10‑second diabetes voice test
Researchers and technologists are touting AI that can flag type‑2 diabetes from a 10‑second voice recording by analyzing speech patterns — the idea: noninvasive screening without a blood draw. The claim has sparked buzz about feasibility and research reproducibility in real‑world settings. (x.com)
A Mayo Clinic Proceedings: Digital Health paper led by Jaycee M. Kaufman analyzed 6–10 second smartphone recordings from 267 adults in India (collected Aug 30, 2021–Jun 30, 2022), producing 18,465 clips and 14 extracted acoustic features for model training. (doaj.org) The Klick-led analysis found sex-specific vocal markers—pitch, pitch SD and relative average perturbation (jitter) were strongest for women while intensity and apq11 shimmer stood out for men—and reported cross-validated prediction accuracies of about 0.75±0.22 for women and 0.70±0.10 for men in age‑ and BMI‑matched samples. (doaj.org) An independent, larger Colive Voice analysis published in PLOS Digital Health (Dec. 19, 2024) tested 607 U.S. participants with gender‑specific algorithms and reported AUCs of 75% for men and 71% for women, correctly identifying 71% of male and 66% of female T2D cases and showing better performance in older women and people with hypertension. (journals.plos.org) The PLOS team used hybrid BYOL‑S/CvT audio embeddings and cross‑validation, and the Colive protocol is registered on ClinicalTrials.gov (NCT04848623), signalling an effort toward pre‑specified methods and reproducible cohorts. (journals.plos.org) Authors and commentators have flagged limitations explicitly: both papers call for external validation in earlier‑stage disease, more diverse languages and recording environments, and larger multicenter cohorts before clinical deployment. (journals.plos.org) The Colive consortium and the Luxembourg Institute of Health describe an ongoing international, multilingual voice database project intended to expand sample diversity and enable the multi‑center validations the papers say are still needed. (lih.lu)
