Google is among several organisations leveraging audio signals to identify early indicators of health issues. The tech giant has trained one of its foundational AI models using 300 million audio clips that include coughs, sniffles, and laboured breathing, enabling it to recognise conditions such as tuberculosis.
To enhance this technology, Google has partnered with Salcit Technologies, an Indian startup that focuses on respiratory healthcare AI.
Salcit Technologies, based in Hyderabad, is utilising Google’s AI model to improve the precision of tuberculosis diagnoses and lung health evaluations. This is achieved by combining Google’s technology with its own machine learning system known as Swaasa, which translates to “breath” in Sanskrit.
The Role of Bioacoustics
Bioacoustics, which merges biology and acoustics, allows for insights derived from the sounds made by humans and animals. Generative AI, the technology behind tools like ChatGPT, is now enhancing this field by adding new functionalities.
Google’s foundational AI model employs sound signals to predict early signs of diseases and presents numerous opportunities for public health. This technology can operate through smartphones, providing a means to monitor high-risk populations in regions where expensive diagnostic equipment, such as X-ray machines, is not readily available.
Addressing Tuberculosis
This AI system is already contributing to the fight against tuberculosis, which claims nearly 4,500 lives daily and affects 30,000 individuals, as reported by the World Health Organization. Although tuberculosis is treatable, millions remain undiagnosed. In India, the disease impacts approximately a quarter-million people annually, making early detection crucial for controlling its spread.
Google’s AI model was trained using a vast array of audio samples, including 100 million cough sounds, sourced from publicly available content like YouTube videos and recordings from a hospital in Zambia, where patients were screened for tuberculosis. These body sounds carry vital information about health, offering subtle clues that can assist in screening, diagnosing, and managing health conditions.
The HeAR (Health Acoustic Representations) AI model utilises these data points to help detect tuberculosis. The AI tool can be easily transported to remote areas, allowing for disease screening based on distinct cough patterns. According to Shravya Shetty, Google’s research director of engineering, this aids in triaging patients and directing them for further evaluation and treatment. The aim is to empower even those with limited training to screen for respiratory illnesses.
Implementation and Accessibility
Leading healthcare providers in India, such as Apollo Hospitals and the Healing Fields Foundation, are employing Swaasa to screen individuals, including those in underserved regions. Salcit has received approval from India’s medical device regulator, marking a first for a software tool used as a medical device.
“The Swaasa app allows users to upload a 10-second cough sample, achieving a 94% accuracy rate in disease detection,” stated by Manmohan Jain, Salcit’s co-founder.
This audio-based assessment is akin to providing a blood sample, but it is processed in the cloud rather than a laboratory. The screening test is priced at 200 rupees, significantly lower than the 3,000 rupees charged for a spirometry test at clinics.
However, the screening tool must gain acceptance among healthcare professionals, and there are concerns about ensuring audio samples are free from excessive background noise. Additionally, rural users unfamiliar with technology may struggle to record coughs accurately using the app. Nevertheless, support is growing, including from organisations like the StopTB Partnership, which aims to eliminate tuberculosis by 2030.
Future Prospects in Bioacoustics
In another bioacoustics initiative, Google is exploring a model based on ultrasound for early breast cancer detection at Chang Gung Memorial Hospital in Taiwan. This AI assists in identifying lesions, with plans for global rollout to provide free breast cancer screenings to underserved populations.
While neither of Google’s AI models is yet ready for commercialisation, sound-based generative AI systems hold the potential to democratise early disease detection, making screening more accessible, affordable, and scalable. Other initiatives, like Montreal-based Ubenwa, are developing models to interpret infant cries for health insights, while additional projects aim to detect autism through specific sound patterns.