Dive Brief:
- Google Cloud expanded its suite of AI tools for healthcare workers on Thursday, rolling out a generative AI search tool that allows payers and providers to comb through patient notes and scan documents and other data for clinical information.
- The tool, Vertex AI Search for Healthcare, was first released for limited availability in March. The tech giant also expanded its Healthcare Data Engine, which provides a longitudinal record of patient data in an industry standard format.
- Google promises the tool has minimal risk of AI “hallucinations” — or outputting inaccurate responses.
Dive Insight:
A survey conducted by Google Cloud and the Harris Poll found that clinicians and claims staff log 27 and 36 hours per week, respectively, on administrative tasks including documentation.
Google says Vertex AI Search attempts to cut down that time by helping workers query multiple aspects of a patients’ medical record. The tool integrates with Gemini 1.5 Flash and MedLM – two of Google’s large language models — to search.
Vertex was first teased during the HIMSS conference in March. It’s been used by health systems including Community Health Systems and Highmark Health.
Google also promises Vertex AI will cite its sources and provide links to internal sources of information — a technique called “grounding” — to increase providers’ confidence in the tool’s answers, according to the release.
The safeguard comes amid growing concern about whether generative AI might output false or misleading information.
An August study, conducted by researchers at the University of Massachusetts at Amherst and scientists at clinical AI company Mendel, found generative AI hallucinated when performing healthcare queries. AI tools studied produced medical record summaries containing false or misleading statements, creating potential for providers to misdiagnose patients or recommend inappropriate treatment plans.
Google’s use of grounding is a “meaningful step,” toward addressing hallucinations, according to Jeffrey Cribbs, a distinguished vice president analyst on consultancy Gartner’s healthcare team. However, citing where the information comes from could make the tool less effective at its original goal of reducing administrative burden.
“What I’ve just given you then is a big research assignment,” Cribbs said. “The whole value proposition here is that we want to be able to save people time. We can’t expect them to go back to every piece of the medical record, or we’re not saving them time. So, over time, we want the citations to be used by exception, rather than by rule.”
In order for providers to trust that the AI got a provider’s query right without manual review, Cribbs said that the industry will want to adopt a “higher level of validation” for generative AI.
Google has launched several generative AI ventures to date, including large language models tuned to search patients’ health records.
The company has become a leader in generative AI testing, according to Cribbs. Google is known in the industry for robust validity testing and checking measures for health equity amid concerns about embedded bias in algorithms.
Still, Cribbs said that generative AI validation assessment is still “an emerging practice,” and some would-be consumers have concerns about search tools like the one Google unveiled Thursday. Of particular concern is whether clinicians might default to trusting AI’s output over their own clinical expertise — a phenomenon called “automation bias.”
“There is very broad concern among healthcare organizations considering adopting AI in clinical workflow about how that the use of that tool may expose [them to] new kinds of risks,” Cribbs said.
Still, when implementing a new AI tool Cribbs stressed that the goal is not perfection.
“We are we are not comparing against perfect information extraction that exists today,” Cribbs said. “The fact of the matter is that information is not extracted from clinical records in healthcare in an efficient or in an accurate way in the market today, and so we’re looking for substantially better.”
link