Gia is built on peer-reviewed speech-biomarker research. Scienza's foundation model for aging was then validated on a sealed test it had never seen.
Updated November 2026
David Kaiser; Eric D. Goldstein, MD (Henry Ford Health)
Submitted to JMIR AI, October 2026 · Manuscript 115801 · Public preprint coming soon
See the resultsTiming, pitch, articulation, word choice. Published studies report that these features of ordinary speech track cognitive and neurological change.
Kiyoshige and colleagues (2025), in 1,461 community-dwelling adults, reported AUC 0.89 (95% CI 0.86 to 0.92) for a model combining voice features with age, sex and education. The Lancet Regional Health, Western Pacific.
Brueckner and colleagues (2025) reported that vocal biomarkers from unconstrained conversational speech separate Parkinson's disease from controls. EMBS-BHI 2025 proceedings.
A 2025 scoping review in Biology maps how voice production engages the autonomic nervous system, giving a physiological basis for vocal markers of stress and cognitive load.
A 2026 summary in Frontiers in Digital Health reports the field's view that voice AI has reached readiness for clinical implementation.
These are the papers' results in their own populations. They are not Gia's results. Gia's own performance is reported below.
Scienza's foundation model for aging reads a person's longitudinal record and estimates risk of defined outcomes over defined horizons. We tested it the way a trial is run.
Sealed-test results, as reported in the manuscript
The rules were locked first.
Five outcomes, horizons, eligible populations, metrics, pass criteria, seeds and analysis code were fixed by hash before the sealed partition was opened.
A sealed partition of 468,756 residents.
Split by person. None contributed to training or calibration. Model weights were frozen before the evaluation.
The sealed job ran exactly once.
A separately registered second look scored comparators only against Gia's unchanged predictions. It did not reopen the original verdicts.
Every claim reported, met or not.
All five outcomes met their discrimination criterion. Four met calibration. Parkinson's calibration was indeterminate, as pre-classified for an underpowered outcome.
How to read these results
AUROC with 95% confidence intervals, resident-level bootstrap. Bars run from 0.5 (chance) to 1.0 (perfect ranking). Calibration held on four outcomes (slopes 0.96 to 1.01). Falls at one year, AUROC 0.848 (0.847 to 0.849), is a non-confirmatory replication: earlier falls results were known before registration. Hospital return is trained and is not part of the sealed-test claims.
Manuscript submitted to JMIR AI. The data come from a single source and a predominantly skilled-nursing population. Independent-population validation and prospective workflow evaluation are the stated next steps.
Titles are reproduced as published. Inclusion here is not a claim that a study evaluated Gia.
Kiyoshige, et al. (2025). Developing and testing AI-based voice biomarker models to detect cognitive impairment among community dwelling adults: a cross-sectional study in Japan. The Lancet Regional Health, Western Pacific. doi:10.1016/j.lanwpc.2025.101598
Brueckner, et al. (2025). Detecting Parkinson's Disease using Vocal Biomarkers based on Speech Foundation Models. EMBS-BHI 2025 (conference proceedings).
(2026). Translating AI Research into Reality: Summary of the 2025 Voice AI Symposium and Hackathon. Frontiers in Digital Health. doi:10.3389/fdgth.2026.1754426
(2025). Mapping the Neurophysiological Link Between Voice and Autonomic Function: A Scoping Review. Biology (MDPI). doi:10.3390/biology14101382
Kaiser D, Goldstein ED (2026). Pre-registered validation of a longitudinal sequence model for cognitive and neurological outcomes in senior-care residents with a post hoc comparator addendum. Submitted to JMIR AI.
Collins GS, Moons KGM, Dhiman P, et al. (2024). TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385:e078378.
Dwork C, Feldman V, Hardt M, Pitassi T, Reingold O, Roth A (2015). The reusable holdout: preserving validity in adaptive data analysis. Science 349(6248):636-638.
