GOOGLE is developing software that can identify early signs of a disease – all you have to do is cough into your phone mic.
The tech behemoth is no stranger to the development of artificial intelligence. If anything, Google is aggressively pushing to integrate AI into existing services from video conferencing to email.
Now the company is taking a different approach. In an August 19 blog post, Google research director of engineering Shravya Shetty teased the company’s latest AI model.
“From cough to speech and even breath, the sounds our bodies make are filled with information about our health,” the post read.
“As researchers at Google, we recognize the potential of sound as a useful health signal, and also that smartphone microphones are widely accessible.”
The company has been examining ways to extract insights from acoustic data.
Theoretically, a user could cough into their phone and receive potential diagnoses, courtesy of AI.
That is the goal of Health Acoustic Representations (HeAR), a model trained on 300 million sound clips.
The “cough model,” in particular, was trained on roughly 100 million cough sounds.
HeAR learns to identify patterns within audio data, laying the grounds for further analysis.
Google claims the model ranks higher than competitors “on a wide range of tasks and for generalizing across microphones.”
As HeAR is just a foundation model, Google hopes researchers will fine-tune it for other purposes, like identifying a specific disease.
Models trained using HeAR already show promise, with Google reporting “high performance” despite sparse training data.
The tech giant has partnered with Salcit Technologies, an Indian healthcare company that developed its own model to analyze cough sounds.
Researchers are particularly interested in finding ways to identify tuberculosis, a devastating bacterial infection that can lie dormant for years.
The immune system captures the germs in what is known as the primary infection stage – but some may survive and multiply.
If the disease progresses to the active stage, it can spread outside the lungs, affecting the kidney, liver, and heart.
Tuberculosis typically presents with cancer-like symptoms including fatigue, shortness of breath, and chest pain.
While doctors can stop the disease in its tracks through early intervention, it can be life-threatening if untreated.
Tuberculosis is uncommon in the Western world but prevalent in parts of Southeast Asia where access to healthcare is limited.
Google intends for HeAR to bolster the development of diagnostic tools and monitoring solutions for TB and other chest and lung diseases.
If everything goes according to plan, the AI will “help improve health outcomes for communities around the globe,” even in the presence of cost barriers or limited healthcare data.
The model is currently available to researchers by request. It is unclear when access will expand to the public.
While many netizens have expressed support for the tool – with one Reddit user claiming, “This is what AI should be” – others are wary.
You may have heard of Google's Gemini AI tool - so what exactly is it?
Google’s Gemini AI launched in December 2023 and is available online, billed as a way to “supercharge your creativity and productivity”.
Gemini is a multimodal model that learns from a variety of data types including images, text, and audio.
When a user enters a prompt into Gemini, it generates a response using information it already knows or pulls from other sources (often Google services).
While training on datasets, it identifies patterns that help it mimic a human response. As it is continuously learning, Gemini also learns from your prompts, responses, and feedback.
Google has admitted that “Gemini will make mistakes and might even say something offensive”.
The program occasionally cites its sources. If it quotes at length from a webpage, for instance, it makes reference to that page. Sometimes, it generates a URL that users can click.
Gemini has usage limits to reduce traffic, meaning it may cap the number of prompts and conversations a user can have within a specific timeframe.
This number depends on factors like how long and complex a user’s prompts are and the length of the conversation with Gemini. Google will alert you when you are close to hitting the limit for a given period.
Data privacy concerns remain a pressing matter in the age of artificial intelligence, as models scrub large swathes of data from users, usually without permission.
“Seems like an excuse for more privacy invasions,” one Reddit user wrote.
“Privacy nightmare continues,” another lamented.
Even for those who recognize the potential in the tool, their optimism has limits.
“This kinda thing is incredibly awesome, but it’s terrifying to think that the most interested industry will be health insurance providers,” one user wrote.
