I’m a heavy Bevel user and would love to see Bevel lead on body‑composition tracking the same way it does on fitness and metabolic health.

Right now, most consumer apps rely on weight + BMI and maybe import rough body‑fat numbers from a bioimpedance scale. The problem is: those scales are not very accurate for fat mass. A 2021 observational study on three popular smart scales (including Withings Body Cardio) found they underestimated fat mass by about 2–4 kg vs DEXA and concluded they should not be used as a DXA replacement for body composition.¹

At the same time, there’s now strong peer‑reviewed evidence that computer‑vision from smartphone photos can match or beat these scales and get much closer to DXA:

An Amazon‑led npj Digital Medicine study validated a “visual body composition” (VBC) algorithm using two smartphone photos against DXA. VBC had a mean absolute error of ~2.2% body fat vs DXA and outperformed all 5 BIA devices and BodPod they tested, with the tightest limits of agreement and no significant bias.²

Spren has shown similar results. Their 2024 white paper reports a mean absolute error ≈2.3 with r ≈0.96 vs DXA, and when they compare against the same smart‑scale study above, Spren’s median fat‑mass error is only 0.6 kg, versus –2.2 to –4.4 kg for consumer scales.¹³

A 2025 npj Digital Medicine paper on an AI 2D‑photo method (1,273 adults) found the AI‑photo approach had the highest agreement with DXA (CCC ≥0.96) compared with multi‑frequency BIA devices, skinfolds, and ultrasound. The authors explicitly describe AI‑photo as potentially interchangeable with DXA for body‑fat estimation in practice.⁴

More recently, Spren’s Pennington Biomedical validation study reported ~2.6% mean absolute error vs three independent DXA machines, placing their smartphone scans within the natural measurement variability of the DXA devices themselves and outperforming most hardware‑based systems.⁵

Given this, I’d love to see Bevel build a camera (and optionally LiDAR)–based “visual body composition” module into the app. Concretely, something like:

A guided 2–3 photo (or short video) flow using the rear iPhone camera (and LiDAR where available) to capture full‑body geometry.

On‑device or cloud processing that estimates body‑fat %, lean mass, and key circumferences (waist, hips, thighs, etc.), calibrated against DXA‑level algorithms.

A simple “DEXA‑like” body‑comp report inside Bevel, with longitudinal tracking and the ability to overlay and correlate with lab markers, CGM data, weight, and training logs.

Privacy‑safe design (all images processed locally or securely, with optional “no image storage” mode) so that this feels acceptable to health‑conscious users and clinicians.

This would let Bevel:

Give users a DXA‑adjacent body‑composition signal without asking them to buy any hardware.

Avoid leaning on low‑accuracy BIA scale data.

Deepen Bevel’s differentiation as the “health OS” that integrates labs, glucose, and now genuinely high‑quality body‑composition trends.

The research and tooling clearly show this is feasible today. If Bevel shipped a well‑validated visual body‑composition feature (especially one that can be used alongside DEXA for calibration), it would be a huge value‑add for users who care about both aesthetics and cardiometabolic health.

If others reading this would also use this, please upvote / comment so the team can see the demand. 🙏

1) Frija-Masson J et al. “Accuracy of Smart Scales on Weight and Body Composition: Observational Study.” JMIR mHealth and uHealth, 2021.

2) Majmudar MD et al. “Smartphone camera based assessment of adiposity: a validation study.” npj Digital Medicine, 2022.

3) Wu S. “Validating Accuracy: A Deep Dive into Spren Vision™ for Body Composition Analysis.” Spren White Paper, 2024.

4) Ferreira TJ et al. “Advances in the estimation of body fat percentage using an artificial intelligence 2D-photo method.” npj Digital Medicine, 2025.

5) Spren. “Spren Achieves Gold Standard Body Composition Accuracy Using Only Your Phone Camera.” Press release, Pennington Biomedical validation, 2025.