Combining artificial intelligence with the sound of someone’s voice may eventually help diagnose patients with potential heart failure or Parkinson's disease.
As AI fever grips healthcare, some providers and digital health companies are using the technology to analyze people's speech patterns so they can detect future heart attacks or better understand a patient's social needs. The concept is promising enough that the National Institutes of Health has budgeted $14 million to create a database of 30,000 voices by 2026 that could be used to train AI for the diagnosis of diseases.
Read more: WHO calls for greater regulation of AI in healthcare
At The Ohio State University Wexner Medical Center in Columbus, Ohio, a team of researchers studied how a remote monitoring system operated through an AI-based smartphone application could detect potential heart failure events in patients.
“The idea was to discern whether or not changes in a patient’s speech could be detected through an AI-based system way before a clinician could hear them,” said Dr. William T. Abraham, a professor of cardiovascular medicine at OSU Wexner and lead author of the study. “The doctor is not listening to the patient every day whereas this system is.”
In a study of 263 patients, OSU researchers found the app was 76% accurate in predicting patients who would have a heart failure event. The app made the detection on average 24 days before hospitalization. In a separate study of 153 patients, the app was 71% accurate in detecting heart failure events on average three weeks in advance. Other methods to predict heart failure, such as a patient’s weight gain, are only 10%-20% accurate, Abraham said.
For the study, patients recorded five sentences every day into the app in their native language. The app was designed to detect whether a patient’s speech changed time by evaluating pitch, volume, dynamics and other characteristics. Voice changes can indicate early increases of lung fluid, a sign of progressing heart failure.
Abraham sees potential for the technology beyond cardiovascular health.
“Think about all the other medical illnesses that might affect speech—respiratory illness, asthma, COPD, chronic pain management, depression, etc.,” Abraham said. “I think the potential for this is quite great.”
While the concept holds promise, experts such as Dr. Yael Bensoussan, director of the University of Southern Florida Health Voice Center, said the technology faces roadblocks in the form of patient privacy.
“Sometimes you talk to your friends about going to a specific restaurant and then you see an ad for that restaurant pop up,” said Bensoussan, who is leading the NIH consortium gathering 30,000 voices. “The problem is if I was going to tell you, your phone will listen to your voice all day, analyze the information and figure out a diagnosis, would you approve? Probably not.”
More voices needed for better data
Beyond privacy concerns, Bensoussan said no industry standards exist concerning how speech data is collected and validated by developers. She also said there needs to be more voices collected to reflect people from different regional and socioeconomic backgrounds. Through the NIH project, Bensoussan and the team are looking to capture at least 30,000 voices for the database. The project has accumulated 200 voices, she said.
Despite potential reservations, Bensoussan said she is excited about the potential areas of medicine where AI-enabled speech analysis could help screen for and potentially diagnose patients with a disease.
For Parkinson’s and Alzheimer’s disease, it also could be used by pharmaceutical companies to monitor how patients respond to treatment, Bensoussan said.
“If you trial a new drug in a patient with Parkinson's, instead of taking a blood draw every week to see how he's doing... you could just measure his voice every week and see if there is improvement or worsening condition based on his voice,” Bensoussan said.
Outside of clinical areas, some companies and systems are using AI-enabled speech analysis to improve access to care. Eskenazi Health, a safety-net health system based in Indianapolis, is using a platform from AI company Authenticx to analyze phone discussions between patients and its customer service call center. The platform listens to a call between a patient and the customer service representative and gathers insights on potential points of friction.
The health system implemented the technology during the early days of the COVID-19 pandemic when call center volumes were skyrocketing, said Rachelle Tardy, director of clinical outcomes and integration. Over time, leaders at the health system realized there were additional opportunities to use AI to improve access to care, she said.
“Healthcare is more than just really coming in for an encounter. There are social constraints that impact care for patients,” Tardy said. “We began to identify some of the barriers that exist for patients. There was a trend we started to notice around transportation, especially around patients who needed to reschedule an appointment.”
Insights from the platform led Eskenazi to changing its scheduling process to identify if patients had a transportation need or if there was another reason why they couldn’t make an appointment, Tardy said. Without AI, providers may not understand a patient's social determinants, she said.
“When you start to see a patient has a significant number of no-shows, the automatic assumption is ‘What’s wrong with you? Why didn’t you come to your appointment,’” Tardy said. "Not every patient has the fundamental resources that we take for granted."
