), each defined from linked structured electronic health record (EHR) data as the first occurrence after the index echocardiogram. Endpoints, and the prevalent disease excluded at baseline, were identified from structured diagnosis records using prespecified International Classification of Diseases (ICD) codes, e.g. Ninth and Tenth Revision, Clinical Modification (e.g., ICD-9, ICD-10-CM) (Table S2). We further evaluated the association of PanAdipo with all-cause mortality in the same three cohorts and in MESA. To mitigate reverse causation due to prevalent disease, we restricted each analysis to individuals free of the respective condition at baseline and adopted a landmark design with a 90-day blanking window, as well as sensitivity analyses for 30 and 365 days.

Statistical analysis

We summarized continuous variables as mean (standard deviation [SD]) or median [25th to 75th percentile] depending on distribution, and categorical variables as counts (percentages). We computed the AUROC of PanAdipo and of each PanEcho output for the binary label of prominent epicardial fat depot, with 95% confidence interval (CI) derived from 500-iteration bootstrap resampling at the study level. To quantify the statistical independence of PanAdipo from existing echocardiographic measurements, we further estimated, for each PanEcho output in turn, the coefficient of determination (R2, defined as the squared Pearson correlation) of that output with PanAdipo and, in parallel, with PanEcho-derived LVEF. For the paired CT analysis, we tested cross-modal associations between the per-study PanAdipo score and each of 833 BOA-derived radiomic features using Spearman rank correlations, and report results at a nominal threshold (P <0.05), after Benjamini-Hochberg false discovery rate correction, and at a Bonferroni-adjusted threshold (P <6.0×10−5) reflecting the 833 features tested.

In MESA, we modeled the cross-sectional association between log-transformed and standardized PanAdipo and each cardiometabolic biomarker (also z-scored) with ordinary least-squares linear regression, sequentially adjusting for age and sex, then additionally for BMI, and separately for waist-hip ratio (WHR), so that the coefficient on PanAdipo represents the standardized beta per SD of log(PanAdipo). In a sensitivity analysis, we repeated all models in the normoglycemic subgroup.

For longitudinal analyses, we excluded individuals with prevalent disease at the index echocardiogram and began follow-up after a blanking window of 30, 90, or 365 days. We fit Cox proportional hazards models in each cohort separately with Huber-White robust standard errors, using the standardized log(PanAdipo) as the exposure, with adjustment for age and sex in the primary model and additional adjustment for BMI in the secondary model. For the four incident cardiometabolic endpoints (T2D, hypertension, hyperlipidemia, MASLD/MASH), we pooled cohort-specific log hazard ratios using a fixed effect inverse-variance meta-analytic estimator. Tests were two-sided with a nominal significance level of 0.05 unless otherwise specified. All analyses were performed in Python 3.11. Reporting follows the TRIPOD+AI statement, with a completed checklist provided in Table S3.26

RESULTS

Cohort composition

Across the four testing cohorts, we analyzed 22,834 individuals in total (Table 1, Figure S1): 4,588 in the YNHHS TTE held-out test set, 10,957 in YNHHS POCUS, 4,549 in MIMIC-IV, and 2,740 in MESA. The mean age ranged from 66±17 years (YNHHS TTE) to 74±9 years (MESA) (Figure S1A), whereas the mean BMI ranged from 28.5 to 29.7 kg/m2 (Figure S1B); and PanAdipo scores followed right-skewed distributions across cohorts (median 0.05 to 0.09; IQR 0.03 to 0.18), with the lowest values in MESA (Figure S1C, log scale). There was a balanced sex distribution, with the proportion of female participants ranging from 46.6% to 52.8% across cohorts (Figure S1D). As expected for their respective settings, the hospital-based cohorts had a substantial baseline burden of cardiometabolic comorbidities, whereas the community-based MESA cohort had a markedly lower burden of overt cardiovascular disease but a high prevalence of hypertension and hyperlipidemia (Figure S1E).

Table 1 |.

Baseline characteristics across the four PanAdipo testing cohorts.

