Machine Learning and NLP Are Emerging as Key Tools for Diagnosing Alzheimer’s Disease


Winterlight Labs, a spinoff from the University of Toronto, is using natural language processing (NLP) and machine learning to identify people with Alzheimer’s disease (AD) and other forms of dementia based on patterns in their recorded speech.

Why There Is a Need

The World Health Organization estimates that 47.5 million people in the world have dementia, which is defined as a chronic or persistent disorder of the brain marked by lapses in memory, personality changes, and impaired reasoning. Alzheimer’s disease accounts for 60% to 70% of the cases of dementia. In developed countries, AD is one of the most costly of all diseases to treat.

Every year in the United States, 236,000 people are diagnosed with AD. There are an additional 100,000 healthy individuals over the age of 60 who seek AD screening. Current paper and pen-based screening tests for Alzheimer’s disease can require five visits, take 15 hours to complete, and cost over $8,000. Examination of brain tissue is still needed for a definitive diagnosis.

Alzheimer’s is a disease that usually starts slowly and gets worse over time. Approved Alzheimer’s medications can help temporarily with memory symptoms but there is no cure. Finding a cure for AD has attracted a significant amount of funding and there are currently more than 200 active clinical trials focused on dementia in North America.  In these clinical trials alone, Winterlight estimates that its technology could save pharmaceutical companies $32 million if the company’s artificial intelligence (AI) technology was used to screen patients in place of current paper-based approaches. As the population ages and new medications are developed for AD, the demand for screening is likely to increase substantially.

How it Works

Dementia is often linked to language, and AI tools such as NLP and machine learning can be used to identify cognitive impairment. Winterlight has developed a tablet-based prototype that asks people to describe what’s happening in an image, encouraging them to speak for anywhere from 1 to 5 minutes. The language sample recording is then transcribed using a speech recognition system. Over 400 features of the sample including acoustic, lexical, syntactic, and semantic measurements are extracted into various categories. A machine learning classifier is then used to analyze the results. Machine learning classifiers are algorithms that have been trained to organize data into categories.  In the case of AD, the classifiers have been trained to distinguish between a healthy individual and a person with the disease. In tests, the prototype currently has an 81% accuracy rate in classifying speech samples.


Besides cost savings, Winterlight’s technology has a high level of accuracy and the approach is applicable to not only screening for dementia but also early detection and monitoring the disease over time.  Currently, the software can only do screening on population-level data, but the company is looking into applying it on a personalized level, analyzing individuals’ changes in speech over time. Winterlight plans to conduct pilot studies involving seniors living in Toronto retirement homes this summer to gather more individual data.


In medical statistics, false positives and false negatives are errors that result in incorrect diagnoses. In a false positive outcome, a person without AD could be classified as having the disease.  In a false negative outcome, a person with AD could be classified as not having it. Since there are multiple forms of dementia, and some cognitive impairment is a natural consequence of aging, false positives seem especially likely. Examination of brain tissue is likely to remain the only way to definitively diagnose the disease. As long as there are no cures, the downside of a false diagnosis seems harmless.  From a business model perspective, the ability to diagnose a disease but not prescribe a cure limits the market for the product.

The Path Forward

As treatments for AD are developed, approved, and promoted, the need to screen patients will grow according. Since systems such as Winterlight’s technology are based in part on machine learning, their results are very likely to improve over time, yielding more accurate results and overcoming many of the current challenges related to false diagnoses.

Comments are closed.