AI Voice Analysis App May Detect Worsening Heart Failure Early

Voice analysis with an app that uses artificial intelligence may provide early warning of acute decompensated heart failure (HF).

The proprietary HearO system and app (Cordio Medical) showed a sensitivity of about 70% for predicting impending acute decompensation about 3 weeks before it occurred in high-risk outpatients with HF in the HearO Community Study.

William T. Abraham, MD, Ohio State University Wexner Medical Center, Columbus, Ohio, reported the findings in a press briefing and a late-breaking trial session on November 13 at the American Heart Association (AHA) 2023 Scientific Sessions.

They show that "novel speech analysis technology may be useful in remote monitoring of HF patients, providing early warning of worsening HF events including impending decompensation," Abraham said. "This approach has the potential to reduce acute decompensated HF hospitalizations and improve patient quality of life and economic outcomes."

The assigned discussant for the trial, David T. Ouyang, MD, the Department of Cardiology and Division of Artificial Intelligence in Medicine, Cedars-Sinai Medical Center, Los Angeles, said he is "cautiously optimistic," but notes that "it is still quite early" research.

This trial was a multicenter, noninterventional, single-arm, unblinded study that enrolled 263 outpatients with HF into a "development" group, used to refine the development of an algorithm that predicts imminent worsening HF, and 153 similar patients in a "test" group that would validate the algorithm at six sites in Israel.

The patients spoke the same five short sentences — for example, "The puppy sat on the ship" — into their smartphone each morning. The system algorithm is being designed to detect voice changes that occur with lung congestion in decompensated HF; in the future, when these changes are detected, it will send an alert to the patient's clinician about imminent risk of worsening HF that requires hospitalization and/or intravenous therapy.

Abraham presented preliminary data from 80 patients in the development group at the Heart Failure Association of the European Society of Cardiology sessions, as previously reported by theheart.org | Medscape Cardiology, which followed an earlier small study.

Other studies have shown that the sensitivity for daily weight change to predict HF hospitalization is only about 10% or 20%, he said, and patients only recognize worsening signs and symptoms 2 or 3 days before HF hospitalization.

The current research aims to develop a tool with better sensitivity and specificity that can predict worsening HF events earlier when added to other clinical information, he said, "so that we have a broad window of opportunity to intervene and keep patients out of the hospital."

The current study is limited by a small number of patients and HF events, particularly in the test group, Abraham acknowledged.

The investigators are now conducting an observational trial that will enroll 250 patients at diverse sites in the United States and 250 patients in Israel, he said. This is an event-driven trial requiring a minimum of 78 HF events to perform the analysis, so it may enroll more than 500 patients.

"We expect complete enrollment of the trial early in 2024," Abraham said, "and are hopeful of having the events and the analysis available by the end of 2024 or early in 2025." The investigators then anticipate the potential regulatory approval from the US Food and Drug Administration (FDA) in about a year, followed by continued research to demonstrate the clinical and economic value of this system in studies including randomized trials.

Promising Technology but Still Early

"It's incredible technology if it works," Ouyang told theheart.org | Medscape Cardiology, given that HF is prevalent and there's a need to prevent people with HF from being readmitted to hospital.

The main limitation of this, and the upcoming study, is that it is not a prospective randomized trial, he noted. Rather, when the patient has an HF event, the researchers look back 31 days to see if, and when, the system detected worsening HF.

"Generally, for a lot of cardiologists, we think that the change [decompensation], while gradual, most acutely happened in the immediate 1 to 2 weeks," Ouyang added. "It would be helpful to find an earlier signal, but the signal could be quite subtle."

In addition, he pointed out, "I get a voice change when I get a cold or some other things that aren't just heart failure," such as spasmodic dysphonia or essential tremors or COVID. "There needs to be more study," he said, "on what other voice changes mean and whether this will create a false positive.

"This is promising technology but very early," Ouyang summarized. He added, "It's a long way off from FDA clearance or getting any kind of insurance to pay for it. To do that, they need proof that the early signals are either modifiable or can improve care."

For example, "AI might pick up readmissions more often, but is there actually a therapy associated with the diagnostic, where we can prove that we can actually avoid some of these outcomes?" he asked rhetorically. "That's still unproven at this point."

A Development Group and a Test Group

The current study enrolled adult outpatients with HF (New York Heart Association class II or III) with reduced or preserved ejection fraction who were at a risk for an HF event (due to prior HF hospitalizations or N-terminal pro–B-type natriuretic peptide levels) into a development group and a test group at six sites in Israel.

Patients in both groups were similar. They had a mean age of 67 years, and 25% were women. Most (57%-72%) patients spoke Hebrew, and the rest spoke Russian or Arabic or, rarely, English or Spanish.

They were instructed to record the same five short sentences on their smartphone each morning.

The patients in the development group made recordings on average 83% of days between March 27, 2018, and November 30, 2021, and were followed for up to 44 months.

The patients in the test group made recordings on average 81% of days between February 1, 2020, and April 30, 2023, and were followed for up to 31 months.

Most patients (70%-75%) made the five daily recordings 70% or more of the time.

In the development group, 58 of the 263 patients had an HF event. On average, the system had detected risk of worsening HF 24 days before the event, 76% of the time (76% sensitivity), with a false-positive rate of 3.15 per patient per year.

In the test group, 14 of 153 patients had an HF event. On average, the system had detected risk of worsening HF 26 days before the event, 71% of the time (71% sensitivity), with a false-positive rate of 2.67 per patient per year.

The study was supported by Cordio Medical, and Abraham has received fees from the company. Ouyang has no financial disclosures relevant to this study.

Credits:
Lead image: E+ / Getty Images

Medscape Medical News © 2023  WebMD, LLC

Send news tips to news@medscape.net.

Cite this: AI Voice Analysis App May Detect Worsening Heart Failure Early - Medscape - Nov 22, 2023.