1. YNHHS TTE 2. YNHHS POCUS 3. MIMIC-IV TTE 4. MESA TTE
Total Counts (n) 4,588 10,957 4,549 2,740
Age, mean ± SD (years) 65.5 ± 16.7 66.1 ± 17.2 70.3 ± 13.2 73.7 ± 8.5
Female, n (%) 2,299 (50.1%) 5,107 (46.6%) 2,337 (51.4%) 1,446 (52.8%)
Race
 Non-Hispanic White 3,324 (72.4%) 6,986 (63.8%) 2,891 (63.6%) 1,116 (40.7%)
 Non-Hispanic Black 547 (11.9%) 2,322 (21.2%) 739 (16.2%) 678 (24.7%)
 Hispanic 448 (9.8%) 1,221 (11.1%) 172 (3.8%) 586 (21.4%)
 Asian 108 (2.4%) 194 (1.8%) 113 (2.5%) 360 (13.1%)
 Other/Unknown 161 (3.5%) 234 (2.1%) 634 (13.9%) -
BMI, mean ± SD (kg/m2) 29.1 ± 7.1 29.7 ± 7.7 29.4 ± 7.2 28.5 ± 5.7
PanAdipo, median [IQR] 0.08 [0.03–0.18] 0.07 [0.03–0.17] 0.09 [0.05–0.16] 0.05 [0.03–0.07]
Hypertension, n (%) 2,643 (57.6%) 7,020 (64.1%) 2,949 (64.8%) 1,808 (66.0%)
Hyperlipidemia, n (%) 2,238 (48.8%) 4,812 (43.9%) 2,422 (53.2%) 1,714 (62.6%)
Type 2 diabetes, n (%) 939 (20.5%) 3,196 (29.2%) 1,272 (28.0%) 651 (23.8%)
MASLD/MASH, n (%) 232 (5.1%) 513 (4.7%) 138 (3.0%) -
Myocardial infarction, n (%) 406 (8.8%) 2,574 (23.5%) 795 (17.5%) 83 (3.0%)
Heart failure, n (%) 1,077 (23.5%) 4,323 (39.5%) 1,257 (27.6%) 64 (2.3%)
Chronic coronary syndrome, n (%) 1,110 (24.2%) 3,328 (30.4%) 1,377 (30.3%) 195 (7.1%)
Stroke, n (%) 459 (10.0%) 1,288 (11.8%) 371 (8.2%) 77 (2.8%)
Atrial fibrillation, n (%) 675 (14.7%) 1,918 (17.5%) 805 (17.7%) 187 (6.8%)
Peripheral vascular disease, n (%) 142 (3.1%) 582 (5.3%) 282 (6.2%) 32 (1.2%)
Open in a new tab

Abbreviations: BMI: body mass index; IQR: interquartile range; MASLD: metabolic dysfunction-associated steatotic liver disease; MASH: metabolic dysfunction-associated steatohepatitis; MESA: Multi-Ethnic Study of Atherosclerosis; POCUS: point-of-care ultrasound; SD: standard deviation; TTE: transthoracic echocardiography; YNHHS: Yale-New Haven Health System.

PanAdipo detects prominent epicardial adiposity independent of structural measurements

In the temporally held-out Yale TTE test set, study-level PanAdipo discriminated the expert reader’s prominent EAT with an AUROC of 0.91 (95% CI 0.88–0.94). At a 90% specificity threshold (score of ≥0.48), the specificity and sensitivity were 91.3% and 67.7%, respectively.

To assess whether this phenotype could already be inferred from existing PanEcho-derived measurements, we benchmarked PanAdipo against each of the original PanEcho outputs for the same binary label. The AUROC of PanAdipo (0.91) substantially exceeded that of every individual PanEcho structural heart disease output (next best AUROC, 0.79) (Figure S2). The best-performing PanEcho outputs were those reflecting septal thickness, left ventricular wall thickness, and components of diastolic and right heart function. To further evaluate this independence, we quantified how much of the variance in each individual PanEcho output was explained by PanAdipo, using PanEcho-derived left ventricular ejection fraction (LVEF) as a reference (Figure S3). LVEF, as a global function metric, explained a substantial proportion of the variance of functional and chamber-geometry outputs (R2 of 0.80 for LV systolic function, 0.74 for LV internal diameter at end-systole, 0.65 for end-systolic volume). In contrast, PanAdipo explained at most 4% of the variance of any individual PanEcho measurement. Taken together, these findings suggest that adverse epicardial adiposity, although partially correlated with structural and diastolic remodeling, is not encoded by traditional echocardiographic models focusing on structural heart disease.

PanAdipo encodes specific radiomic phenotypes of visceral adiposity

We next examined what PanAdipo had learned to recognize. Grad-CAM++ saliency maps concentrated over the anterior epicardium in PLAX, over the anterior right ventricular wall in PSAX, and over the basal and apical epicardial regions in A4C views (Figure 2). As an independent test of the inferred phenotype, we further identified individuals across YNHHS who had a CCTA within one year of the index echocardiogram (n=5,594 studies from 5,481 unique participants; n=2,911 [52.0%] female, age 61 [IQR: 50–69] years) and computed the Spearman correlation between the per-study PanAdipo score and 833 radiomic features extracted from automated body and organ segmentation of the CT (Figure 3). PanAdipo was most strongly associated with fat fraction within the pericardium (ρ = +0.75, P <1×10−300), i.e., the percentage of the volume surrounded by the pericardium that is occupied by epicardial fat. Among adipose depot volumes, the association was strongest for absolute EAT volume (ρ = +0.68, P <1×10−300), exceeding those observed for visceral adipose tissue volume (ρ = +0.40, P = 1.4 × 10−208) and subcutaneous adipose tissue volume (ρ = +0.45, P = 2.3 × 10−276), suggesting that PanAdipo learns a phenotype of epicardial-specific adiposity.

Figure 2 |. PanAdipo class activation maps attend to the epicardial area.

