AI Reads Chest X-rays to Spot Early Heart Dysfunction in Obesity and Diabetes

An artificial intelligence tool that reads ordinary chest x-rays may be able to detect early, symptom-free heart dysfunction in people living with combined obesity and type 2 diabetes, and may even track how that dysfunction responds to widely used incretin medications such as GLP-1 receptor agonists. That is the tentative conclusion of a small exploratory study published in Cardiovascular Diabetology – Endocrinology Reports, in which researchers at the Mayo Clinic analyzed chest x-rays taken before and after at least twelve months of incretin-based therapy in fifteen patients. The authors emphasize that the work is a proof of concept, not a definitive trial, and that its findings await confirmation in a larger, controlled, prospective study.

The condition the researchers set out to probe is sometimes called diabesity, the combination of obesity and type 2 diabetes mellitus. Together, these two metabolic problems have become recognized as a major cause of heart failure. When they occur in combination, they are particularly damaging to the heart muscle. Fat tissue around and within the abdomen releases free fatty acids that accumulate inside cardiomyocytes, the contractile cells of the heart, at levels exceeding what the cells can store or burn through fatty-acid oxidation driven by insulin resistance. This internal leakage of harmful lipid intermediates injures the endoplasmic reticulum and mitochondria, triggers inflammatory responses, and drives cell death. Associated fibrotic reactions further stiffen the energy-disrupted myocardium, producing subclinical left ventricular diastolic dysfunction, a condition in which the heart’s main pumping chamber becomes stiff and fills improperly, often long before any symptoms appear.

Fat deposits located immediately around the heart add to this injury. Epicardial and pericardial adipose tissue, because of their central location and unique composition, exert relatively greater lipotoxic effects on left ventricular function than fat stored elsewhere. They also worsen diastolic dysfunction mechanically, by physically restraining the filling of the ventricle during diastole. Incretin-based therapies, including glucagon-like peptide-1 receptor agonists such as dulaglutide and semaglutide, and the combined GLP-1/GIP receptor agonist class, are already well recognized for their cardiovascular benefits in obesity and diabetes, and they improve symptoms in patients who already have diabesity-associated HFpEF. What has not been described, according to the authors, is their impact on the asymptomatic pre-clinical phase of the disease.

The Mayo Clinic team hypothesized that a validated artificial intelligence method for staging pulmonary venous hypertension on chest x-rays, also known as pulmonary congestion, could fill this gap. The AI model assigns each image a prediction of absent pulmonary venous hypertension, Stage 1 disease characterized by vascular distention and redistribution with minimal interstitial edema, or Stage 2 or greater disease marked by vascular congestion with at least mild interstitial or alveolar edema. These stages serve as a surrogate for tendencies toward elevation of mean left atrial pressure, which rises in diastolic dysfunction. The staging method had previously been validated for tracking grades of diastolic dysfunction measured by Doppler echocardiography, the standard ultrasound-based assessment, although no concurrent echocardiographic data were available for the fifteen patients in the present study.

To assemble their study group, the investigators searched the shared electronic medical record of the Mayo Clinic enterprise, which spans three major quaternary medical centers and more than seventy smaller facilities internationally, under an institutional review board approval that waived individual consent. The inclusion criteria were strict. Patients needed at least twelve months of continuous GLP-1 or GLP-1/GIP receptor agonist therapy, non-portable digital chest x-rays taken within one month before therapy began and again within one month of completing or after at least twelve months of ongoing therapy, and no known or suspected ventricular or valvular dysfunction that could complicate the assessment. After excluding a handful of patients with potentially confounding conditions, one each with atrial fibrillation, volume overload from advanced chronic kidney disease, secondary pulmonary hypertension, and pulmonary fibrosis, fifteen cases remained.

The fifteen patients spanned a wide range of obesity severity. Sorted by decreasing body mass index within a week of starting therapy, six cases fell into Class 3 obesity with BMI values from 41.0 to 62.6, five into Class 2 with values from 35.4 to 39.8, and three into Class 1 with values from 31.4 to 33.7, while one pre-diabesity case was considered overweight at a BMI of 25.3. Within two months of starting therapy, fourteen cases showed abnormally elevated hemoglobin A1c levels ranging from 6.2 to 12.3 percent, indicating concurrent diabetes in twelve cases and pre-diabetes in two. Therapy lasted twelve to forty-six months between the x-ray monitoring points and consisted of a GLP-1 receptor agonist alone in twelve cases or a combined GLP-1 and GLP-1/GIP regimen in three.

