The competitive din in the vocal sound diagnostic space rose to new levels as Sonde Health Inc. launched a developer portal to enable other companies to incorporate its vocal biomarker-based health check technology into their apps. Sonde joins a growing number of companies and academic labs around the world talking up the ability of voice recordings to detect a surprising range of diseases. Current efforts extend beyond diagnosis of ailments with frequent respiratory components such as COVID-19, pneumonia, asthma, chronic obstructive pulmonary disorder (COPD), and pulmonary hypertension to include dementia, depression, psychosis, hypertension, and somnolence.
Boston-based Sonde is an independent spin out of Puretech Health (LSE:PTCHF), which incubated the company and provided its initial funding. Puretech continues to be involved through Sonde’s board.
Sonde’s voice technology was built on 1 million health-labeled voice samples obtained over the last five years through smartphone recordings provided through partnerships with hospitals and health providers as well as participants in the company’s research. Sonde’s August 2020 acquisition of Neurolex Laboratories Inc., a web-enabled voice survey and analysis company, added hundreds of thousands of voice samples.
“It’s very difficult to obtain high-quality, clinical, health-labeled voice samples and metadata, but we have more than anyone in the world,” Sonde CEO David Liu told BioWorld. “Without large amounts of health-labeled voice data, it’s impossible to create and validate the machine learning models necessary to detect changes in voice due to changes in health.”
How it works
The artificial intelligence underlying the platform was trained on thousands of voice samples from individuals with respiratory conditions such as asthma, pneumonia, persistent cough, and more. Speech pathologists, engineers and data scientists identified specific, objective acoustic features that correlated with respiratory conditions and indicated their severity.
The platform works best to detect changes in symptoms rather than making an initial diagnosis or distinguish between two diseases with respiratory symptoms, such as asthma, COPD, and COVID-19.
“Our testing is extremely rigorous, which is why we can say with certainty that our accuracy is over 70%,” Liu noted. The 70% rate refers to the sensitivity and specificity for a trio of symptoms – shortness of breath, chest pain or tightness, and cough.
That’s not stunning accuracy, but it has the advantage of being available for commercial use today.
The company is launching its application programming interface (API) as a product for other developers to use in their own iOS and Android apps. The HIPAA-compliant API comes with full documentation of its vocometers and allows customization with health-based questions and demographics. Sonde charges 0.15 per analysis to organizations using the platform and says the tool can be integrated into other systems in less than one day.
“At Sonde, we’re building products and technologies that empower people to proactively understand and manage their health, and we want to share them with the world,” said Liu. “We created the Developer Portal and self-serve API to spur innovation outside of Sonde so more people can benefit from our voice-enabled health check capabilities. We can’t wait to see what new use cases are imagined to deliver actionable health data and insights to people.”
Use in COVID-19
Sonde’s own development efforts are currently focused on marketing Sonde One, a COVID-19 detection app, to companies keen to conduct daily population screening to reduce the spread of COVID-19.
Sonde One uses the company’s advanced audio signal processing and machine learning platform to analyze a six-second recording, combined with a symptom check list, and a temperature check.
Before being used for screening or monitoring, the system is calibrated to each user through four initial recordings that establish a baseline, so that underlying conditions are factored into the analysis. When applying the technology to COVID-19, the calibration reduced the rate of false positives by 3%, Liu said.
Still, the app does not detect asymptomatic individuals with COVID-19, who are estimated to account for about 40% of positive cases. And, its 70% sensitivity and specificity means that at least 30% of the remaining 60% of cases are missed. Together, then, the app would tell about 58% of individuals who actually have COVID-19 that they are cleared to go to work.
On the other hand, the app is available now, requires no other materials, and is low cost.
Other companies and academic institutions are likely to give Sonde a run for the money in this application. Vocalis Health and MIT Open Voice both report significantly greater accuracy in detecting COVID-19, with MIT’s platform boasting it can detect asymptomatic cases 100% of the time. Beyond Verbal and Winterlight Labs/Pear Therapeutics have advanced voice-analysis tools for behavioral health.