Figure 2 |

Open in a new tab

Grad-CAM++ explainability maps from the PanAdipo video encoder, overlaid on representative frames across PLAX, PSAX, and A4C views. Attention largely concentrates over the epicardial area in each view. Abbreviations: A4C, apical four-chamber; Grad-CAM++, gradient-weighted class activation mapping; PLAX, parasternal long-axis; PSAX, parasternal short-axis.

Figure 3. PanAdipo correlates with CT-derived radiomic phenotypes of ectopic adiposity.

Figure 3.

Open in a new tab

(A) Schematic of the cross-modal analysis in a TTE cohort at YNHHS with CCTA performed within one year of the index echocardiogram (n=5,594 studies). The wide field-of-view of cardiac-focused CCTA allowed cross-sectional inclusion of subcutaneous, visceral, intermuscular, pericardial, muscle, and bone adiposity in addition to the cardiac chambers. PanAdipo scores from the echocardiogram were correlated with 833 size, density, texture, and shape features extracted from automated body and organ analysis modules (BOA pipeline). (B) Spearman ρ plot organized by anatomical compartment, showing the direction and magnitude of cross-modal associations between PanAdipo and BOA-derived CCTA features. The strongest positive associations were observed for fat fraction within the pericardium (ρ ≈ +0.75), EAT volume (ρ ≈ +0.68), whereas associations with visceral and subcutaneous adipose tissue volumes were more modest (ρ ≈ +0.40–0.45). Abbreviations: BOA: body and organ analysis; CCTA, coronary computed tomography angiography; CT, computed tomography; EAT: epicardial adipose tissue; TTE: transthoracic echocardiography; YNHHS: Yale-New Haven Health System.

Cross-sectional cardiometabolic associations in MESA are independent of BMI

Across cohorts, PanAdipo was only modestly correlated with BMI (Spearman ρ=0.40 in YNHHS TTE test, 0.34 in MIMIC-IV, 0.23 in YNHHS POCUS, and 0.19 in MESA; Figure S4). To understand the BMI-independent implications of PanAdipo phenotyping, we evaluated cross-sectional associations with cardiometabolic biomarkers in the deeply phenotyped MESA cohort. Per SD of log(PanAdipo), there were significant associations with higher triglycerides, lower HDL cholesterol, and higher log(HOMA-IR), and modestly higher systolic and diastolic blood pressure, with little or no association with LDL or total cholesterol (Figure 4A). These associations persisted after additional adjustment for BMI and WHR, including in a pre-specified normoglycemic subgroup (Figure 4B), indicating a depot-specific signal not explained by anthropometric body size or by overt dysglycemia.

Figure 4 |. Cross-sectional cardiometabolic associations in MESA, independent of BMI.

Figure 4 |

Open in a new tab

Standardized beta coefficients (with 95% CIs) of log(PanAdipo) for cardiometabolic biomarkers measured at MESA Exam 6, in (A) the full cohort and (B) the pre-specified normoglycemic subgroup (HbA1c <5.7% and fasting glucose <100 mg/dL). Three nested models are shown: adjustment for age and sex; additional adjustment for BMI; and additional adjustment for WHR. Color scale reflects effect direction and magnitude (red = positive association; blue = inverse). Abbreviations: BMI, body mass index; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HDL, high-density lipoprotein cholesterol; HOMA-IR, homeostatic model assessment for insulin resistance; LDL, low-density lipoprotein cholesterol; SBP, systolic blood pressure; WHR, waist-hip ratio.

PanAdipo predicts incident cardiometabolic disease in opportunistic cohorts

To evaluate the prognostic implications of opportunistic PanAdipo deployment, we performed landmark analyses across YNHHS (TTE and POCUS) and MIMIC-IV. We excluded individuals with prevalent disease (T2D, hypertension, hyperlipidemia, MASLD/MASH) at the index echocardiogram and applied blanking windows of 90 days (with additional sensitivity analyses with 30- and 365-day blanking windows). Each SD of log(PanAdipo) was associated with higher age- and sex-adjusted risk of newly-documented T2D, hypertension, hyperlipidemia, and MASLD/MASH (90-day pooled HRs 1.21 [1.11–1.31], 1.07 [1.01–1.14], 1.17 [1.11–1.23], and 1.37 [1.20–1.55], respectively; Figure 5, Table S4). After additional adjustment for BMI in the subset with available anthropometrics, the associations with hyperlipidemia (HR 1.11, 95% CI 1.04–1.18) and MASLD/MASH (HR 1.25, 95% CI 1.08–1.44) persisted, indicating a BMI-independent contribution. By contrast, mortality analyses showed an inverse association in the emergency department POCUS cohort, consistent with an obesity paradox pattern,27 but no association in the YNHHS TTE, MIMIC-IV, or community-based MESA cohorts (Table S5).

Figure 5 |. PanAdipo predicts incident metabolic disease in cohorts free of baseline disease.

