Google got together last week with an Indian AI startup to roll out a bioacoustics health-care model to detect disease from human sounds.
Bioacoustics is the thrilling combination of biology and acoustics, helping us gain insights from sounds produced by animals and humans. Generative AI, the kind of tech that brought ChatGPT to 200 million users worldwide, is now adding a new level of functionality to this field.
One foundation AI model built by Alphabet Inc.’s Google uses sound signals to predict early signs of disease, opening up a world of possibilities. The technology can ride in a smartphone and track high-risk populations in tricky geographies.
Where expensive diagnostic hardware such as X-ray machines or technical expertise isn’t within easy reach, AI along with a handset’s microphone could come to the rescue.
That AI system is already helping tackle the world’s top infectious killer. Nearly 4,500 people die daily and 30,000 fall ill with tuberculosis, according to the World Health Organization. It is treatable, but millions go undiagnosed. In India, the disease fells nearly a quarter-million people yearly, and early detection is key to halting its spread.
Google trained its foundation AI model with 300 million pieces of audio from around the world, including coughs, sniffles, sneezes and breathing. Two-second audio clips were amassed from non-copyrighted, publicly-viewable content such as YouTube videos and even cough sounds recorded in a hospital in Zambia, where patients came for TB screening. Body sounds are filled with information about our well being, containing near-imperceptible clues that can help screen, diagnose and manage health conditions.
The data feeding Google’s HeAR (short for Health Acoustic Representations) AI model included 100 million cough sounds that now help detect TB.
The AI tool loaded on a smartphone is easily carried to the remotest populations to screen for the disease. The AI detects early signs based on subtle differences in cough patterns, helping triage patients and lining them up for further investigation and treatment, said Shravya Shetty, Google’s Mountain View, California-based research director of engineering. The goal is to power the tools to help even people with limited training screen for respiratory illnesses, she says.
The tech giant’s Indian collaborator, Hyderabad-based Salcit Technologies, is a respiratory health care AI startup. Salcit is using the AI model to improve the accuracy of TB diagnosis and lung health assessments by combining with its own machine learning AI called Swaasa, which is the word for breath in the ancient Indian language of Sanskrit.
Leading Indian health care providers like Apollo Hospitals and the nonprofit Healing Fields Foundation are using Swaasa to screen people, including in remote areas. Salcit has India’s medical device regulator’s approval, a first for a software tool to be deployed as a medical device. In its mobile app, Swaasa allows users to upload a 10-second cough sample (just cough near your mobile) and test for diseases with a 94% accuracy, Salcit’s co-founder Manmohan Jain said.
The cough sound is the equivalent of giving a blood sample, only this particular sonic sample is processed on the cloud rather than in a laboratory. The screening test retails at 200 rupees ($2.40), compared with a spirometry test that could cost 3,000 rupees at a clinic.
But there are challenges.
While the new tech is exciting doctors in the field by opening up a new frontier, it’s not easy to change routine clinical practices. The screening tool will need to find acceptance.
There’s also the problem of ensuring audio samples don’t come with an abundance of background noise. Rural users, unfamiliar with technology, may be unable to record coughs on the app. Yet, the tech is finding supporters, including those like the StopTB Partnership, a UN-backed organization, which aims to end TB by 2030.
In another bioacoustics venture, Google is researching a model based on ultrasound for early breast cancer detection at the Chang Gung Memorial Hospital in Taiwan. The AI assists in lesion detection and Google aims to roll out it out globally, offering free breast cancer screening for populations that can’t afford costly mammograms.
Neither of Google’s models is yet near commercialization. But sound-based generative AI systems could democratize early disease detection, making screening accessible, affordable and scalable.
\Montreal-based Ubenwa has built a foundation model for infant cries, and interprets infant’s needs and health by analyzing the biomarkers in their cry sounds. And others are working on AI tools that can detect autism based on oohs, aahs and gurgling sounds. Voice and sound are the new frontiers in medicine, says Salcit’s Jain.
