Google Gemini health check: how its AI services aim to help the medical industry (and make us fitter)

According to Google, around 30% of all data being generated today is coming from the healthcare industry. Google is hoping it can capitalise on the data growth with generative AI, and used its The Check Up event this week to unveil a number of updates to artificial intelligence products that are in the works.  

Perhaps unsurprisingly, Google announced it’s working on a version of Gemini – its general AI model released in December 2023 –  for the healthcare industry. It added it’s already “researching” how a healthcare-specific Gemini might work across various data modalities, including medical imaging such as CT scans and X-rays. According to the company, Gemini could be used to review the scans and look for evidence of pathologies. This generates information that’s then used by doctors to diagnose and treat patients. 

Google is also targeting medical imaging with the release of MedLM for chest X-rays. At The Check Up, it announced that the product is now available for “trusted testers in an experimental preview on Google Cloud”.  

The company hasn’t provided further details on how it will work, beyond the generic statement that it could be used for “the classification of chest X-rays for a variety of use cases”. Chest X-rays are some of the most commonly requested imaging in healthcare, with nearly 200,000 ordered per year in the UK by GPs alone.

MedLM is a group of healthcare-focused AI models built on the company’s large language model Med-Palm2. Med-Palm2 was originally developed for extracting useful information from ever-expanding repositories of medical documentation – making summaries of medical notes, for example – but has since been expanded to become multi-modal.

Read more: Google Gemini is testing with 10 million tokens | TechFinitive

Gemini smart enough to ace medical exams

Google also announced that Gemini has been brushing up on its medical knowledge by learning to answer “USMLE-style” questions. The USMLE is a series of three exams sat by medical school graduates to get their US medical licence. Pass rates among flesh-and-blood medics that take USMLE exams is 95% to 98%, depending on which exam they take, while Gemini managed 91%.

Initiatives aimed at mitigating the biases of AI in the healthcare space were also unveiled at The Check Up, including a Google-developed framework called Health Equity Assessment of machine Learning performance (HEAL). This assesses whether AIs perform equally well across different subpopulations, including marginalised communities.

Google has already used HEAL while testing out a dermatology model. The framework spotted that the model wasn’t performing as well in one particular group of patients, the over 70s, when identifying non-cancerous skin problems.

Alongside its efforts targeting healthcare providers, Google is also tweaking consumer-facing health products. Google Research and Fitbit are building a Personal Health LLM that will be used to make the Fitbit app more personalised. The model might be used “to analyse variations in your sleep patterns and sleep quality, and then suggest recommendations on how you might change the intensity of your workout based on those insights,” Google said.

New AI features will be released to Fitbit’s Premium users first, via Fitbit Labs.

Worth a read

• AWS and IBM each partner with Nvidia in AI push
• Shea Martin, CEO, Brew Ninja: “You will need the passion to get past the naysayers”
• Stephen Lester, CTO at Paragon: “Though the spotlight has remained on AI it will be some time before we see genuinely groundbreaking use cases”