The way someone speaks can tell a lot about them. A voice can depict someone’s origins and serve as a window to their emotion. Now, researchers at the University of Wisconsin–Madison are using speech to pick up on an under-diagnosed genetic condition called fragile X (FX) premutation.

FX premutation is a genetic abnormality that causes neurodegenerative disorders like infertility, and can lead to having children with FX syndrome (FXS). “FXS is the most common inherited cause of intellectual disability, and leading known genetic cause of autism spectrum disorder,” said Arezoo Movaghar, PhD student of biomedical engineering and co-author of the study.

The research team includes Krishanu Saha, senior author of the study and assistant professor of biomedical engineering at UW; Marsha Mailick, professor of social work and UW’s vice chancellor for research and graduate education; Jan Greenberg, professor of social work and associate vice chancellor for research and graduate education; and Audra Sterling, an assistant professor of communication sciences and disorders.

Full mutation of FXS is rare, but millions of people carry FX premutation, and most are unaware of their condition. Up to this point, the phenotypes (physical expression of a gene, or trait) of FX premutation had not been explored.

“Early detection has significant potential for improving public health and has an impact on family planning,” said Movaghar. “Characterizing the genotype-phenotype in this condition is essential to develop prognosis and diagnosis methods for FX-related conditions.”

In one of the studies, researchers found a correlation between age and difficulty speaking in women that carried FX premutation. “Considering the fact that language impairment is a hallmark of other neurocognitive disorders, including Parkinson’s disease, and Alzheimer’s disease, we sought to use cognitive and linguistic phenotypes to identify FX premutation carriers,” explained Movaghar. “In a novel approach based on machine learning, we have developed a computational framework using language samples and cognitive assessments to rapidly pre-screen the population for FX premutation status.”

Researchers compared 100 five-minute voice recordings from mothers with FX premutation talking about their FXS-positive children, with an additional 100 recordings of mothers talking about their children with autism spectrum disorder. They used these recordings to create an algorithm to differentiate mothers with FX premutation and those without.

The researchers used transcripts of the recording and the machine learning algorithms to create a list of linguistic features, like the average sentence length and the amount of pauses in a sentence, to quickly pre-screen for FX premutation. “Since this framework does not rely on genetic data, it is simple, fast and cost effective,” said Movaghar.

In fact, the machine learning algorithms correctly distinguished women with FX premutation from those without with 81 percent efficiency.

Computational approaches can revolutionize phenotype discovery, preventive care and diagnosis of neurodegenerative disorders, and have a significant effect on public health and family planning.

Current studies are limited to FX premutation-positive women with FXS-positive children. But, soon, research will include a more diverse crowd. “Future research is needed where the framework is tested with larger number of participants with more diversity in race, age, education and geographic location,” said Movaghar. The researchers plan to expand their testing sample to men with FX premutation, FX premutation carriers whose children do not have FXS or other disabilities, and a comparison group from the general population of people at risk for neurocognitive disorders.

The researchers note that machine learning algorithms do not have to be limited to testing for FX premutation. They can be expanded to test for other disorders, as well.

Movaghar is working to create a mobile app that would make testing simple and personal. Users could test themselves for FX premutation by answering questions posed by the app and recording a five-minute voice sample.