A Body-Fat Scale Added Nothing to Insulin-Resistance Prediction. A Phone Photo Nearly Matched a DXA Scan. Here's Everything Else That Shipped.

A bioelectrical-impedance scale, tested against insulin-resistance risk, added nothing over demographics alone. A smartphone photo, tested the identical way, lifted classification to within 1.3 AUROC points of a full DXA body scan. That result gets the full breakdown below. Five more results shipped this week: an AMIE video-consultation AI rated on par with primary-care physicians across 300 simulated visits, and significantly better at guiding a physical exam; two same-day preterm-birth papers that reach opposite conclusions on the identical measurement problem depending on dataset size and how strictly patient leakage is controlled; a prostate-MRI reconstruction model that passes a 1,000-case blinded reader study on three sequences and fails it on a fourth; an in-frame indel pathogenicity classifier scored across more than 372,000 variants; and an external validation showing four LLMs get 74% of their generated gene-disease associations right, with the errors concentrated in rare diseases and SNP-level claims.

HealthTech: the wrong sensor can measure the right thing and still add nothing

Full breakdown in today's feature. Short version: Google Research's PhotoScan model estimates body fat, android-to- gynoid ratio, and visceral-to-subcutaneous ratio from a single phone photo, pretrained on 35,323 UK Biobank participants and validated on an independent 132-person cohort. Added to a demographics-only baseline (AUROC 0.692), the photo model lifted insulin-resistance classification to 0.760 — a full DXA scan reached 0.773. A bioelectrical-impedance scale, tested the same way, produced no improvement in AUROC or reclassification at all.

Clinical & translational AI: an AI held its own on video, and did better at the exam

Google's " Advancing AMIE towards expert-level audio-visual clinical consultations" (Google Research, published August 11, 2026) ran 300 simulated consultations across 100 clinical scenarios, comparing AMIE's video mode against 10 board-certified primary care physicians, scored blind by 20 independent PCP evaluators. AMIE (Video) rated on par with PCPs on history-taking, diagnostic accuracy, management, and communication — and significantly higher on eliciting physical signs and guiding virtual exam maneuvers, an area text-only AI consultation cannot touch. The caveat the team states plainly: every consultation used professional patient actors in a simulated setting, not real patients with their own conditions.

Clinical & translational AI: two papers, one dataset, opposite conclusions

Posted the same day, two independent teams built electrohysterography-based classifiers for preterm birth and reached opposite readings of how solved the problem is.

Neither paper cheated: both used record-grouped splits meant to prevent the same patient from leaking between training and test folds. The divergence comes from scale and feature choice — the full 300-record benchmark, evaluated with 92 elastic-net features, finds a signal indistinguishable from chance and has to abstain on nearly three-quarters of cases to keep any usable accuracy; the 26-recording curated subset, evaluated with 14 decomposition-based features across nine classifiers, finds a strong signal. Read alone, either paper tells a clean story. Read together, they are a benchmark-validity finding: how much you trust an EHG classifier depends heavily on which of these two evaluation designs produced the number in front of you.

MedTech: a reconstruction model that discloses exactly where it fails

"A cross-modal generative model for incomplete and degraded prostate MRI with multicentre clinical validation" (Ma, He, Zhu, et al.; bioRxiv, DOI 10.64898/2026.08.16.745066, posted August 18, 2026) introduces MSCNet, which reconstructs missing or degraded MRI sequences from the ones a scan did capture. Across a 200-case diagnostic set, AUC for detecting clinically significant prostate cancer was 0.860 on real images, 0.841 on MSCNet-completed images, and 0.797 for a baseline generative model. In a blinded 1,000-case reader study, MSCNet met non-inferiority for DWI, ADC, and T2W sequence completion — and failed it for T1W, a disclosed limitation rather than a buried one.

Genomics: a calibrated classifier for a variant class other tools skip

"INDELVAR: structure-informed prediction of in-frame indel pathogenicity with calibrated PP3/BP4 thresholds" (Ji, Oh, Kim; bioRxiv, DOI 10.64898/2026.08.13.737497, posted August 18, 2026) scores in-frame insertions and deletions of 1–10 amino acids using AlphaFold-derived structural context, evolutionary conservation, and gene constraint. It reaches an AUROC of 0.980 in cross-validation and 0.977 on an independent test set, and publishes precomputed scores for more than 372,000 GRCh38 indels. The honest ablation: against a recent protein-language-model baseline, INDELVAR's edge narrows to a difference the authors call not statistically significant — the gain over older, non-neural baselines is real and large; the gain over the best neural baseline is not yet proven.

Genomics: checking what an LLM says it knows against what is actually known

"Assessing the Reliability of LLM-Generated Phenotype-Genotype Associations Through External Validation" (Sun, Xin, Zeng, et al.; bioRxiv, DOI 10.64898/2026.08.13.744701, posted August 18, 2026) had four LLMs generate 4,196 phenotype-genotype associations, then ran every claim through external verification against the GWAS Catalog, OMIM, and Ensembl. 74.19% matched an external knowledge base, with Claude Sonnet the strongest model. The gap that matters: accuracy held up well for phenotype-gene associations but dropped sharply for phenotype-SNP claims and for rare-disease associations specifically — exactly the cases where an external check is hardest to skip and most necessary to run.

What this means for reproducible, local-first science

The through-line this week is the same one this digest keeps returning to: a number is only as trustworthy as the condition it was measured under. A sensor's physical closeness to the biology it is measuring does not guarantee it carries signal for the outcome that matters. A benchmark's headline accuracy depends on which of two equally-defensible evaluation designs produced it. A reconstruction model is only honest if it discloses the one sequence type it cannot yet handle. A classifier's real edge is the gap left after comparing it to the best available baseline, not the gap to the oldest one. And an LLM's stated knowledge is only as good as the external source that was actually checked against it, particularly on the rare cases where checking is easy to skip.