Do brain signatures of BMI predict cardiometabolic and respiratory disease?

A model trained on weight, tested on disease

The paper asks a narrow question with a broad implication: if a neural network learns to read body mass index (BMI) from a structural brain scan, does the representation it builds carry information about systemic disease? BMI here is the standard weight-for-height measure, split into underweight (below 18.5 kg/m²), healthy weight (18.5–24.9), overweight (25.0–29.9) and obesity (30.0 or above).

The authors trained a 3D convolutional neural network on T1-weighted MPRAGE scans to predict BMI, then reused the network's final 64-value feature map as a "digital biomarker" for disease. That feature map is the model's compressed summary of a scan, produced by five convolutional blocks that shrink a 160×192×160 voxel input to a 5×6×5 volume with 64 channels. Disease classifiers were then fit on those 64 numbers rather than on the image.

The design matters because the network never saw a diagnosis during BMI training. Any disease signal in those 64 values is a by-product of learning to estimate body mass, not of being optimized for disease.

Six cohorts, three of them held out

The pooled dataset contained 45,702 subjects with one image each: UK Biobank (42,754), HCP Young Adult (1,109), HCP-Aging (719), and healthy controls from ADNI (822), PPMI (142) and PREDICT-HD (156). UK Biobank, HCP and HCP-Aging were used for training and within-cohort evaluation; ADNI, PPMI and PREDICT-HD were held out entirely as external test cohorts.

Within the three training cohorts, 70% of samples (n = 31,206) trained the model, 20% (n = 8,917) tuned hyperparameters, and 10% (n = 4,459) formed a blind within-cohort test set. Training used balanced sampling — 12,000 UK Biobank images plus 2,000 each from HCP and HCP-Aging per epoch — so the largest cohort would not dominate and BMI would not be tied to a particular age range.

Rather than regressing a single number, the model divided the BMI range 10–60 into 50 equal bins and output a probability for each; the final estimate is the probability-weighted average of the bins. This is the operational definition of "predicted BMI" throughout the paper.

Strong inside cohorts, weaker outside them

The model performed strongly on within-cohort test data but lost accuracy when transferred to the held-out cohorts. The paper reports this as a consistent pattern rather than a single number: strong within-cohort performance, reduced external performance. The authors note that the external cohorts are screened healthy controls drawn from neurodegenerative-disease studies and skew older, so they are not general-population samples — a factor that may itself contribute to the drop.

That asymmetry is the first boundary on everything that follows. BMI prediction was tested in fully external cohorts; the disease analyses were not.

A separate longitudinal check used a follow-up subset and found the model tracked BMI change over 2.3 years. Sensitivity was stronger for BMI increases than decreases, and stronger in obese participants than in the healthy-weight group, while the overweight and underweight groups showed no significant difference from healthy weight. The paper also reports that the standard deviations of BMI change grow progressively from healthy weight to obesity, which is part of why the obese group is easier to track.

The 64 features outperform BMI on all eight diagnoses

The disease analysis was confined to the UK Biobank subset with clinical labels (n = 34,524) and covered eight conditions with at least 1% prevalence: atrial fibrillation/flutter, coronary artery disease, type 2 diabetes, hypercholesterolemia, hypertension, myocardial infarction, heart failure and chronic obstructive pulmonary disease. Each was encoded as a binary label.

For each disease, a logistic regression classifier was trained on either BMI or the model-derived biomarkers, with tenfold cross-validation; 90% of samples trained the classifier and 10% were held out in each fold. The biomarker-based models achieved significantly higher AUCs than BMI-based models in all eight diseases (p < 0.05, Wilcoxon signed-rank test). AUC here is the area under the receiver-operating-characteristic curve, a discrimination measure — not a risk score or a calibrated probability.

The authors also report that matched-dimension comparisons confirm the advantage is not explained by representational capacity alone, and that calibration was assessed only post hoc with the Brier score. The disease models were not optimized for absolute-risk estimation.

White matter in cerebellum, callosum and brainstem drives predictions

To see where the model was sensitive, the authors used integrated gradients, which quantify how perturbing individual voxel intensities would shift the predicted BMI, producing a per-subject importance map. Maps were warped to MNI space, normalized, and summarized by principal component analysis.

The first principal component separated subjects almost cleanly by BMI status: the positive cluster consisted almost exclusively of overweight and obese individuals. Its loadings were dominated by white matter following anatomical boundaries, with the largest clusters along the corpus callosum and cerebellum and additional contributions from brainstem pathways (cerebral peduncles, medial lemniscus), limbic structures (mamillary bodies, fornix, cingulum) and major tracts such as the optic radiation and uncinate fasciculus. White matter dominance held across most principal components.

The paper states that these white matter regions on their own detected disorders as well as the full model. It also cautions that importance maps are a simplification — the decision is made by non-linear relationships across all voxels — and should be read as where the model is sensitive overall, not as validated anatomical localization.

Smoking, not alcohol, shows up in the BMI space

Projecting the biomarkers into a two-dimensional UMAP space — used only for visualization, never as a prediction feature — showed orderly structure by BMI category and by sex. Mapping lifestyle variables onto that space gave a split result: alcohol intake showed no association with BMI along the first dimension, while daily smoking separated into two distributions.

Heavy smokers (more than 10 cigarettes per day) clustered predominantly on the left side of the space, and people on that side had higher BMI values and increased risk of obesity-related disease. Individuals smoking 10 or fewer cigarettes per day were evenly distributed. The authors read this as evidence that the MRI biomarkers reflect smoking status in addition to disease status.

That reading is a caution as much as a finding: a representation built from BMI is not a pure BMI signal, and the disease classifiers may be drawing on correlated lifestyle information.

Shared mechanisms are proposed, not demonstrated

The authors suggest the results are consistent with shared mechanisms linking metabolic state and brain structure, and name candidates: inflammation, which has been reported to mediate brain effects of obesity; cerebrovascular dysfunction and blood–brain barrier disruption; and expression of obesity-related genes in brainstem regions such as the nucleus tractus solitarius. They explicitly call for future studies to disentangle these possibilities.

The brainstem localization is given functional context: the medulla oblongata and pons house nuclei that generate and modulate breathing rhythm and autonomic cardiovascular control, integrating chemosensory and baroreceptor input. The paper frames this as alignment with existing literature, not as a demonstrated pathway.

The authors are direct that the white matter observations are specific to this cross-sectional BMI model and should not be interpreted as mechanistic claims. The gap between "the model is sensitive here" and "this region causes the disease link" is not closed by this study.

The comparison holds inside one cohort, not yet beyond it

The paper's own limitations section sets the boundary precisely. All disease-prediction analyses were confined to the UK Biobank (n = 34,524) and were therefore not externally validated, even though BMI prediction was. Residual confounding by age, vascular burden, smoking or socioeconomic factors cannot be fully excluded, because the brain–BMI representation is more similar to BMI than to age but observational neuroimaging cannot rule these out.

The model used a single architecture without exhaustive comparison across alternatives, and the robustness of importance maps to preprocessing, template choice and architecture — plus test–retest reliability of the latent biomarker — was not assessed. Generalizability to other scanners, acquisition protocols and clinical populations remains to be established.

Given those constraints, the paper's most decision-relevant claim is a comparison rather than a deployment: for the same eight diagnoses in the same UK Biobank sample, a representation learned from brain structure alone discriminated disease better than the BMI value used to train it. That reframes brain imaging as a readout of whole-body physiology and makes the next question mechanistic — which process, inflammation or vascular burden, writes that trace — rather than whether the AUC gap exists.

Sources used in this answer