What if AI could predict how long you would live just by listening to your voice? Perhaps it could hear the tremor in your speech, and foresee that you may be weakening and approaching the end of your life. What might your doctor do with all this information?
Population estimates show that, by 2060, 23% of all U.S. residents will be aged 65 and older, while 4.7% will be 85 and older. The cost of meeting their healthcare needs has driven researchers and clinicians across private and public sectors to find ways to better assess the pace of aging in older adults.
To help address this issue, Stony Brook University’s Ph.D. students Yunting Yin and Douglas William Hanes, along with Steven Skiena, Director of Stony Brook’s AI Institute, and Sean A. P. Clouston, Professor of Family, Population, and Preventive Medicine at the university, recently published a study, “Quantifying Healthy Aging in Older Veterans Using Computational Audio Analysis,” in the Journal of Gerontology.
The idea behind the project was based on the human ability to make inferences about a person’s age by listening to their speech. We do this unconsciously by not only noticing linguistic but also paralinguistic (non-verbal) factors. Like changes in voice timbre, speaking volume, and wavering while talking. However, research in the study of age-related vocal dysfunction, termed “presbyphonia,” suggests that these human predictions are off (on average) by 10 years, and even more so when the speaker is over 65 years of age.
To study this issue, Stony Brook researchers analyzed interviews provided by over two thousand male U.S. World War II veterans. They used publicly available information* about 60% of these veterans, as well as their voice recordings, to train the AI, and the other 40% to validate the predictions made by the model.
The results of the study were surprising, suggesting that the computer could reliably estimate a person’s biological age within 3 years, representing a 72% reduction in human estimation error. It was also noticed that when the person was younger than their predicted biological age, they had a greater life expectancy (10% for every five years).
“There are several studies that have found correlations between human voice and aging. However, these studies are still nascent,” says Ph.D. student Yunting Yin. “We need to study the subject further so we can fully define the relationship between presbyphonic and pathological changes.”
Postdoctoral Research Associate Douglas William Hanes, also involved with this study, adds, “This study adds to a growing literature that is dedicated to supporting older individuals during a precarious time. We’re glad to see results suggesting that voice samples can effectively predict biological aging.”
AI Institute Director Steven Skiena said, “Following the shift to telemedicine after COVID-19, we have extended prior research efforts in this area by trying to resolve the need for these methods to help supplement everyday listeners’ and geriatricians’ clinical intuition at a distance. Future research will not only help us refine our results but also mark an advance for the study of biomarkers of aging, potentially serving as a non-invasive—and hence less costly—method in measuring biological age.”
*Read about the Data Set Preparation Method here.
Ankita Nagpal
Communications Assistant