Smartphone-based speech analysis app may predict worsening heart failure
Key takeaways:
- The novel speech analysis technology can predict when the condition of a patient with HF is about to worsen.
- If successfully developed, the technology could help keep patients with HF out of the hospital.
PHILADELPHIA — A novel smartphone-based speech analysis technology predicted worsening heart failure that could lead to hospitalization, according to results of a study presented at the American Heart Association Scientific Sessions.
“Clinicians can recognize changes in patients’ voices or speech patterns at the time they are admitted to the hospital with decompensated heart failure. You can hear the breathlessness, the effort they are making to breathe,” William T. Abraham, MD, FACP, FACC, FAHA, FESC, FRCPE, professor of medicine, physiology and cell biology and a College of Medicine Distinguished Professor in the division of cardiovascular medicine at The Ohio State University Wexner Medical Center, told Healio. “The notion came up that perhaps there is a way to detect those changes early, much earlier than presentation to the hospital. Then maybe we ultimately could use that data to keep patients well and out of the hospital. We thought the solution might be in speech processing and machine learning and artificial intelligence. And that led down this road to develop a very simple, smartphone-based application that listens to speech and can look for changes in things like pitch or tone or speech dynamics that are indicative of accumulating fluid in the lungs and in the upper airwaves.”
Image: Adobe Stock
For the Cordio HearO Community Study, Abraham and colleagues gathered 158,024 speech recordings using the app (HearO, Cordio Medical) by 263 outpatients with congestive HF (mean age, 69 years; 25% women) for the development group and 94,202 speech recordings by 153 outpatients with congestive HF (mean age, 66 years; 24% women) for the test group. Patients were asked to speak five prespecified sentences into the app each morning.
The rate of adherence to daily recordings was 83% in the development group and 81% in the test group, Abraham said during a press conference, noting that 75% of the development group and 70% of the test group had at least 70% recording compliance.
In the development group, the sensitivity for all HF events was 76% and the sensitivity for first HF events was 81%, Abraham said, noting that detecting of impending HF events occurred an average of 24 days before the event.
The false-positive rate was 3.15 per patient per year (95% CI, 3.01-3.31), or approximately one every 3.8 months, he said.
In the test group, the sensitivity for all HF events was 71% and the sensitivity for first HF events was 77%, Abraham said, noting that detecting of impending HF events occurred an average of 26 days before the event.
The false-positive rate was 2.67 per patient per year (95% CI, 2.48-2.89), or approximately one every 3 months, he said.
“We ask our patients to weigh themselves every day and to tell us when their weight changes by a couple or three pounds overnight or a few pounds over a few days,” Abraham told Healio. “The sensitivity for weight change for predicting a heart failure hospitalization is only about 20%. The sensitivity for this HearO application and speech processing system approaches 80%. So it’s got a much higher sensitivity [and] a lower false-positive rate ... than the things that are currently considered standard of care for monitoring heart failure patients.
“The next steps include an ongoing U.S. pivotal trial to extend the observations in this current study performed outside the U.S. and lead to FDA regulatory approval, so we can begin to put this tool in the hands of clinicians and patients,” Abraham told Healio. “And then subsequently to do additional studies to help us understand just how large the value proposition here is. And by that, what I mean is how well can we use this technology to keep patients feeling well and to keep them out of the hospital.”
Perspective
Back to Top
The SPEECH trial is provocative. The technology works. Voice recognition and a change in speech pattern may indeed serve as a harbinger of impending decompensation in those with heart failure. That may already qualify as a “win”. Yet, there are several situations that might either allow this system to soar or sink. One is cost. If this system is truly inexpensive, something like the cost of an app, it beats everything else, including implantable pulmonary artery monitors and impedance monitors. The question is will it come forward at a very reasonable cost or not.
Another is a different kind of diversity. Is this technology language agnostic? What happens when the system is trained on patients with languages that are non-English or with varying dialects within a given primary language? How do we know that five sentences per day is enough to train the system for someone who speaks Creole, or one of the many Asian or African languages? Even though there are big language buckets, within the languages there are multiple dialects, and five sentences may not be sufficient. The training time may need to be a lot longer than what was done in this study, or not as much. This will require more work in language discordant geographies.
Telemedicine/telehealth and remote management are here to stay. We need more tools. Right now, we must rely on whatever adroitness and cleverness we have when evaluating a patient by phone or video; determining sick or well is not always clear. We need something else besides a subjective “how do you feel?” The idea that when we are talking to someone, we can understand how they feel and interrogate voice patterns is a potential tool in a toolbox that is essentially empty. We need at least one really good tool. Is this it? Let’s see how the data evolve.
Collapse
Abraham WT, et al. LBS.08. Artificial intelligence at the bedside. Presented at: American Heart Association Scientific Sessions; Nov. 11-13, 2023; Philadelphia.