A Body-Fat Scale Added Nothing to Insulin-Resistance Prediction. A Phone Photo Came Within 1.3 Points of a Full DXA Scan.

A bioelectrical-impedance scale — the sensor already built into millions of bathroom scales and gym check-ins, designed to measure body composition directly — was tested against insulin-resistance risk and added nothing: not one point of AUROC, not one correctly reclassified patient. A single smartphone photo, tested the identical way on the identical cohort, lifted classification from an AUROC of 0.692 to 0.760 — just 1.3 points behind a full-body DXA scan's 0.773. Both devices are trying to estimate the same thing: how much fat someone is carrying, and where. Only one of them moved the number that actually predicts disease risk.

What PhotoScan actually does

"Beyond BMI: Smartphone Body Composition Phenotyping for Cardiometabolic Risk Assessment" (Zhou, Charton, Blanchard, et al., Google Research) trains a deep-learning model to predict three DXA-derived body-composition measurements — total body fat percentage, the android-to-gynoid fat ratio, and the visceral-to-subcutaneous fat area ratio — directly from a photo taken with an ordinary phone camera. The model is pretrained on 35,323 UK Biobank participants who have both a DXA scan and body imagery on file, then fine-tuned on a purpose-built clinical cohort, PhotoBIA (n=677), and tested on a second, independent cohort, MetabolicMosaic (n=132), that the model never saw during training. Reported August 17, 2026 on the Google Research blog, the headline number is accuracy: on the validation cohort, the photo-based estimate of body fat percentage carried a mean absolute error of 2.13–2.15%, with the android-to-gynoid and visceral-to-subcutaneous ratios both accurate to roughly 0.09.

Three ways to estimate body composition, compared

Accuracy against a DXA scan is one thing. Whether that accuracy translates into a better clinical prediction is a separate, harder question, and it is the one the paper actually answers. The team took a baseline model using only age, sex, and BMI — the information already sitting in almost every electronic health record — and tested what each body-composition method added to it, on the independent MetabolicMosaic cohort, for classifying insulin resistance.

Demographics alone scored an AUROC of 0.692. Adding the photo-derived body-composition estimate moved that to 0.760, with a net reclassification improvement of 0.593 — meaning the added information correctly moved a meaningful share of patients into a more accurate risk category. Adding the full DXA scan, the actual gold-standard measurement, moved AUROC to 0.773 and NRI to 0.748 — better, as it should be, but not by much. Adding the bioelectrical-impedance measurement, in the same cohort, under the same test, produced no improvement in AUROC or NRI over demographics alone.

The number that undercuts the intuition

The intuitive ranking going into this study would put DXA first, BIA a close second — it is, after all, a direct electrical measurement of body tissue, not an inference from a photograph — and a phone camera a distant third, useful mainly for its convenience. The insulin-resistance numbers invert the middle of that ranking. BIA estimates body composition by passing a small current through the body and measuring resistance, which depends heavily on hydration, recent food and water intake, and time since exercise — conditions a single scale reading cannot control for and does not know it is missing. A photo, by contrast, is capturing regional shape: where fat sits on the body, not just how much of it there is in total, which is closer to what the android-to-gynoid and visceral-to-subcutaneous ratios are actually measuring, and closer to what predicts cardiometabolic risk in the first place. The paper does not run the mechanistic study that would fully explain the gap; what it does show, cleanly, is that a device's physical proximity to the underlying biology is not the same claim as a device carrying useful signal for the specific outcome being predicted.

The honest limits

MetabolicMosaic, the cohort every headline number above comes from, has 132 people in it — enough to support the comparison the paper makes, not enough to call this a settled clinical result. The pretraining population, UK Biobank, skews toward a narrower age and demographic band than a general primary-care population, and body-composition estimation from images is a category with a known history of uneven accuracy across skin tones and body types when training data is not deliberately balanced for it; the paper does not report a stratified breakdown on that axis. PhotoScan has not been through regulatory clearance and is not a diagnostic device today — it is a research result, published by the company that built it, on cohorts the same company assembled, which is a real result but not yet an independently replicated one. And the gap that matters most for anyone deciding whether to trust this: DXA's NRI advantage over the photo method (0.748 vs. 0.593) is larger than its AUROC advantage (0.773 vs. 0.760) suggests on its own — DXA is still meaningfully better at correctly reclassifying individual patients, even though the two methods look close on the headline curve.

What this means for reproducible, local-first science

The finding worth sitting with is not "phones can replace DXA scanners" — they cannot, and the NRI numbers say so directly. It is that the sensor everyone already has in their pocket captured more of the signal that matters than a purpose-built consumer health device did, on the exact task both were tested against. That is a useful correction for any team building on top of a device's marketing claim instead of its measured output: a scale that reports a body-fat percentage to one decimal place is not the same claim as a scale whose number moves a real clinical prediction, and the only way to know which one you have is to run the comparison this paper ran — the same outcome, the same cohort, every candidate sensor tested against the same baseline, with the result exported alongside the exact model and cohort it came from rather than asserted on its own.