Science2 publishers2 min readPublished
Speech clock trained on 2,928 Spanish speakers rates cognitively impaired people as older
Agustín Ibáñez and colleagues built an AI speech clock from 2,928 Spanish speakers that rates people with cognitive impairment as older than their age. Its evidence so far compares diagnosed patients with healthy volunteers, so its promise as a cheap dementia screen depends on tests that have not been run yet.
The Scientist · Science desk
What happened
- The clock's output is a speech age gap, the difference between the age it predicts from someone's voice and that person's chronological age.
- Machine-learning tools pulled more than 700 features, including pitch and vocabulary range, from recorded speech tasks to train the model to predict age.
- Large speech age gaps were strongly associated with cognitive problems such as those that arise in dementia, according to Nature's report.
- The results were published in Science Advances, and Adolfo Ibanez University neuroscientist Agustin Ibanez is one of the co-authors.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- capability Clinics in low-resource regions could get an ageing measure without brain scans or blood tests, the use Jed Meltzer of the University of Toronto sees for it.
- decision Every participant spoke Spanish, so using the clock on English or other speakers means retraining and retesting it, since features such as vocabulary range are specific to a language.
- precedent With speech now a third basis for ageing clocks beside neuroimaging and DNA methylation, the next useful study would run all three on the same people and weigh cost against accuracy.
Ageing clocks already exist for brains and for DNA. Brain clocks read signatures from neuroimaging and epigenetic clocks read patterns of methyl tags on DNA, but until this study nobody had built one from speech, according to Nature's report [7]. Ibáñez's reason for trying is that speaking involves a "huge amount of brain work", as he put it [8].
The healthy participants are the control, and they came out about where they should: their speech age generally matched their chronological age [12]. "We can see a huge predictive value, just with a very simple four minutes of speech recordings," Ibáñez said [4].
The denominator is 2,928 people recruited in Argentina, Chile, Colombia, Mexico and Peru [9]. Against more than 700 speech features [11], that is at most about four participants per feature [1]. With a ratio like that, the first figure I would want is accuracy on people held out of training. The report does not give it, nor the split between healthy and impaired participants, nor the size of the gap in each group.
Separating people with a known diagnosis from healthy volunteers [10] is the first test a screening tool has to pass, and the easiest one. In people who are already impaired, the gap is a correlate of that impairment. A screen has to find the healthy-seeming person who will decline among many who will not. When most of the people tested are well, even a small false-positive rate sends many of them for follow-up they did not need.
In my view the speech gap has earned a prospective trial. Jed Meltzer, a University of Toronto cognitive neuroscientist who specializes in language and was not involved in the work, went further. "It is a very impressive piece of work," he said [6].
What to watch
- A follow-up that records people while they are still healthy and checks whether a large speech age gap comes before a later diagnosis.
- Whether the gap is larger in dementia than in mild cognitive impairment; a gap that scales with severity could be used to track progression.