Figure 5 |

Open in a new tab

(A) Landmark design: individuals with baseline (prevalent) disease were excluded; follow-up for new-onset disease began after a 90-day blanking window following the index echocardiogram. (B) Forest plot of cohort-specific and pooled (fixed effect inverse-variance meta-analysis) hazard ratios per SD increase in log(PanAdipo) for incident type 2 diabetes, hypertension, hyperlipidemia, and MASLD/MASH at 90 days, adjusted for age and sex, in the YNHHS (TTE and POCUS) and MIMIC-IV cohorts. Abbreviations: MASLD: metabolic dysfunction-associated steatotic liver disease; MASH: metabolic dysfunction-associated steatohepatitis; MIMIC: Medical Information Mart for Intensive Care; POCUS: point-of-care ultrasound; TTE: transthoracic echocardiogram.

DISCUSSION

In this study, we developed PanAdipo, a video-based deep learning model that derives a phenotype of adverse epicardial adiposity from routine echocardiographic acquisitions, and evaluated its transportability and clinical relevance across diverse cohorts. Through complementary analyses incorporating benchmarking against existing echocardiographic outputs, gradient-based saliency mapping, cross-modal radiomic correlations with paired cardiac CT, detailed cardiometabolic phenotyping in a community-based cohort, and landmark longitudinal analyses in health system-based populations, we show that PanAdipo (i) identifies expert reader-defined prominent epicardial adiposity in a temporally held-out test cohort; (ii) captures information not well-represented by traditional anthropometric or echocardiographic measures of cardiac structure and function; (iii) shows its strongest cross-modal correspondence with CT-defined epicardial adiposity; and (iv) is associated with insulin resistance, dyslipidemia, and subsequently documented MASLD/MASH and hyperlipidemia beyond conventional anthropometric measures.

Our findings carry several implications for the early recognition and management of CKM syndrome.1 First, echocardiography is the most widely performed cardiac imaging modality, and every standard transthoracic acquisition incidentally captures the epicardial fat compartment.16 PanAdipo enables scalable, opportunistic identification of adverse visceral adiposity from imaging already obtained for unrelated indications, including abbreviated point-of-care studies performed by frontline clinicians, thus supporting both retrospective screening of existing echocardiographic archives and prospective phenotyping at the bedside via POCUS.28 Second, PanAdipo dissociates adverse adiposity from anthropometric body size. It correlates only modestly with BMI; its cross-sectional associations with HOMA-IR, triglycerides, and HDL persist after adjustment for anthropometrics; and its predictive value for incident MASLD/MASH and hyperlipidemia is independent of BMI. Third, by identifying individuals with adverse epicardial adiposity who are missed by anthropometric risk assessment, our biomarker may provide a biologically grounded substrate for the allocation of adipose-modifying therapies, including GLP-1 RA,29 dual incretin agonists,30 SGLT2 inhibitors, and other CKM-targeted agents that act on the visceral and ectopic adipose milieu, and for trial enrichment in cardiometabolic prevention studies. Together, these properties define a complementary axis of CKM risk stratification, deliverable through imaging that is already in place.

Our observations sit naturally within the broader literature on epicardial and visceral adiposity. The prevalence of adverse epicardial fat accumulation in unselected clinical populations is substantial,5 and EAT volume and composition have repeatedly been linked to insulin resistance, dyslipidemia, and incident metabolic disease independent of overall body size.8,31 The signature we observed, namely robust associations with HOMA-IR, triglycerides, HDL cholesterol, and HbA1c, but minimal associations with LDL cholesterol, total cholesterol, or blood pressure, maps specifically to the insulin resistance and triglyceride-rich-lipoprotein axis of CKM, rather than to global, non-specific cardiovascular risk factors. Most associations persist after adjustment for BMI and WHR and within a normoglycemic subgroup, further supporting a depot-specific signal not captured by anthropometric or glycemic measures alone. PanAdipo thus serves as a proof of concept that the metabolically active adipose compartments can be inferred from routine 2D echocardiography.

Certain limitations merit consideration. First, the training label was an expert-reader assignment of a prominent epicardial fat depot extracted from routine clinical reports, which is more specific than sensitive. Clinical reports flag overt accumulation rather than applying a standardized echocardiographic threshold for adverse adiposity. The continuous biomarker score nevertheless tracks a graded spectrum of severity, as supported by its cross-modal radiomic correspondence and dose-response associations with cardiometabolic biomarkers and incident disease. Second, prognostic inference for hard cardiovascular outcomes is intrinsically difficult in clinically referred hospital cohorts. Patients undergoing TTE or POCUS are selected on cardiac symptoms or known disease, conditioning on the index echocardiogram introduces selection and collider bias, and competing risks and reverse confounding by frailty and cachexia produce paradoxical short-term associations between adiposity and event rates.27 Our landmark analyses for incident T2D, hypertension, hyperlipidemia, and MASLD/MASH mitigate some of these biases, but prospective evaluation in lower-risk populations and over longer horizons will be needed to further clarify the prognostic value of this AI biomarker. Third, PanAdipo is not a replacement for cardiac CT-based radiomic profiling of EAT. With millimetric resolution and dedicated radiomic pipelines, CT captures quantitative and qualitative aspects of the epicardial fat compartment, including textural and density signatures that are not accessible to echocardiography.7,11 Finally, whether AI-derived phenotyping of adverse epicardial adiposity changes physician behavior, downstream investigation, and patient outcomes will require prospective evaluation.

