(Bloomberg) -- Nini Loomis, a registered nurse with almost two decades of experience, was tasked with testing a new Google-powered tool that uses artificial intelligence to analyze patient medical records. The technology is designed to generate a concise report that can be seamlessly handed from one caregiver to the next at the end of a shift.
But when Loomis scrutinized a readout on one of her patients, she spotted a glaring, potentially life-threatening omission: The software failed to note whether the person had any drug allergies.
A patient’s allergy history can help avert deadly disasters, like anaphylactic reactions to medication. “This should say, ‘no known drug allergies,’” said Loomis, whose Florida hospital is part of the HCA Healthcare group, as she pointed to a section of the report where information was missing. The details of HCA’s nurse-handoff pilot testing, as well as tests of an AI-powered documentation app for emergency room physicians in use at HCA hospitals, are being reported on in-depth for the first time by Bloomberg.
AI is a once-in-a-generation technological breakthrough, like the automobile or the internet, that has the power to transform everything. For the past two years, Google’s parent company, Alphabet Inc., has been working furiously to bring generative AI to the public with a chatbot and more conversational search results. But the technology could also help the company finally conquer what has become one of its most elusive goals: making inroads in health care. And breaking through in the $4 trillion US health-care industry just might give Alphabet a leg up against other tech competitors in the AI race.
After years of strategy shifts, billions of dollars in investments and a long list of health-care moonshots that have failed to take off — like glucose-sensing contact lenses, cancer-detecting pills and Google Glass for surgery — Alphabet needs a win. The nurse handoff note is one of the many areas of health care that providers want artificial intelligence to improve, cutting down on tedious, clerical parts of the job to leave them more time with patients. But Google’s ambitions are even grander. The company sees promise in AI that, one day, might be trained on all of humanity’s medical knowledge and could deliver personalized health insights to patients.
AI has the “potential to transform health on a planetary scale,” Karen DeSalvo, Google’s chief health officer, said at an annual company event in March.
But right now, the technology simply isn’t good enough.
There are still questions about whether Google’s AI can consistently deliver high-quality health information, according to more than 30 physicians, nurses, health-care executives, industry analysts and experts that Bloomberg interviewed for this story. And patients and health-care workers are skeptical about artificial intelligence in general. Over the past few months, nurses and other caregivers have protested the use of AI in hospitals, citing concerns about the technology’s potential to harm patients and replace human jobs. Some 60% of Americans would be uncomfortable with medical providers relying on AI to deliver health care, according to a survey conducted last year by the Pew Research Center.
Then there’s the wide chasm between Alphabet’s utopian vision of using AI to democratize health care and the tools it’s piloting, which for the moment are largely focused on streamlining industry inefficiencies. The nurse handoff technology that Loomis tested, which was developed by HCA using Google’s AI, is currently being used with real patients at four HCA Healthcare hospitals in Tennessee and Florida. Google has also partnered with medical-transcription company Augmedix on a tool for automatically documenting ER visits. Other projects include an experiment with Deloitte to use chatbots to help health systems answer people’s questions about accessing care and a family of large language models, akin to ChatGPT, called MedLM that doctors could deploy in a number of ways.
“We are incredibly intentional about our pacing when it comes to developing AI models for health care,” a Google spokesperson said in a statement in response to questions from Bloomberg News. “We publish research, conduct studies with health-care organizations, pilot new models or tools, gain feedback and iterate. There is a big opportunity but we need to do it safely to get it right.”
Google isn’t the only tech giant that has stumbled on its way to disrupting the stodgy and splintered US health-care industry. Apple Inc. ditched or delayed countless initiatives, including a glucose monitor and health clinics. Amazon.com Inc. shuttered a telehealth service and a project that aimed to lower employer health insurance costs. Microsoft Corp. shut down a 12-year effort to improve medical records management. Oracle Corp.’s $28 billion acquisition of Cerner Corp., which the company said would transform health-care technology, led instead to lost customers and slipping sales. The companies have all poured billions of dollars into giving health care a Silicon Valley makeover; so far none of them have produced much in the way of results.
“If Google’s overall ambition is to completely disrupt the health-care industry, well, nobody in the big tech world has succeeded,” said Rajiv Leventhal, a senior analyst with Emarketer researching digital health. “Health care is a unique beast.”
