Sundar Pichai, CEO of Google and Alphabet, speaks about artificial intelligence during a Bruegel think tank conference in Brussels, Belgium, on January 20, 2020.
Yves Herman | Reuters
Google on Wednesday announced MedLM, a suite of new healthcare-specific artificial intelligence models designed to help clinicians and researchers conduct complex studies, summarize doctor-patient interactions and more.
The move marks Google’s latest effort to monetize AI tools in the healthcare industry as competition for market share remains fierce among rivals such as Amazon and Microsoft. CNBC spoke with companies testing Google’s technology, such as HCA Healthcare, and experts say the potential for impact is real, though they are taking steps to implement it carefully.
The MedLM suite includes a large and a medium-sized AI model, both based on Med-PaLM 2, a large language model trained on medical data that Google first announced in March. It’s been generally available to eligible Google Cloud customers in the US since Wednesday, and Google said that while the cost of the AI suite varies depending on how companies use the different models, the mid-sized model is less expensive to run.
Google also said it plans to introduce versions of Gemini, its newest and “most capable” healthcare AI model, to MedLM in the future.
Aashima Gupta, Google Cloud’s global director of healthcare strategy and solutions, said the company has found that different medically tuned AI models can perform some tasks better than others. That’s why Google decided to introduce a range of models instead of trying to create a one-size-fits-all solution.
For example, Google said its larger MedLM model is better for performing complex tasks that require deep knowledge and a lot of computing power, such as conducting a study using data from a healthcare organization’s entire patient population. But if companies need a more flexible model that can be optimized for specific functions or real-time functions, such as summarizing a doctor-patient interaction, the medium-sized model should work better, according to Gupta.
Real use cases
A Google Cloud logo at the Hannover Messe industrial technology trade show in Hanover, Germany, Thursday, April 20, 2023.
Krisztian Bocsi | Bloomberg | Getty Images
When Google announced Med-PaLM 2 in March, the company initially said it could be used to answer questions like “What are the early warning signs of pneumonia?” and “Can incontinence be cured?” But as the company has tested the technology with customers, the use cases have changed, according to Greg Corrado, Google’s head of health AI.
Corrado said that clinicians don’t often need help with “accessible” questions about the nature of a disease, so Google hasn’t seen much demand for these features from customers. Instead, healthcare organizations often want AI to help solve more support or logistical problems, such as managing paperwork.
“They want something that helps them with the real pain points and slowdowns that exist in their workflow that only they know,” Corrado told CNBC.
For example, HCA Healthcare, one of the largest health systems in the US, has been testing Google’s AI technology since the spring. The company announced a official collaboration with Google Cloud in August aiming to use generative AI to “improve workflows on time-consuming tasks.”
Dr. Michael Schlosser, senior vice president of care transformation and innovation at HCA, said the company is using MedLM to help emergency medicine physicians automatically document their interactions with patients. For example, HCA uses an ambient speech documentation system from a company called Augmedix to transcribe doctor-patient encounters. Google’s MedLM suite can then take these transcripts and break them down into the components of an ER provider note.
Schlosser said HCA is using MedLM in emergency rooms at four hospitals, and the company wants to expand use over the next year. By January, Schlosser added, he expects Google’s technology to be able to successfully generate more than half of the notes without help from carriers. For doctors who can spend up to four hours a day in offices, Schlosser said the savings in time and effort make a significant difference.
“This was a huge leap forward for us,” Schlosser told CNBC. “Now we think we’re going to be at a point where AI, on its own, can generate 60+ percent of the note correctly on its own, before we have humans do the review and editing.”
Schlosser said HCA is also working to use MedLM to develop a handover tool for nurses. The tool can read the electronic health record and identify relevant information for nurses to pass on to the next shift.
Handles are “laborious” and a real pain point for nurses, so it would be “powerful” to automate the process, Schlosser said. Nurses at all HCA hospitals perform approximately 400,000 handoffs per week, and two HCA hospitals are piloting the nurse handoff tool. Schlosser said nurses conduct a side-by-side comparison of a traditional handover and an AI-generated handover and provide feedback.
With both use cases, however, HCA has found that MedLM is not infallible.
Schlosser said the fact that AI models can spit out incorrect information is a big challenge, and HCA is working with Google to find best practices to minimize those fabrications. He added that token limits, which limit the amount of data that can be fed into the model, and managing AI over time are additional challenges for HCA.
“What I would say now, is that the hype surrounding the current use of these AI models in healthcare is outpacing the reality,” Schlosser said. “Everybody is facing this problem, and nobody has really let these models loose in a scalable way in health care systems because of it.”
Still, Schlosser said providers’ initial response to MedLM has been positive, and they acknowledge they’re not yet working on the final product. He said HCA is working hard to implement the technology responsibly to avoid putting patients at risk.
“We’re very careful with how we approach these AI models,” he said. “We don’t use these use cases where the results of the model can somehow affect someone’s diagnosis and treatment.”
Google also plans to introduce healthcare-specific versions of Gemini to MedLM. Its shares rose 5% after the launch of Gemini earlier this month, but Google faced scrutiny over its demo video, which was not conducted in real time, the company confirmed Bloomberg.
In a statement, Google told CNBC: “The video is a demonstration of what Gemini can do to interact, based on real multimodal prompts and test output. We can’t wait to see what people create when Gemini Pro opens on December 13 .”
Google’s Corrado and Gupta said Gemini is still in the early stages and needs to be tested and evaluated with customers in controlled healthcare settings before the model is rolled out more widely through MedLM.
“We’ve been testing Med-PaLM 2 with our customers for months and now feel comfortable using it as part of MedLM,” said Gupta. “The Gemini will follow suit.”
Schlosser said HCA is “very excited” about Gemini and the company is already working on plans to test the technology, “We think it can give us an extra layer of performance when we get it,” he said.
Another company using MedLM is BenchSci, which aims to use artificial intelligence to solve problems in drug discovery. Google is one investor on BenchSci, and the company has been testing its MedLM technology for a few months now.
Liran Belenzon, co-founder and CEO of BenchSci, said the company has merged MedLM’s artificial intelligence with BenchSci’s own technology to help scientists identify biomarkers, which are key to understanding how a disease progresses and how it can be cured.
Belenzon said the company spent a lot of time testing and validating the model, including providing feedback to Google about needed improvements. Now, Belenzon said BenchSci is in the process of bringing the technology to market more broadly.
“[MedLM] it doesn’t work out of the blue, but it helps accelerate your specific efforts,” he said in an interview with CNBC.
Corrado said research around MedLM is ongoing, and he believes Google Cloud healthcare customers will be able to coordinate models for many different use cases within an organization. He added that Google will continue to develop domain-specific models that are “smaller, cheaper, faster, better.”
Like BenchSci, Deloitte tested MedLM “over and over again” before rolling out the technology to healthcare clients, said Dr. Kulleni Gebreyes, Deloitte US Head of Life Sciences and Healthcare;
Deloitte uses Google technology to help health systems and health plans answer members’ questions about access to care. If a patient needs a colonoscopy, for example, they can use MedLM to search for providers based on gender, location or benefit coverage, as well as other criteria.
Gebreyes said customers have found MedLM to be accurate and effective, but, like other models, AI isn’t always great at deciphering a user’s intent. It can be a challenge if patients don’t know the correct word or spelling for colonoscopy or use other colloquial terms, he said.
“Ultimately, this is not a substitute for a diagnosis by a trained professional,” Gebreyes told CNBC. “It brings expertise closer and makes it more accessible.”