What caught my attention isn't the headline accuracy — MAE around 3.8 % across BMI 18–35 — it's the architecture choices that make this deployable on-device. The backbone is a MobileNetV3-small variant stripped to 1.2 M parameters, quantized to INT8 with TensorFlow Lite. That puts inference at ~35 ms on the Tensor G2 NPU, leaving plenty of thermal budget for the camera pipeline. Google didn't publish the full training recipe, but the supplementary material hints at a two-stage curriculum: first a synthetic pre-train on 3D human meshes (SMPL-X + randomized lighting/backgrounds), then fine-tune on the real DEXA pairs with heavy mixup and cutout augmentation. Smart way to sidestep the scarcity of labeled medical imaging data.
Privacy model is the usual on-device inference pitch: image never leaves the phone, only the scalar output hits the Health Connect API if the user opts in. Still, the attack surface is non-trivial. A malicious app with camera permission could theoretically reconstruct a rough body-shape embedding from repeated queries — membership inference on the latent space, basically. Google says they add Gaussian noise (σ = 0.02) to the logits before returning the value, but the noise budget isn't user-configurable and there's no formal DP guarantee in the whitepaper.
Clinical utility? Debatable. DEXA remains gold standard for visceral vs. subcutaneous differentiation; a 2D projection fundamentally cannot separate the two. For population-level screening or longitudinal self-tracking it's probably fine — think "am I trending up or down over three months" rather than "do I have metabolic syndrome." The error bars widen sharply above 35 % body fat and below 18 %, exactly where clinical decisions matter most.
Integration path looks straightforward for Android devs: new BodyCompositionManager
class in Play Services 24.20, single estimateBodyFat(bitmap)
call returns a BodyFatResult
with confidence interval. iOS support is "under evaluation" — read: waiting for Core ML conversion to finish.
Biggest unanswered question: generalization across ethnicity and age. The DEXA cohort was 72 % White, 14 % Asian, 9 % Black, 5 % Other, mean age 34. If the model systematically underestimates visceral fat in South Asian phenotypes (known higher cardiometabolic risk at lower BMI), that's a health-equity issue waiting to happen. Google says a broader validation study is underway; no timeline given.
Would I ship a feature on top of this today? For a consumer wellness dashboard with clear "not medical advice" disclaimers — yes. For anything that touches insurance underwriting or clinical pathways — hard no until external replication lands.
Google drops twelve billion on Marvell for next-gen TPU work 1h ago Google is buying up Spirit Airlines data for some reason 1d ago
Google just bought Spirit Airlines' data at auction to feed its 2d ago
Google is dropping $10 million on Spirit Airlines business data 2d ago
Google just snagged a goldmine of Spirit Airlines data from a 2d ago
Google is buying Spirit Airlines' data to feed its AI models 2d ago
Next Rethinking LLM scaling after Jie Tang's latest breakdown →