In conclusion, AI-enabled echocardiography provides a scalable biomarker of adverse epicardial adiposity that captures information beyond conventional echocardiographic and anthropometric measures, is associated with cardiometabolic dysfunction and subsequently documented CKM disease, and highlights a new role for routinely acquired echocardiograms in opportunistic metabolic risk stratification.

Supplementary Material

Tables S1-S5

Figures S1-S4

Data S1

TRIPOD+AI Checklist

CLINICAL PERSPECTIVE.

What Is New?

  • AI-enabled echocardiography derives a scalable, view-agnostic biomarker of adverse epicardial adiposity from standard transthoracic and point-of-care ultrasound acquisitions.

  • The biomarker captures information not represented by conventional echocardiographic measurements and corresponds most strongly to CT-defined epicardial adiposity on paired cross-sectional imaging.

  • Higher biomarker values are associated with insulin resistance, dyslipidemia, and subsequently documented metabolic disease beyond body mass index and waist-hip ratio.

What Are the Clinical Implications?

  • Routine echocardiography may provide an opportunistic window into adverse adiposity and extend cardiac imaging to metabolic risk stratification.

  • This approach provides a portable platform for prospective evaluation of point-of-care adiposity phenotyping, integration with complementary cardiometabolic data, and enrichment of prevention trials with participants at increased metabolic risk.

ACKNOWLEDGMENTS

The authors thank the participants and investigators of the Multi-Ethnic Study of Atherosclerosis (MESA), the MIMIC-IV team at Beth Israel Deaconess Medical Center, and the clinicians and staff of the Yale-New Haven Health System whose imaging studies made this work possible. The authors further wish to acknowledge the contributions of the consortium working on the development of the NHLBI BioData Catalyst® ecosystem, which enabled these analyses.

SOURCES OF FUNDING

E.K.O. acknowledges research support from the American Heart Association (AHA; award no. 26CDA1612298), the Robert A. Winn Excellence in Clinical Trials Career Development Award, the Wiesman Award for Excellence in Early-Career ATTR Research, a Pepper Scholar Award through the Claude D. Pepper Older Americans Independence Center at Yale School of Medicine (P30AG021342), and a Yale Center for Clinical Investigation KL2 award through a CTSA Grant Number UL1 TR001863 from NCATS, a component of the NIH. R.K. acknowledges support from the National Heart, Lung, and Blood Institute (R01HL167858 and K23HL153775), and the National Institute on Aging (R01AG089981). C.A. acknowledges support through the British Heart Foundation (CH/F/21/90009, TG/19/2/34831, RG/F/21/110040 RE/24/130024), NHS-AI awards (AI_AWARD02013, AI_AWARD02013), Innovate UK (#104472, #104688), EU Research and Innovation Action MAESTRIA (965286), and NIHR Oxford Biomedical Research Centre (Cardiac and Imaging themes). The contents are solely the responsibility of the authors and do not necessarily represent the official views of the funders or the NIH.

DISCLOSURES

E.K.O. is a named co-inventor on patent applications (filed through Yale University) and granted patents licensed through the University of Oxford to Caristo Diagnostics Ltd, outside the scope of this work. He is a co-founder of Evidence2Health LLC, and has previously consulted for Caristo Diagnostics Ltd and Ensight-AI Inc. He has also received honoraria from Clinical Education Alliance, and serves as an Associate Editor for the European Heart Journal. R.K. is an Associate Editor of JAMA and receives research support, through Yale, from the Blavatnik Foundation, Bristol-Myers Squibb, Novo Nordisk, and BridgeBio. He is a co-inventor of pending patent applications (unrelated to this study), and a co-founder of Ensight-AI Inc and Evidence2Health LLC. C.A. declares several patents (US10695023B2, PCT/GB2017/053262, GB2018/1818049.7, GR20180100490, and GR20180100510) licensed to Caristo Diagnostics, and honoraria from Amarin, Silence Therapeutics, Abcentra, Amgen, Nodthera, UCB, and Caristo Diagnostics. C.A. is the immediate past Chair of the British Atherosclerosis Society, and a founder, shareholder and non-executive director of Caristo Diagnostics. H.M.K. has received stock options from Element Science, OpenEvidence, Arboretum, and Identifeye and payments from Arboretum and F-Prime for advisory roles; was a cofounder of and held equity in Hugo Health; is a cofounder of and holds equity in Refactor Health, Carabel, and ENSIGHT-AI; is a cofounder of medRxiv; is on the Board of openRxiv (nonpaid, volunteer); and is associated with research contracts through Yale University from Janssen, Kenvue, Novartis, and Pfizer. The remaining authors have no pertinent disclosures.

