September 19, 2022 — Imagine this: you think you might have COVID. You speak a few sentences into your phone. Then an app gives you reliable results in less than a minute.
“You look sick” is what we humans might say to a friend. Artificial intelligence, or AI, could take this to new frontiers by analyzing your voice to detect COVID infection.
An inexpensive and simple app could be used in low-income countries or to screen crowds at concerts and other large gatherings, researchers say.
It’s just the latest example of a growing trend to explore the voice as a diagnostic tool to detect or predict disease.
Over the past decade, AI speech analysis has been shown to help detect Parkinson’s disease, post-traumatic stress disorder, dementia, and heart disease. The research has been so promising that the National Institutes of Health has just launched a new initiative to develop AI to use voice to diagnose a wide range of conditions. These range from respiratory diseases such as pneumonia and COPD to cancer of the larynx and even stroke, ALS and psychiatric disorders such as depression and schizophrenia. Software can detect nuances that the human ear cannot, researchers say.
At least half a dozen studies have taken this approach for COVID detection. In the most recent breakthrough, researchers at Maastricht University in the Netherlands report that their AI model was accurate 89% of the time, compared to an average of 56% for various lateral flow tests. The voice test was also more accurate at detecting infection in people without symptoms.
One catch: Lateral flow tests show false positives less than 1% of the time, compared to 17% for the voice test. Still, since the test is “virtually free”, it would still be practical to ask those who test positive to take further tests, said researcher Wafaa Aljbawi, who presented the preliminary results at the International Congress of the European Respiratory Society in Barcelona, Spain.
“I am personally excited about the possible medical implications,” says Visara Urovi, PhD, researcher on the project and associate professor at the Institute of Data Science at Maastricht University. “If we better understand how the voice changes with different conditions, we could potentially know when we’re about to get sick or when to seek more testing and/or treatment.”
A COVID infection can change your voice. It affects the airways, “resulting in lack of speech energy and loss of voice due to shortness of breath and congestion in the upper airways,” the preprint document, which has not yet been evaluated by peers. The dry cough typical of a COVID patient also causes changes in the vocal cords. And previous research has found that dysfunction of the lungs and larynx due to COVID changes the acoustic characteristics of a voice.
Part of what makes the latest research remarkable is the size of the dataset. The researchers used a University of Cambridge crowdsourced database that contained 893 audio samples from 4,352 people, 308 of whom tested positive for COVID.
You can contribute to this database – all anonymous – through Cambridge’s COVID-19 Sounds app, which asks you to cough three times, take a deep breath through your mouth three to five times and read a short sentence three times .
For their study, researchers at Maastricht University “focused only on spoken sentences,” Urovi says. The “signal parameters” of audio “provide information about the energy of speech,” she says. “It is these numbers that are used in the algorithm to make a decision.”
Audiophiles may find it interesting that researchers have used mel spectrogram analysis to identify characteristics of the sound wave (or timbre). AI enthusiasts will note that the study found long-short-term memory (LSTM) to be the type of AI model that performed best. It is based on neural networks that mimic the human brain and is particularly good at modeling signals collected over time.
For laypersons, it’s enough to know that advances in the field can lead to “reliable, effective, affordable, convenient, and easy-to-use” technologies for disease detection and prediction, according to the paper.
Building this research into a meaningful application will require a successful validation phase, Urovi says. Such “external validation” – testing how well the model works with another set of sound data – can be a slow process.
“A validation phase can take years before the app can be made available to the general public,” says Urovi.
Urovi points out that even with Cambridge’s large data set, “it’s difficult to predict how well this model might work in the general population.” If speech tests turn out to work better than a rapid antigen test, “people might prefer the cheap, non-invasive option.”
“But further research is needed to determine which voice features are most useful for selecting COVID cases and to ensure models can differentiate between COVID and other respiratory conditions,” the paper says.
So, is pre-gig app testing in our future? This will depend on cost-benefit analyzes and many other considerations, Urovi says.
Nevertheless, “it can still provide benefits if the test is used in support or in addition to other well-established screening tools such as a PCR test”.