In battle with Microsoft, Google bets on medical AI program to beat healthcare industry

ChatGPT, a computer program that can smoothly respond to a variety of cross-subject queries, was released in November, marking the beginning of early experiments to use the underlying technology for patient care in healthcare systems across the United States.

Google is betting that its own medical chatbot technology, called Med-PaLM 2, will be able to conduct conversations about medical issues better than general-purpose algorithms, as it inputs questions and answers for medical licensing exams. The company began testing the system with customers, including the research hospital Mayo Clinic, in April, according to people familiar with the matter.

According to Google executives and research published by the company, Med-PaLM 2 can be used to generate answers to medical questions and perform tasks such as summarizing documents and organizing health data sets. It is possible.

The healthcare industry has become the new frontier in the battle between tech giants and smaller start-ups for customers with AI products, but past efforts such as IBM’s Watson Health Initiative At times, it struggled to turn technology into sustainable profit.

Medical leaders and ethicists believe that generative AI has the potential to transform medicine, but patients need to be educated about how their health data will be used in new ways, and as new tools are rolled out. said it needed to be evaluated. Alphabet’s Google has faced scrutiny in the past over how it handles sensitive medical data through partnerships with hospitals.

AI algorithms are already being used in hospitals for specialized tasks, such as predicting heart disease from a patient’s ECG. Generative AI tools can be used to generate authoritative-looking answers to medical questions, introducing a new set of risks and injecting patients in ways doctors don’t recommend. can affect.

Google executives said customers testing Med-PaLM 2 would maintain control over their data in an encrypted setting that the technology companies couldn’t access, and the program wouldn’t capture that data.

A Google spokesperson declined to say when the program will be broadly available to customers or the general public.

Google’s competitors are quickly incorporating AI advances into patient interactions. Microsoft, OpenAI’s largest investor and closest business partner, partnered with health software company Epic in April to create a tool that can automatically draft messages to patients using the algorithm behind ChatGPT. built.

These services could fuel cloud computing operations, a focus area for tech giants pushing the potential of AI programs. In 2021, Google opened an office in Rochester, Minnesota, near the Mayo Clinic headquarters, to work on projects using hospital data. The hospital announced in June that it would build a new internal search tool to query patient records using a Google AI model.

Both Google and Microsoft are also interested in building a virtual assistant to answer medical questions from patients around the world with larger ambitions, especially in areas with limited resources, according to company documents.

An internal email reviewed by The Wall Street Journal, citing researchers working on the project, said in April that the AI ​​model trusted as a medical assistant would have “restricted access to doctors.” He told employees that it could be “very valuable in the country where it is located.” .

In a paper published in March, Microsoft and OpenAI said algorithms such as the GPT-4 program behind ChatGPT “could be leveraged to provide information, communication, screening, and decision support in underserved areas.” has potential,” he said.

Google senior research director Greg Collard, who worked on Med-PaLM 2, said the company is still in the early stages of developing products using the technology and working with customers to understand their needs.

“I don’t think I’m at a stage where I would want this type of technology in my family’s healthcare efforts just yet,” Corrado said. But Med-PaLM 2 “occupies a place where AI becomes beneficial and its use expands in healthcare,” he said.

Google has withheld some of its cutting-edge AI programs from the public due to concerns about their safety and potential impact on its core online search business. This vigilance created an opportunity for Microsoft and his OpenAI, and both companies moved more quickly to release the popular ChatGPT chatbot to the public and give customers access to the underlying AI system.

Hospitals have begun testing OpenAI’s GPT algorithms via Microsoft’s cloud service on tasks like summarizing doctors’ notes and generating reminders. In these cases, Microsoft hosts and controls the AI ​​systems, a spokesperson said.An independent study released by the companies found that Google’s Med-PaLM 2 and OpenAI’s GPT-4 are each used in medical trials. A similar score was recorded on the question.

Physicians and medical executives said further development and testing are still needed before programs such as Med-PaLM 2 can be used to diagnose patients and suggest treatments.

The World Health Organization, citing the large language model underpinning chatbots in May, is concerned that “the usual attention given to new technologies has not been consistently given to LLMs.” said.

A study released by the company in May found that physicians who reviewed the answers provided by Med-PaLM 2 to more than 1,000 consumer health questions scored 8 out of 9 categories of ratings defined by Google. He said he preferred the system’s answer to the doctor-generated answer along the lines of

However, doctors found that Med-PaLM 2’s responses contained more inaccurate or irrelevant content than other chatbots. This suggests that the program shares similar problems with other chatbots that tend to generate off-topic or false statements in confidence.

Google researchers said the program’s ability to avoid inaccurate or irrelevant information did not show significant improvement from the first version announced in December.

“There’s no way to measure these things at scale,” said Deb Dash, a clinical assistant professor at Stanford University School of Medicine who studies AI applications in medicine. “This is a very ongoing study.” says.

Patients need to be educated about new ways AI tools can use their health data, said Kelly Owens, a medical ethicist at New York University’s Grossman School of Medicine.

Ideally, “it should be a human-to-human conversation” between patients, doctors and medical staff, rather than disclosures embedded in consent forms, Owens said.

Neither Google nor Microsoft said they used patient data to train their algorithms. Corrado said Google could allow healthcare companies to create custom versions of Med-PaLM 2 using patient records and other data, but that’s not possible at this time.

Email Miles Kruppa (miles.kruppa@wsj.com) and Nidhi Subbaraman (nidhi.subbaraman@wsj.com).