NON-STANDARD ABBREVIATIONS AND ACRONYMS

BOA

Body and Organ Analysis

EAT

epicardial adipose tissue

FoV

field of view

Grad-CAM++

gradient-weighted class activation mapping

MESA

Multi-Ethnic Study of Atherosclerosis

MIMIC-IV

Medical Information Mart for Intensive Care, version IV

WHR

waist-hip ratio

YNHHS

Yale-New Haven Health System

DATA AND MODEL SHARING

Model weights and an inference pipeline for PanAdipo are intended for release at the time of publication on our Lab’s GitHub repository. Underlying clinical imaging and electronic health record data from the Yale-New Haven Health System are not publicly redistributable. MIMIC-IV is available through PhysioNet credentialed access. This manuscript was prepared using MESA Research Materials obtained through the NHLBI Biologic Specimen and Data Repository Information Coordinating Center (BioLINCC) and/or the NHLBI BioData Catalyst® ecosystem under approved data use agreements. The content is solely the responsibility of the authors and does not necessarily represent the official views of the MESA investigators, BioLINCC, BioData Catalyst, the NHLBI, or the National Institutes of Health (NIH).

REFERENCES

  • 1.Ndumele CE, Neeland IJ, Tuttle KR, Chow SL, Mathew RO, Khan SS, Coresh J, Baker-Smith CM, Carnethon MR, Després J-P, Ho JE, Joseph JJ, Kernan WN, Khera A, Kosiborod MN, Lekavich CL, Lewis EF, Lo KB, Ozkan B, Palaniappan LP, Patel SS, Pencina MJ, Powell-Wiley TM, Sperling LS, Virani SS, Wright JT, Rajgopal Singh R, Elkind MSV, Rangaswami J, American Heart Association. A synopsis of the evidence for the science and clinical management of cardiovascular-kidney-metabolic (CKM) syndrome: A scientific statement from the American heart association. Circulation. 2023;148:1636–1664. [DOI] [PubMed] [Google Scholar]
  • 2.Cornier M-A, Després J-P, Davis N, Grossniklaus DA, Klein S, Lamarche B, Lopez-Jimenez F, Rao G, St-Onge M-P, Towfighi A, Poirier P, American Heart Association Obesity Committee of the Council on Nutrition, Physical Activity and Metabolism, Council on Arteriosclerosis, Thrombosis and Vascular Biology, Council on Cardiovascular Disease in the Young, Council on Cardiovascular Radiology and Intervention, Council on Cardiovascular Nursing, Council on Epidemiology and Prevention, Council on the Kidney in Cardiovascular Disease, and Stroke Council. Assessing adiposity: a scientific statement from the American Heart Association: A scientific statement from the American heart association. Circulation. 2011;124:1996–2019. [DOI] [PubMed] [Google Scholar]
  • 3.Ndumele CE, Rodriguez F, Dixon DL, Khan SS, Mukherjee D, Bajaj M, Bangalore S, Bozkurt B, Breathett K, Clarke SL, de Boer IH, Ellison DH, Evangelista LS, Heffron SP, Kazi DS, Kulshreshtha A, Lingvay I, Low Wang CC, Mercado CA, Morton JM, Neeland IJ, Pagidipati N, Powell-Wiley TM, Rangaswami J, Rao G, Reza N, Saeed A, St Peter W, Starks JB, Sterling M, Talbot AW, Tran AH, Tuttle KR, VanWagner LB, Vest AR, Virani SS. 2026 AHA/ACC/ADA/ASN guideline for the Prevention, Detection, Evaluation, and Management of cardiovascular-kidney-Metabolic Syndrome: A report of the American college of cardiology/American heart association joint committee on clinical practice guidelines. J Am Coll Cardiol. 2026;87:e1889–e2007. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Oikonomou EK, Antoniades C. The role of adipose tissue in cardiovascular health and disease. Nat Rev Cardiol. 2019;16:83–99. [DOI] [PubMed] [Google Scholar]
  • 5.Neeland IJ, Ross R, Després J-P, Matsuzawa Y, Yamashita S, Shai I, Seidell J, Magni P, Santos RD, Arsenault B, Cuevas A, Hu FB, Griffin B, Zambon A, Barter P, Fruchart J-C, Eckel RH, International Atherosclerosis Society, International Chair on Cardiometabolic Risk Working Group on Visceral Obesity. Visceral and ectopic fat, atherosclerosis, and cardiometabolic disease: a position statement. Lancet Diabetes Endocrinol. 2019;7:715–725. [DOI] [PubMed] [Google Scholar]