When the researchers grouped the patients by their weight-loss response to therapy, three equal-sized categories emerged. Significant Responders lost 9 to 30 percent of their body weight, consistent with reported target levels of 10 to 20 percent weight loss for incretin-based therapy. Insignificant Responders lost only 0 to 2 percent. Non-Responders actually gained weight, with increases of 2 to 12 percent. The categories also differed in their diabetes control. All Significant Responders showed hemoglobin A1c reductions, ranging from mild decreases of 0.5 to 0.9 percentage points to moderate decreases of 1.0 to 1.9 percentage points, with normal levels achieved twice. Insignificant Responders showed varied changes, including two decreases and two moderate increases, while Non-Responders were largely stable except for two reductions.

The x-ray findings tracked these metabolic patterns. Before therapy, all fifteen cases showed evidence of diastolic dysfunction-related left atrial pressure elevation on AI-enabled staging, eleven at Stage 1 and four at Stage 2 or greater. After therapy, four cases improved, five remained stable, and six worsened. Improvement was concentrated among the patients with the greatest weight loss or hemoglobin A1c reduction, including one Significant Responder who lost 30 percent of body weight and transitioned from Stage 2 or greater down to Stage 1. By category, Significant Responders collectively showed stable-to-decreased dysfunction, Insignificant Responders showed stable-to-increased dysfunction in 60 percent of cases, and Non-Responders showed increased dysfunction in 80 percent, the exception being the initially pre-diabesity case. The authors caution that these observations are semi-quantitative, because the tiny sample size risks spurious and imprecise results if formal statistical significance were claimed.

Several limitations temper the excitement. The study was retrospective and descriptive, drawn from a single healthcare enterprise, and involved only fifteen heterogeneous cases. Because the patients were asymptomatic, Doppler echocardiography had not been consistently performed, and when it was, the examinations were not synchronized with the x-rays at the therapy endpoints, so the AI staging results could not be validated against concurrent ultrasound in this cohort. The authors also note that echocardiography itself can be inconclusive about diastolic dysfunction, in roughly 14 percent of cases under the 2016 recommendations, and that the AI staging could misrepresent the current status of left ventricular filling at any particular moment, since left atrial pressure elevations are likely intermittent due to fluctuating loading conditions. The study also did not image changes in visceral, epicardial, or pericardial fat deposits.

Despite these caveats, the authors argue that the observed pattern justifies further development. They note that chest x-ray-based evaluation could become a practical, affordable, and widely available enhancement to diabesity care, easily integrated into routine monitoring, and that representatives of all three initial BMI subgroups appeared in all three therapeutic categories, suggesting the approach may apply across the obesity spectrum. Whether the initial BMI levels affected the diagnostic value of the AI staging remains unclear. The team plans a larger, definitive, and controlled prospective study to confirm or disprove these preliminary results, incorporating revalidation of the AI staging against Doppler echocardiographic grading under newer recommendations. Until then, the message is deliberately restrained: an ordinary chest x-ray, read by an algorithm, may one day offer an early warning of stiffening hearts in obesity and diabetes, and a way to watch incretin therapy work, but that promise remains unproven.

Subject of Research: AI-enabled chest x-ray detection of subclinical diastolic dysfunction in obesity and type 2 diabetes and its response to incretin therapy

Article Title: AI-enabled chest x-ray potentially detects subclinical diastolic dysfunction in diabesity and monitors its functional responses to GLP-1 or GLP-1/GIP receptor agonists: initial exploration and proof-of-concept

Article References: White, R. D., Demirer, M., Erdal, B. S., & Lin, G. (2026). AI-enabled chest x-ray potentially detects subclinical diastolic dysfunction in diabesity and monitors its functional responses to GLP-1 or GLP-1/GIP receptor agonists: initial exploration and proof-of-concept. Cardiovascular Diabetology – Endocrinology Reports, 12(1), Article 51. https://doi.org/10.1186/s40842-026-00312-5

Image Credits: AI Generated

DOI: 10.1186/s40842-026-00312-5

Keywords: artificial intelligence, chest x-ray, diastolic dysfunction, obesity, type 2 diabetes, GLP-1 receptor agonists, incretin therapy, pulmonary venous hypertension, heart failure, proof-of-concept study, AI-enabled, chest