Research using artificial intelligence to analyze voice features such as speaking speed, pauses and pitch to flag Alzheimer's risk is gaining momentum. In a small preliminary study, a voice-based model sorted participants into risk groups with up to 77 percent accuracy, well above the 47 percent achieved by a conventional cognitive-test model.
According to the Korea Economic Daily, which on the 16th cited a New York Times report from Nov. 11, a team led by Dr. Marziye Eshghi at Mass General Brigham's Institute of Health Professions in Boston is analyzing older adults' voices and speech-related muscle movements to classify Alzheimer's risk.
광고 문의 · 300×250The preliminary study involved 49 people without cognitive impairment. Researchers divided them into low, medium and high-risk groups based on blood levels of Alzheimer's-related proteins and genetic risk variants, then analyzed voice recordings without knowing the groupings to test how accurately the AI sorted them.
Some 130 measures, from speed to pitch
The analysis used about 130 measures, including speaking speed, pauses between words, pitch and articulatory muscle movement. The voice model reached up to 77 percent classification accuracy, compared with 47 percent for the cognitive-test model. The figure reflects how well the model sorted risk groups, however, not a direct prediction of whether an individual will develop dementia.
Researchers believe subtle differences that people cannot hear or see may appear years before cognitive decline. Dr. Maria Luisa Gorno-Tempini of the University of California, San Francisco, called voice and language an important window into brain health.
Earlier research by Boston University professor Rhoda Au found that AI analysis of voices recorded during neuropsychological testing could help predict whether patients with mild cognitive impairment would progress to Alzheimer's within six years. Au noted that speaking is a cognitively complex task that nearly everyone performs.
Large three-country validation under way
The nonprofit Alzheimer's Drug Discovery Foundation is running the SpeechDX study, collecting three years of English, Spanish and Catalan voice data from about 2,000 people in the United States, Australia and Spain. It is also gathering blood tests, brain scans and cognitive assessments, aiming to identify which people with positive blood biomarkers will actually decline.
Voice analysis is cheaper and less physically burdensome than brain imaging or spinal fluid tests, and recordings can be stored and reanalyzed with new techniques. With the preliminary study limited to 49 people, however, caution remains that large-scale validation is needed before it can serve as a screening tool in super-aged societies such as South Korea.