  • 6.Hu T, Freeze J, Singh P, Kim J, Song Y, Wu H, Lee J, Al-Kindi S, Rajagopalan S, Wilson DL, Hoori A. Artificial intelligence prediction of cardiovascular events using opportunistic epicardial adipose tissue assessments from computed tomography calcium score. JACC Adv. 2024;3:101188. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.West HW, Siddique M, Williams MC, Volpe L, Desai R, Lyasheva M, Thomas S, Dangas K, Kotanidis CP, Tomlins P, Mahon C, Kardos A, Adlam D, Graby J, Rodrigues JCL, Shirodaria C, Deanfield J, Mehta NN, Neubauer S, Channon KM, Desai MY, Nicol ED, Newby DE, Antoniades C, ORFAN Investigators. Deep-learning for epicardial adipose tissue assessment with computed tomography: Implications for cardiovascular risk prediction. JACC Cardiovasc Imaging. 2023;16:800–816.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Ng ACT, Strudwick M, van der Geest RJ, Ng ACC, Gillinder L, Goo SY, Cowin G, Delgado V, Wang WYS, Bax JJ. Impact of epicardial adipose tissue, left ventricular myocardial fat content, and interstitial fibrosis on myocardial contractile function. Circ Cardiovasc Imaging. 2018;11:e007372. [DOI] [PubMed] [Google Scholar]
  • 9.Antonopoulos AS, Margaritis M, Verheule S, Recalde A, Sanna F, Herdman L, Psarros C, Nasrallah H, Coutinho P, Akoumianakis I, Brewer AC, Sayeed R, Krasopoulos G, Petrou M, Tarun A, Tousoulis D, Shah AM, Casadei B, Channon KM, Antoniades C. Mutual regulation of epicardial adipose tissue and myocardial redox state by PPAR-γ/adiponectin signalling. Circ Res. 2016;118:842–855. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Carena MC, Badi I, Polkinghorne M, Akoumianakis I, Psarros C, Wahome E, Kotanidis CP, Akawi N, Antonopoulos AS, Chauhan J, Sayeed R, Krasopoulos G, Srivastava V, Farid S, Walcot N, Douglas G, Channon KM, Casadei B, Antoniades C. Role of human epicardial adipose tissue-derived miR-92a-3p in myocardial redox state. J Am Coll Cardiol. 2023;82:317–332. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Oikonomou EK, Chan K, Patel P, Wahome E, Dangas K, Desai R, Sobirov I, Khera R, Petersen SE, Pugliese F, Rajani R, Nicol E, Kardos A, Adlam D, Kelion AD, Sabharwal N, Screaton N, Greenwood JP, Rodrigues J, Huck D, Shirodaria C, Tomlins P, Siddique M, Sohan Y, Fry S, van Assen M, Blankstein R, Desai MY, Neubauer S, Channon KM, Deanfield J, Antoniades C, ORFAN Consortium. Early prediction of heart failure from routine cardiac CT using radiomic phenotyping of epicardial fat. J Am Coll Cardiol 2026;87:3539–3555. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Eisenberg E, McElhinney PA, Commandeur F, Chen X, Cadet S, Goeller M, Razipour A, Gransar H, Cantu S, Miller RJH, Slomka PJ, Wong ND, Rozanski A, Achenbach S, Tamarappoo BK, Berman DS, Dey D. Deep learning-based quantification of epicardial adipose tissue volume and attenuation predicts major adverse cardiovascular events in asymptomatic subjects. Circ Cardiovasc Imaging. 2020;13:e009829. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Kramer CM, Borlaug BA, Zile MR, Ruff D, DiMaria JM, Menon V, Ou Y, Zarante AM, Hurt KC, Murakami M, Packer M, SUMMIT Trial Study Group. Tirzepatide reduces LV mass and paracardiac adipose tissue in obesity-related heart failure: SUMMIT CMR substudy. J Am Coll Cardiol. 2025;85:699–706. [DOI] [PubMed] [Google Scholar]
  • 14.Bao Y, Hu Y, Shi M, Zhao Z. SGLT2 inhibitors reduce epicardial adipose tissue more than GLP-1 agonists or exercise interventions in patients with type 2 diabetes mellitus and/or obesity: A systematic review and network meta-analysis. Diabetes Obes Metab. 2025;27:1096–1112. [DOI] [PubMed] [Google Scholar]
  • 15.Iacobellis G, Assael F, Ribaudo MC, Zappaterreno A, Alessi G, Di Mario U, Leonetti F. Epicardial fat from echocardiography: a new method for visceral adipose tissue prediction. Obes Res. 2003;11:304–310. [DOI] [PubMed] [Google Scholar]
  • 16.Iacobellis G. Epicardial adipose tissue in contemporary cardiology. Nat Rev Cardiol. 2022;19:593–606. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Chong B, Jayabaskaran J, Ruban J, Goh R, Chin YH, Kong G, Ng CH, Lin C, Loong S, Muthiah MD, Khoo CM, Shariff E, Chan MY, Lajeunesse-Trempe F, Tchernof A, Chevli P, Mehta A, Mamas MA, Dimitriadis GK, Chew NWS. Epicardial adipose tissue assessed by computed tomography and echocardiography are associated with adverse cardiovascular outcomes: A systematic review and meta-analysis. Circ Cardiovasc Imaging. 2023;16:e015159. [DOI] [PubMed] [Google Scholar]
  • 18.Holste G, Oikonomou EK, Tokodi M, Kovács A, Wang Z, Khera R. Complete AI-enabled echocardiography interpretation with multitask deep learning. JAMA. 2025;334:306–318. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Vukadinovic M, Chiu I-M, Tang X, Yuan N, Chen T-Y, Cheng P, Li D, Cheng S, He B, Ouyang D. Comprehensive echocardiogram evaluation with view primed vision language AI. Nature. 2026;650:970–977. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Johnson A, Bulgarelli L, Pollard T, Gow B, Moody B, Horng S, Celi LA, Mark R. MIMIC-IV [Internet]. 2024. [cited 2025 Oct 24];Available from: 10.13026/kpb9-mt58 [DOI] [Google Scholar]
