Skip to main content
 
Notice: Users may be experiencing issues with displaying some pages on stanfordhealthcare.org. We are working closely with our technical teams to resolve the issue as quickly as possible. Thank you for your patience.
Stanford Health Care

Voiceprints of cognitive impairment: analyzing digital voice for early detection of Alzheimer's and related dementias.

Voiceprints of cognitive impairment: analyzing digital voice for early detection of Alzheimer's and related dementias. NPJ dementia Rezaii, N., Wong, B., Aisen, P., Beckett, L., Dage, J. L., Eloyan, A., Foroud, T., Womack, K., Carrillo, M. C., Kramer, J. H., Jack, C. R., Koeppe, R., Kukull, W. A., Nudelman, K., Polsinelli, A. J., Raman, R., Rumbaugh, M., Taurone, A., Thangarajah, M., Toga, A., Vemuri, P., Atri, A., Clark, D., Day, G. S., Duara, R., Graff-Radford, N. R., Grant, I., Honig, L. S., Johnson, E. C., Jones, D. T., Masdeu, J. C., Mendez, M. F., Musiek, E., Onyike, C. U., Riddle, M., Rogalski, E., Salloway, S., Sha, S., Scott Turner, R., Wingo, T. S., Wolk, D. A., Hammers, D., Kirby, K., Rabinovici, G. D., Apostolova, L. G., Dickerson, B. C., LEADS Consortium, La Joie, R., Blazhenets, G., Touroutoglou, A., Ghetti, B., Grinberg, L. T., Murray, M. E., Newell, K. 2025; 1 (1): 35

Abstract

Early detection of Alzheimer's disease (AD) is critical yet challenging, particularly in younger individuals. This study leverages artificial intelligence to analyze digital voice recordings from the Craft Story Recall task within the Longitudinal Early-onset AD Study (LEADS) to (1) detect cognitive impairment and (2) differentiate early-onset AD (EOAD) from early onset non-AD cognitive impairment (EOnonAD). Using speech samples from 120 patients and 68 cognitively unimpaired controls, we employed two classification approaches: feature-engineered machine learning and end-to-end deep learning incorporating a Large Language Model. To detect mild cognitive impairment, the feature-engineered model, using acoustic and linguistic features, achieved an AUC of 0.945 on the holdout test set, while the end-to-end model yielded an AUC of 0.988. For differentiating EOAD from EOnonAD, the feature-engineered model achieved an AUC of 0.804, and the end-to-end model yielded an AUC of 0.904 on the holdout set. Explainability analyses revealed reduced linguistic informativeness as a key AD indicator.

View details for DOI 10.1038/s44400-025-00040-0

View details for PubMedID 41195302