Voice Signal Analysis Can Help Detect Highest Risk Patients with Congestive HF
Voice signal analysis was done on a large group of patients with chronic health conditions registered to a telemedicine call center at Sheba Medical Center. The use of this non-invasive biomarker may help identify patients who are at an increased risk for extended hospitalization or mortality.
According to research published in the Journal of the American Heart Association, voice signal analysis is a significant emerging biomarker that may help to detect which patients with congestive HF are at the highest risk for mortality or 20-month hospitalization.
The research study, a collaborative effort between Sheba Medical Center and Mayo Clinic, was performed on a group of 10,583 subjects with chronic health conditions, including congestive heart failure, who were registered to a telemedicine call center at Sheba. With the aid of voice processing techniques developed by Vocalis Health, researchers analyzed 20 seconds of audio from each patient and derived 223 acoustic characteristics. Then, based on a training cohort of 8,316 patients without congestive HF, they developed a vocal biomarker and evaluated its use in a cohort of 2,267 patients with congestive HF, who were classified into biomarker quartiles (data markers that are divided into four equal groups); the mean age of the patients with congestive HF was 77, and 63% of them were men.
“This is a very straightforward, noninvasive, transparent tool,” explained Prof. Elad Maor, MD, PhD, associate professor of cardiology at Sheba Medical Center and Tel Aviv University, and also a former fellow at Mayo Clinic. “Patients record themselves all the time. They make phone calls; they have telemedicine visits with clinics, especially now with the COVID-19 pandemic; patients record their voice when they use Siri, Alexa or Google to do searches.”
Nurses at Sheba’s telehealth center collected the 20-second voice recordings from periodic phone conversations with patients. Prof. Maor and other members of the research team described how they used the voice processor to extract low-level acoustic features from the conversation samples via a resolution of 100 points per second and forming a matrix with 2000 columns. The features included pitch and formant measures, mel-frequency cepstrum representation, loudness, jitter, and shimmer. Then, further high-level characteristics were extracted from the matrix, leading to 223 valid features for each voice sample. Finally, the training cohort was used to build a machine learning linear model that was blinded to the cohort of congestive HF patients.
The primary outcome of the study was all-cause mortality, which was found to occur in 36% of the patients over an average follow-up span of 20 months. According to the Kaplan-Meier survival analysis, the cumulative probability of death was greater with increasing quartiles – ranging from 23% in quartile one and up to 54% in quartile four. Additionally, each standard deviation increase in the vocal biomarker was associated with a 32% greater risk of death during follow-up.
In the group of patients with congestive HF, each standard deviation in the vocal biomarker demonstrated close to a 25% greater risk for hospitalization during follow-up.