  • 21.Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, Pollard TJ, Hao S, Moody B, Gow B, Lehman L-WH, Celi LA, Mark RG. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10:1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Blaha MJ, DeFilippis AP. Multi-Ethnic Study of atherosclerosis (MESA): JACC focus seminar 5/8. J Am Coll Cardiol. 2021;77:3195–3216. [DOI] [PubMed] [Google Scholar]
  • 23.Chattopadhyay A, Sarkar A, Howlader P, Balasubramanian VN. Grad-CAM++: Improved visual explanations for deep convolutional networks [Internet]. arXiv [cs.CV]. 2017. [cited 2026 May 19];Available from: 10.48550/arXiv.1710.11063 [DOI] [Google Scholar]
  • 24.Haubold J, Baldini G, Parmar V, Schaarschmidt BM, Koitka S, Kroll L, van Landeghem N, Umutlu L, Forsting M, Nensa F, Hosch R. BOA: A CT-based body and organ analysis for radiologists at the Point of care: A CT-based body and organ analysis for radiologists at the Point of care. Invest Radiol. 2024;59:433–441. [DOI] [PubMed] [Google Scholar]
  • 25.Wasserthal J, Breit H-C, Meyer MT, Pradella M, Hinck D, Sauter AW, Heye T, Boll DT, Cyriac J, Yang S, Bach M, Segeroth M. TotalSegmentator: Robust segmentation of 104 anatomic structures in CT images. Radiol Artif Intell. 2023;5:e230024. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Collins GS, Dhiman P, Andaur Navarro CL, Ma J, Hooft L, Reitsma JB, Logullo P, Beam AL, Peng L, Van Calster B, van Smeden M, Riley RD, Moons KG. Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ Open. 2021;11:e048008. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Antonopoulos AS, Oikonomou EK, Antoniades C, Tousoulis D. From the BMI paradox to the obesity paradox: the obesity-mortality association in coronary heart disease. Obes Rev. 2016;17:989–1000. [DOI] [PubMed] [Google Scholar]
  • 28.Oikonomou EK, Vaid A, Holste G, Coppi A, McNamara RL, Baloescu C, Krumholz HM, Wang Z, Apakama DJ, Nadkarni GN, Khera R. Artificial intelligence-guided detection of under-recognised cardiomyopathies on point-of-care cardiac ultrasonography: a multicentre study. Lancet Digit Health. 2025;7:e113–e123. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Lincoff AM, Brown-Frandsen K, Colhoun HM, Deanfield J, Emerson SS, Esbjerg S, Hardt-Lindberg S, Hovingh GK, Kahn SE, Kushner RF, Lingvay I, Oral TK, Michelsen MM, Plutzky J, Tornøe CW, Ryan DH, SELECT Trial Investigators. Semaglutide and cardiovascular outcomes in obesity without diabetes. N Engl J Med. 2023;389:2221–2232. [DOI] [PubMed] [Google Scholar]
  • 30.Jastreboff AM, le Roux CW, Stefanski A, Aronne LJ, Halpern B, Wharton S, Wilding JPH, Perreault L, Zhang S, Battula R, Bunck MC, Ahmad NN, Jouravskaya I, SURMOUNT-1 Investigators. Tirzepatide for obesity treatment and diabetes prevention. N Engl J Med. 2025;392:958–971. [DOI] [PubMed] [Google Scholar]
  • 31.Reijrink M, de Boer SA, Spoor DS, Lefrandt JD, Lambers Heerspink HJ, Boellaard R, Greuter MJ, Borra RJH, Hillebrands J-L, Slart RHJA, Mulder DJ. Visceral adipose tissue volume is associated with premature atherosclerosis in early type 2 diabetes mellitus independent of traditional risk factors. Atherosclerosis. 2019;290:87–93. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

Model weights and an inference pipeline for PanAdipo are intended for release at the time of publication on our Lab’s GitHub repository. Underlying clinical imaging and electronic health record data from the Yale-New Haven Health System are not publicly redistributable. MIMIC-IV is available through PhysioNet credentialed access. This manuscript was prepared using MESA Research Materials obtained through the NHLBI Biologic Specimen and Data Repository Information Coordinating Center (BioLINCC) and/or the NHLBI BioData Catalyst® ecosystem under approved data use agreements. The content is solely the responsibility of the authors and does not necessarily represent the official views of the MESA investigators, BioLINCC, BioData Catalyst, the NHLBI, or the National Institutes of Health (NIH).

Articles from medRxiv are provided here courtesy of Cold Spring Harbor Laboratory Preprints

RESOURCES