Can a Smartphone App Use Voice Patterns to Predict CAD Events?

Voice patterns detected by a smartphone app may help to predict subsequent CAD events, a small, single-center study suggests. Overall, patients with the highest values of this biomarker had an increased risk of experiencing ACS, presenting to the emergency department with chest pain, or being admitted to the hospital with chest pain.

In the same patients, all of whom had been referred for clinically indicated angiography at the Mayo Clinic, researchers previously showed a link between voice characteristics on the app, developed by the company Vocalis Health, and the presence of CAD. And other studies using the same app drew a link between voice and indices of pulmonary hypertension among patients undergoing cardiac catheterization, as well as death and hospitalization among patients with congestive heart failure.

The new results, published in Mayo Clinic Proceedings, are set to be presented at the upcoming American College of Cardiology (ACC) 2022 Scientific Sessions in Washington, DC.

Lead investigator Jaskanwal Deep Singh Sara, MBChB (Mayo Clinic, Rochester, MN), who spoke with reporters at an ACC media briefing earlier this week, made the case for the technology’s role in improving risk prediction. According to an ACC press release, Vocalis Health based the app’s algorithm on a “training set of over 10,000 voice samples collected in Israel.”

“Identifying at-risk groups forms the cornerstone of successful preventative strategies,” Sara said. “However, traditional multivariable risk-prediction scores have important shortfalls, and nonconventional risk measures”—such as calcium scoring and measures of endothelial function—“often require in-person, lengthy, and costly evaluations. So simple, inexpensive, and noninvasive methods to identify people at risk would be of great value.”

Also welcome in the COVID-19 era are tools that could be used without face-to-face clinical assessments, to reduce the spread of disease and put less the strain on overburdened healthcare systems, the researchers point out in their paper.

William T. Abraham, MD (The Ohio State University, Columbus), recently led a study testing a different app with a different purpose: monitoring “wet” versus “dry” voice patterns among heart failure patients, in the hopes of providing an early alert that they’ve begun to develop pulmonary edema.

He told TCTMD he’s “struggling to understand the mechanism here,” where the Vocalis Health app is being used as a screening tool among a broader swath of patients. Two explanations proposed by the authors, which relate to the inflammatory hypothesis and cardiac autonomic regulation, are “plausible,” said Abraham. “But right now I think they would have to do a lot more mechanistic work to understand [this] better.”

Yet Abraham said he was still glad to see the paper by Sara et al. “I do think that it shows all of the energy that’s in this space of speech analysis to understand cardiovascular disease,” he commented.

More Than Double the Risk

For their study, 108 patients (mean age 59.5 years; 54.6% men) treated at the Mayo Clinic who were slated for angiography. Participants were asked to record three 30-second voice samples using the Vocalis Health app: one simply by reading a text and the two others by describing a positive and a negative emotional experience. For each recording, the Vocalis Health algorithm was used to analyze features related to voice intensity and volume. The researchers then calculated the mean biomarker value for each patient.

Around half of these individuals (56.5%) presented with stable angina, while 14.8% had ACS and 28.7% were being evaluated preoperatively.

Median follow-up time was 24 months (range 1 to 60 months), during which 39.8% of patients had a primary outcome event: developing ACS, presenting to the ED with chest pain, and/or being admitted to the hospital for chest pain.

Patients in the highest biomarker tertile had a greater risk of experiencing the primary endpoint than those in the lower two tertiles (58.3% vs 30.6%; P = 0.006). Adjusted for CAD grade on baseline angiography, the relationship between voice and outcome remained statistically significant (HR 2.61; 95% CI 1.42-4.80). The study’s secondary outcome, either a positive stress test result or the presence of CAD on coronary angiography at follow-up, also was increased among those with the highest biomarker value, though the difference wasn’t significant after adjustment for baseline CAD grade (HR 3.13; 95% CI 1.13-8.68).

Yet there was no link between conventional risk factors (eg, age, sex, comorbidities, and smoking status) and CAD events.

“These results may have important clinical implications for telemedicine and the remote and noninvasive screening of patients to identify those at risk of coronary disease and its complications,” Sara said in the virtual press conference.

But What’s the Mechanism?

As to what’s behind the various voice patterns, Sara and colleagues suggest it could relate to the “systemic nature of atherosclerosis and inflammation,” with effects on blood vessels both in the heart and in organs related to speech. Another possibility relates to the vagus nerve, which is involved in cardiac autonomic regulation as well as voice production.

Moreover, post hoc analyses showed that individual patients’ biomarker levels held steady over time. “This finding suggests that CAD or aberrations in the vascular biology of the coronary bed may not be the sole mechanism accounting for the observed associations seen in this study,” the researchers propose.

Could this mean, as the paper hints, that the voice patterns linked to CAD are simply a signifier of worse overall health?

To TCTMD, Sara acknowledged that this could be possible. “It may relate to a sicker patient, a more-unwell substrate so to speak, and it begs the question: what is it that we’re actually picking up? Is it coronary disease per se? Or is it a milieu of inflammation, of altered autonomic nervous system reactivity, or other things as well? It could be a combination, but we need to do some more studies to try to tease that apart” in larger, more-diverse populations, he said.

Katie Berlacher, MD (University of Pittsburgh, PA), who moderated the discussion, described the report by Sara et al as “thought-provoking.”

Over the past 2 years of the COVID-19 pandemic, clinicians have had to rely on telemedicine like never before, so these results are welcome, Berlacher said. “I think for many of us in the field [who have] seen a number of patients through screens, we’ve wondered about ways we can assess people without having them come into the office or the emergency room,” she noted, adding that it will be exciting to see the themes of this small but creative study explored in larger populations.

For Abraham, there are some caveats to the current report. “Perhaps this paper, for this particular app and technology, raises more questions than maybe it provides answers in terms of mechanism and what they really were evaluating with their endpoint,” he said. For instance, it’s not clear whether the “chest pain” aspect of the endpoint refers exclusively to angina rather than possibly including noncardiac causes, Abraham pointed out.

The fact that voice-signal analysis has been used in so many different conditions, including Parkinson’s disease, means “we really have to work out the diagnostic utility—the sensitivity, the specificity—for each of these disease states,” he stressed.