Even before he trained as an otolaryngologist, researcher Anthony Law understood that all of us are natural voice scientists with the intuitive power to hear peoples’ voices and know something’s wrong. ]
“When I got COVID in 2021,” he says, “within the first five seconds of our telephone conversation my mom knew something was wrong. She could hear it in my voice.”
Prof. Anthony Law
“A gentleman came to my clinic a couple of months ago,” he says. “I knew just from hearing his rough, strained and gravelly voice that there’s a high probability that he has laryngeal cancer. The severity of these voice changes mean his cancer is likely in the advanced stages.”
Early-stage laryngeal cancers are easy to treat if primary care physicians can catch them in time. The problem is that those changes in the voice, called dysphonia, can be a symptom of many things besides cancer. Law calls it a needle in a haystack hunt for the untrained.
“Almost everyone that has laryngeal cancer has voice change,” Law says. “We make voice by closing our voice box. it’s almost like two hands clapping together. When it closes, it vibrates and makes a sound. When you have a cancer, oftentimes that closure is incomplete. That gives a very breathy rough strain and sound in a voice.”
Law divides time between treating patients and doing more technical research, using math and computation to help people he sees in the clinic. He earned both an MD and a PhD in biophysics and learned about the power of artificial intelligence (AI) from friends at Microsoft while doing his residency in the northwest.
Today, he uses an AI model called a deep neural network, which broadly mimics the architecture of the brain, to make his expertise in voice analysis available to primary care clinicians so they can use voice to diagnose cancer of the larynx as accurately as a specialist.
“It's a very easy sign for trained laryngologists to hear that someone has a mass in their larynx,” Law says. “But not for primary care doctors. They don't have that extra training to hear concerning voice versus non-con concerning voice and screen out people who have voice changes due to a cold versus voice changes due to cancer.”
Training the AI model to detect voice changes
Law’s current AI model is about 93 percent successful at identifying who has a mass in their larynx, which is considered a good proxy for the underlying cancer. But it took years of dedicated, tedious work to build the database of 15,000 voice recordings needed to train the AI model.
“We've been really mindful to make sure that we're being fair and equitable when we build our models,” he says. “If we’re really good at detecting laryngeal cancer in women, but not men, then we've set up bias. Or if we're good at detecting laryngeal cancer in Midwesterners without an accent, but really bad at detecting the same in Southerners, we’ve left out a population.”
Early tests worked better in the pristine environment of the lab than the noisy environment of clinics. Seeking to improve that, Law developed a phone app that doctors can use to easily analyze patient voice data, while also fitting seamlessly into the busy workflow of a medical clinic.
“The whole process of using the app takes about four or five minutes,” he says. “A minute or two for the consent. After that, there's some demographic data that helps us understand who the patient is. Then there's a series of 10 voice prompts. Right now, we're trying to get more diverse data. We're hoping in six to eight months to do a randomized trial, with certain clinics using it and certain clinics that aren't. We'll see if we can change referral times. When patients come to see us, if they've been screened by our app, do they have smaller tumors? Are we catching them earlier? And eventually, does that change survival for patients who have laryngeal cancer?”
Law says machine learning in medicine is at an inflection point and his experimental app is part of it. If the trial is successful, he believes the voice app could be expanded to analyze other health problems as well. “We can go one of two ways,” he says. “Either all of that power for this really cool tool can be locked up in the hands of a few. Or if it’s used right, it can take a lot of the expertise that we have in big universities and democratize it and increase access for patients. That's why we continue to be excited.”