Machine learning mapped brown fat in 8,709 patients from over 100,000 clinical scans
A new study in Nature uses a machine-learning model to map brown fat's health associations at a scale that was not practically possible with manual analysis. The study links detectable brown fat to lower rates of type 2 diabetes and heart disease — and a blood pressure drug emerges as a candidate for further research, though so far only in cell cultures and mice.
The study, published in Nature on 14 October 2026, builds on a classic insight in human physiology: brown adipose tissue (BAT) burns calories and appears to be linked to metabolic health. But the evidence base has been scattered, because active brown fat can only be detected with PET/CT scans, and the volume of imaging data has made large studies impractical. That is where machine learning comes in.
How AI made the study possible
The researchers used 18F-FDG-PET/CT scans from 101,243 adult patients at Memorial Sloan Kettering Cancer Center (MSKCC), taken between June 2009 and July 2024. The scans were clinical, not taken for research purposes; the study is an example of "phenotyping" existing patient data at large scale.
An AI image-analysis tool developed at MSKCC quantified brown fat activity semi-automatically in 8,709 patients with detectable BAT. The measure combines two elements: uptake of the glucose tracer (a sign that the fat tissue is "switched on" and burning energy) and tissue volume. The tool measured both total activity in each patient and activity in six anatomical depots — for example axillary, in the armpit.
First author Nicolás Gómez-Banoy, an endocrinologist and instructor in clinical investigation in Paul Cohen's group at Rockefeller University, applied the tool developed by colleagues at MSKCC. The analysis covered more than 500 diagnoses, compared with roughly 50,000 patients in the group's earlier study from 2021 — a doubling that would not have been practically possible with manual quantification. For statistical control, the researchers used propensity-score matching on age, sex, BMI and outdoor temperature, since cold affects brown fat activity.
What the study found
According to News-Medical, detectable BAT was associated with lower rates of cardiometabolic and psychiatric conditions, including type 2 diabetes, depression and anxiety. According to UA.News, citing News-Medical, people with detectable brown fat had 58% lower odds of type 2 diabetes and 47% lower odds of chronic ischemic heart disease.
The effects varied between depots: according to Gómez-Banoy, axillary brown fat was linked to lower risk of high blood pressure compared with other regions. "But overall, patients with brown fat have better glycemic parameters, lipid profiles and echocardiographic measures, as well as lower inflammation markers," he says in the press release from Rockefeller University.
The losartan lead
Further analyses showed that patients taking angiotensin receptor blockers (ARBs) — a class of blood pressure medications — had more brown fat on their scans. The researchers tested losartan directly in cell cultures: according to the reporting, the drug increased levels of the UCP1 protein and oxygen consumption in brown fat cells, via mechanisms described as different from its blood-pressure-lowering effect. In mice, losartan increased energy expenditure both at room temperature and under cold exposure, without measurable weight reduction.
The limits are important: the effect of losartan on human brown fat has not been tested in clinical trials. This is a preclinical finding, not an established therapeutic option.
The limitations
Irfan Lodhi, a molecular biologist at Washington University in St. Louis, points out in Nature's news coverage that the study is retrospective: "Because this is a retrospective study, more work is needed to prove that brown fat actually causes all these benefits," he says.
The study is correlational, not causal — it shows associations, not that brown fat prevents disease. The data also come from a single cancer center (MSKCC), which limits generalizability to healthy populations. While machine learning makes it possible to analyse these scans, it cannot replace future studies that follow patients over time or intervene directly.
Why the AI method matters
The value here lies less in the individual findings than in the method: a machine-learning tool made it possible to extract a physiological signature — brown fat activity across six depots — from a clinical imaging archive of over 100,000 scans, and then test it against more than 500 diagnoses. Manual quantification would not have been practically feasible at this scale.
This provides a template for future research: large, retrospective cohorts built on already existing clinical images, where machine learning identifies patterns and generates hypotheses that can later be tested in prospective studies. The study's biggest contribution may not be the answers it gives, but the proof that such a pipeline is possible.