Google flexes its health care AI muscle

Google flexes its health care AI muscle
Illustration of a robot arm lifting a medical cross

Illustration: Annelise Capossela/Axios

Google showed off an array of new artificial intelligence (AI)-driven health care tools on Tuesday, from a souped-up chatbot that can shed light on your medical symptoms to enhanced search features that tell you if a doctor takes Medicaid.

Why it matters: There’s an arms race among big tech companies to infuse their products with AI — but the results, particularly in health care, can have unwanted consequences or pitfalls, like racial bias, privacy concerns and ethical problems.

Driving the news: The “large language model” that Google has been building for the medical world — an AI chatbot called Med-PaLM 2 — now consistently passes medical exam questions with a score of 85{bf0515afdcaddba073662ceb89fbb62b6b1bf123143c0e06b788e1946e8c353f}, placing it at “expert” doctor level, the company said.

  • That’s an 18{bf0515afdcaddba073662ceb89fbb62b6b1bf123143c0e06b788e1946e8c353f} improvement from the system’s previous performance, per the company, and “far surpasses similar AI models.”
  • A rival generative AI tool, ChatGPT, also passed the medical exams — but just barely. (ChatGPT’s creator, OpenAI, just released a new, more powerful version of its underlying tech.)
  • Google’s system is being trained to “retrieve medical knowledge, accurately answer medical questions, and provide reasoning,” the company says.

Yes, but: Google acknowledges AI’s shortcomings in the medical realm.

  • “There’s still a lot of work to be done to make sure [Med-PaLM 2] can work in real-world settings,” reads a blog post from Yossi Matias, a Google vice president of engineering and research, and Greg Corrado, its head of health AI.
  • Google found “significant gaps” when the tool was “tested against 14 criteria — including scientific factuality, precision, medical consensus, reasoning, bias and harm,” per the post.
  • “We look forward to working with researchers and the global medical community to close these gaps and understand how this technology can help improve health delivery.”

Meanwhile: Google’s conversational AI technology Duplex has called hundreds of thousands of U.S. health care providers to see if they accept Medicaid. The results are being displayed in Google Search, ahead of a March 31 re-enrollment deadline.

  • Google Search results will also soon highlight “providers that identify as community health centers offering free or low-cost care,” the company said.
  • And an improved version of Fitbit’s health metrics dashboard will soon be available — in some cases without a subscription — to help people map changes and trends in their breathing rate, skin temperature and blood oxygen.

What they’re saying: “The future of health is consumer-driven,” Karen DeSalvo, Google’s chief health officer, told reporters.

  • “People will expect a mobile-first experience with more personalized insights, services and care.”

Zoom out: Google is also deploying AI tools to help offer high-quality, low-cost medical diagnostics globally.

  • Ultrasound devices with Google AI are being used to detect breast cancer in Taiwan and determine gestational age in expectant mothers in Kenya.
  • Another Google AI tool that checks chest X-rays for signs of tuberculosis is being rolled out in Sub-Saharan Africa.
A schematic image of an ultrasound of a pregnant uterus and a screen that shows gestational age.
Google’s AI-powered ultrasound system. Image: Courtesy of Google

Between the lines: The company is acutely aware of criticisms that “Dr. Google” can sometimes lead users to misleading or dangerous health guidance.

  • Another new YouTube feature helps people “find human answers” to their health questions, said Garth Graham, director and head of health care and public health at Google/YouTube.

Google also announced a partnership with ThroughLine, which connects users to free mental health crisis support in more than 100 countries.

  • “This will increase the number of crisis helplines that appear at the top of search results,” reads a blog post from DeSalvo. The feature is aimed at “searches related to suicide, domestic violence and other personal crisis topics.”

The big picture: Even as they take breathtaking strides toward improving health care with AI, tech companies and others have been stumbling with what they’ve unleashed.

  • Older people covered by Medicare Advantage are finding their benefits cut off by AI algorithms no matter how dire their medical needs, a Stat News investigation found.
  • Verily Life Sciences, a unit of Google parent company Alphabet that focuses on treatment personalization, started laying off staff in January, per the Wall Street Journal.

On the plus side, AI is starting to detect breast cancers that radiologists miss, as the New York Times reports.

  • The development is “one of the most tangible signs to date of how AI can improve public health,” per the Times — yet the technology produces too many false positives and still needs to “show it can produce accurate results on women of all ages, ethnicities and body types.”

The bottom line: Medicine will always need to rely on a healthy mix of technology and human know-how.

  • “AI on its own cannot solve all of health care’s problems,” Alan Karthikesalingam, the surgeon-scientist who leads Google Health’s machine learning research group, said Tuesday.
  • “Medicine, after all, is about caring for people.”