Clinical Research

Detect the signs of dementia faster, more easily, and more objectively through speech with the help of artificial intelligence.

Work with us

Backed by clinical evidence

Linguistic features identify Alzheimer’s disease in narrative speech

Detecting Alzheimer’s disease

Analysis of spontaneous speech data elicited through Cookie Theft picture description (1-5 minute samples) from 240 people probable AD category and 233 healthy controls. Linguistic and acoustic variables used to train a machine learning classifier to distinguish between the groups, with 82% accuracy.

Fraser, K. C., Meltzer, J. A., & Rudzicz, F. (2016). Linguistic features identify Alzheimer’s disease in narrative speech. Journal of Alzheimer's Disease, 49(2), 407-422.
Using Linguistic Features Longitudinally to Predict Clinical Scores for Alzheimer's Disease and Related Dementias

Predicting MMSE for AD monitoring

A temporal Bayes network trained on 182 lexicosyntactic, 210 acoustic, and 85 semantic features extracted from 393 spontaneous speech samples elicited through Cookie Theft picture description can predict MMSE scores with a mean absolute error of 3.8, comparable to within-subject interrater (clinician) standard deviation of 3.9 to 4.8.

Yancheva, M., Fraser, K., & Rudzicz, F. (2015). Using Linguistic Features Longitudinally to Predict Clinical Scores for Alzheimer's Disease and Related Dementias.
Automated classification of primary progressive aphasia subtypes from narrative speech transcripts

Subtyping primary progressive aphasia

Syntactic and semantic features were automatically extracted from transcriptions of narrative speech for three groups: semantic dementia (SD), progressive nonfluent aphasia (PNFA), and healthy controls. Machine learning classifiers trained on these features were able to distinguish between the three participant groups with up to 100% accuracy.

Fraser, K.C., Meltzer, J.A., Graham, N.L., Leonard, C., Hirst, G., Black, S.E., & Rochon, E. (2014). Automated classification of primary progressive aphasia subtypes from narrative speech transcripts. Cortex, 55, 43-60.

Our technology

Drug development in Alzheimer's disease and dementia is hampered by our ability to identify at risk groups before the onset of clinically significant symptoms. Winterlight Labs is addressing this problem by pioneering a speech-based AI technology which could help accurately predict risk for dementia years before a clinical diagnosis is obtained. Our technology can help detect and monitor subtle changes in cognition by assessing individuals more frequently and more objectively than the assessments used today.

Unique benefits

Easy to administer

The assessment is fast and easy to administer: the patient only needs to describe a picture or recall words as instructed by an iPad. Traditional cognitive assessments like the MMSE and MoCA are stressful for the patient and can be undermined by learning effects. Our assessment however, is designed to be administered frequently (as often as monthly) giving a more granular picture of changes in cognition over time. As a result our assessment can be routinely completed to monitor cognitive health and track variables that might be impacted by treatment.

Automatic and objective scoring

The assessment can be completed objectively across patients and administrators. Standard cognitive assessments include components which are subjectively scored and often too coarse. We analyze hundreds of linguistic cues and can detect Alzheimer’s disease and other conditions with accuracies between 82% and 100%, as reported in our peer-reviewed academic studies.

Discover Insights

Our analysis captures both the acoustic and lexical properties of speech to produce a set of 475 variables. These variables are then grouped into aggregate metrics such word-finding difficulties, syntactic complexity, lexical richness, discourse mapping and more. They can be compared with an existing metric like an MMSE score.


Learning Multiview Embeddings for Assessing Dementia

Pou-Prom C., Rudzicz F. (2018)

Proceedings of Empirical Methods in Natural Languages Processing 2018.


A Swedish Cookie-Theft Corpus

Kokkinakis, D., Lundholm Fors, K., Fraser, K.C., and Nordlund, A. (2018)

Proceedings of the 11th edition of the Language Resources and Evaluation Conference (LREC). Miyazaki, Japan.


On the Importance of Normative Data in Speech-based Assessment

Noorian Z., Pou-Prom C., Rudzicz F (2017)

ML4H, Machine Learning for Health Workshop at NIPS 2017.


Identifying and Avoiding Confusion in Dialogue with People with Alzheimer’s Disease

Chinaei H, Danks A, Mehta T, Chan-Currie L, Lin H, Rudzicz F (2017)

Computational Linguistics, 43(2):377-406.


Rhetorical Structure and Alzheimer’s Aisease

Abdalla M., Rudzicz F., Hirst G. (2017)

Aphasiology, 32(1):41-60.


Speech Recognition in Alzheimer’s Disease and in its Assessment

Zhou L, Fraser K, Rudzicz F (2016)

Proceedings of Interspeech 2016.


Toward Dementia Diagnosis via Artificial Intelligence

Rudzicz F. (2016)

Today’s Geriatric Medicine, 9(2):8, March.


Vector-space Topic Models for Detecting Alzheimer's Disease

Yancheva, M., Rudzicz, F. (2016)

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).


Get in touch

Mailing Address

Winterlight Labs Inc.
Attn: Liam Kaufman
JLABS @ Toronto,
MaRS DD, West Tower
661 University Avenue
Suite 1300
Toronto, ON, M5G 0B7

Send us a message