Mitchell AJ, Meader N, Pentzek M (2011). Clinical recognition of dementia and cognitive impairment in primary care: a meta-analysis of physician accuracy. Acta Psychiatrica Scandinavica. doi:10.1111/j.1600-0447.2011.01730.x
Mattke S, Jun H, Chen E, Liu Y, Becker A, Wallick C (2023). Expected and diagnosed rates of mild cognitive impairment and dementia in the U.S. Medicare population: observational analysis. Alzheimer's Research & Therapy. doi:10.1186/s13195-023-01272-z
Manly JJ, Jones RN, Langa KM, Ryan LH, Levine DA, McCammon R, Heeringa SG, Weir D (2022). Estimating the Prevalence of Dementia and Mild Cognitive Impairment in the US: The 2016 Health and Retirement Study Harmonized Cognitive Assessment Protocol Project. JAMA Neurology. doi:10.1001/jamaneurol.2022.3543
Schrag A, Ben-Shlomo Y, Quinn N (2002). How valid is the clinical diagnosis of Parkinson's disease in the community?. Journal of Neurology, Neurosurgery & Psychiatry. doi:10.1136/jnnp.73.5.529
Pringsheim T, Wiltshire K, Day L, Dykeman J, Steeves T, Jette N (2012). The incidence and prevalence of Huntington's disease: a systematic review and meta-analysis. Movement Disorders. doi:10.1002/mds.25075
Mitchell AJ, Vaze A, Rao S (2009). Clinical diagnosis of depression in primary care: a meta-analysis. The Lancet. doi:10.1016/S0140-6736(09)60879-5
Olariu E, Forero CG, Castro-Rodriguez JI, Rodrigo-Calvo MT, Alvarez P, Martin-Lopez LM (2015). Detection of Anxiety Disorders in Primary Care: A Meta-Analysis of Assisted and Unassisted Diagnoses. Depression and Anxiety. doi:10.1002/da.22360
Magruder KM, Frueh BC, Knapp RG, Davis L, Hamner MB, Martin RH, Gold PB, Arana GW (2005). Prevalence of Posttraumatic Stress Disorder in Veterans Affairs Primary Care Clinics. General Hospital Psychiatry. doi:10.1016/j.genhosppsych.2004.11.001
Hirschfeld RM, Lewis L, Vornik LA (2003). Perceptions and Impact of Bipolar Disorder: How Far Have We Really Come? Results of the National Depressive and Manic-Depressive Association 2000 Survey of Individuals with Bipolar Disorder. Journal of Clinical Psychiatry.
Powell JM, Ferraro JV, Dikmen SS, Temkin NR, Bell KR (2008). Accuracy of mild traumatic brain injury diagnosis. Archives of Physical Medicine and Rehabilitation. doi:10.1016/j.apmr.2007.12.035
Porter J, Boyd C, Skandari MR, Laiteerapong N (2023). Revisiting the Time Needed to Provide Adult Primary Care. Journal of General Internal Medicine. doi:10.1007/s11606-022-07707-x
The validation summary is available to evaluating organizations. Registration artifacts and full tables accompany the manuscript.
Published studies report that acoustic and linguistic features of ordinary speech carry signal about cognitive impairment, Parkinson's disease and other conditions. Kiyoshige and colleagues (2025) reported AUC 0.89 (95% CI 0.86 to 0.92) for a model combining voice features with age, sex and education in 1,461 community-dwelling adults. Those are the papers' results, not Gia's.
Scienza's foundation model for aging was evaluated on a sealed partition of 468,756 residents never used in training. Five outcomes, horizons and pass thresholds were registered before the partition was opened. All five met their discrimination criterion and four met calibration. Falls replicated at AUROC 0.848. The manuscript has been submitted to JMIR AI.
No. The references show that the science Gia is built on is real and published. Gia's own performance is reported separately, from the pre-registered sealed test, and every pre-specified claim is reported as met, not met or indeterminate.
No. Every output is an option for a clinician to review. Gia does not issue orders and nothing enters the record without a clinician's decision.
Gia is a Digital Human® with Voice AI. Every result is an option for a clinician to review, not a diagnosis. See also our dataset, governance and security and AI transparency.
