The US National Institutes of Health (NIH) has earmarked as much as $14 million in funding to support the training of AI software that can analyze patients’ voices to diagnose and study illness.
Twelve research institutions led by the University of South Florida (USF) will receive the money to varying degrees over four years.
Their goal will be to collect, in a privacy conscious way, a training database of people’s voices that can be used to train applications that doctors can use to potentially detect diseases and neurological disorders by examining a person’s speech.
The Voice as a Biomarker of Health project will focus on software that can pick up on these five types of diseases:
- Voice disorders: (laryngeal cancers, vocal fold paralysis, benign laryngeal lesions)
- Neurological and neurodegenerative disorders (Alzheimer’s, Parkinson’s, stroke, ALS)
- Mood and psychiatric disorders (depression, schizophrenia, bipolar disorders)
- Respiratory disorders (pneumonia, COPD)
- Pediatric voice and speech disorders (speech and language delays, autism)
“Our team chose the five categories of diseases based on existing work in voice AI that has been published over the last 20 years,” Yael Bensoussan, the project leader and assistant professor at USF’s Department of Otolaryngology, told The Register.
Recent progress in machine learning algorithms to analyze voice and speech data have shown how technologies can be used to assess physical and mental health. A study led by researchers at MIT, for example, connected jitters and tremors in speech to depression and anxiety.
Academics believe the results are promising enough that listening and processing the sound of speech or breathing using AI could provide a low-cost method to detect diseases and disorders at an earlier stage.
“Voice is one of the cheapest bio-markers to study,” Bensoussan told us.
“When you think of biomarkers such as genetic testing or imaging like MRIs or scans, they are all quite resource-intensive and can be invasive in a sense. CT scans cause radiation for patients, for example. Voice is the easiest biomarker to collect, does not cause any physical risk for patients, and can be collected in very low resource settings especially with modern technology.”
The NIH will give $3.8 million in the first year to the Voice as a Biomarker of Health initiative for participants to construct a large, diverse voice database that can be assessed alongside other data gathered from medical imaging and genomics. Speech data will be recorded from selected patients in clinical settings in a pilot study in the first year.
The database will be shared among researchers to train AI algorithms to recognize common features in the voices of patients diagnosed with specific diseases. To make sure the sensitive data is kept private and secure, the models will be trained using federated learning supported by Owkin, a startup focused on assisting biomedical research using machine learning software.
“Federated learning technology – a novel AI framework that allows machine learning models to be trained on data without the data ever leaving its source – will be deployed across multiple research centers by Owkin to demonstrate that cross-center AI research can be conducted while preserving the privacy and security of sensitive voice data,” a spokesperson representing the company, told El Reg.
More money, as much as $14 million, may be granted to the initiative with congressional approval.
Voice as a Biomarker of Health is part of a wider effort from the NIH to accelerate the adoption of AI in R&D in the hopes that new technologies will revamp US healthcare. The medical research org promised to invest as much as $130 million over four years to numerous projects aimed at creating flagship biomedical datasets, universal software tools, and resources to train healthcare researchers in AI. ®