An AI revolution
Google’s dream of a tech-enabled health revolution began almost two decades ago. In 2008, after a year and a half of development, the company launched a free personal health record platform, dubbed Google Health, that attracted high-profile partners including Walgreens, the American Heart Association and the Cleveland Clinic.
For Google’s founders, Larry Page and Sergey Brin, health care is deeply personal. Several years ago, Page fell ill and his vocal cords were paralyzed, weakening his voice. Brin’s mother has battled Parkinson’s disease for 25 years, and Brin carries a genetic mutation that puts him at higher risk of eventually developing the debilitating brain disorder himself.
Standardizing health records and bringing them online, they believed, would help people better manage their health and allow the industry to treat sick people more efficiently. But the company shuttered the records platform just four years after it launched. It would be the first of many health projects to enter the Google graveyard.
Now, AI has created a fresh opportunity for Alphabet in health. In late 2022, after the tech giant declared a “code red” over the popularity of a new AI program called ChatGPT, a small team within Google’s AI division sprinted over the Thanksgiving and Christmas holiday stretch to publish research on its new health-focused AI model, called Med-PaLM — the first AI system to achieve a passing score when answering US medical exam-style questions. Google said the AI had the potential to transform health care by allowing doctors to retrieve medical knowledge quickly to support their clinical decisions. More advanced versions of the model, called MedLM, are now available to the tech giant’s health-care cloud clients, including some hospitals and medical centers.
“The pace at which things are moving is really a revolution,” said Yossi Matias, head of Google Research in an interview with Bloomberg this March.
Generative AI is developed by ingesting and learning from vast amounts of data, which enables it to, for example, automate tasks like summarizing a health record or interpreting an X-ray. Google thinks that someday, its health-focused AI could create personalized care plans, responding to people’s health as they age or even anticipating disease. Theoretically, an AI system that can perfectly recall the details of any published medical research would fare better than a doctor who can’t possibly remember everything. This AI could also be infinitely scaled and deployed anywhere, including places with few doctors or nurses.
“This technology has the ability to be a care multiplier,” said Greg Corrado, cofounder of an internal research lab, Google Brain, in an interview. For the last seven years he has focused on applications of artificial intelligence to health. A Google spokesperson later added that the generative AI technologies that Google is testing in hospitals are not meant to substitute for human care, and said its tools are being built to keep humans as part of the process.
But Alphabet has a long way to go before its AI is diagnosing patients. It has struggled to develop a unified health strategy, with fractured efforts dispersed across different divisions. Verily, a sister company of Google, works on life sciences initiatives, including a recent push into testing wastewater for viruses. Another Alphabet arm, Calico, focuses on aging and age-related diseases such as neurodegenerative disorders and cancer.
Within Google itself, the health efforts are wide-ranging. In 2021 it acquired fitness tracking company Fitbit for $2.1 billion, a bid to challenge Apple and others in the wearable health-tech space. Another group works on clinical tools, like software that can interpret retinal scans to hunt for signs of blindness. Its medical research includes work on sequencing and interpreting genomes using AI. Through the cloud, Google offers enterprise health-care companies a platform to develop their own AI apps, like chatbots and other software that can help search through health records, using the company’s flagship AI model, Gemini.
The company said there are also several efforts across Google that already utilize AI to democratize access to health care, such as AI-powered screenings for lung and breast cancer; work on expanding access to information on health conditions via products like Google search and its visual search app Lens; and scientific research through efforts like AlphaFold, a landmark AI tool for predicting protein structures.
But the company’s generally scattered approach has hindered Alphabet’s ability to make meaningful progress. “Google has put their hats in many different areas, but they’ve struggled to figure out how they can best help the industry,” said Leventhal, the Emarketer analyst.
Google said its work in health care is meant to support different parts of the overall health ecosystem and that individual groups within Alphabet pursue their own set of objectives and strategies in health. “Our North Star is to help people everywhere live longer, healthier lives, to contribute meaningfully to scientific research and to improve health-care delivery for all,” a company spokesperson said.
In recent years, in order to test its technology and explore market opportunities for it, Google has turned to partnership deals like the one with Nashville-based HCA Healthcare, the largest health system in the country by bed size, with 182 hospitals across the US. At Google’s health event in March, leaders from both companies stood side by side to present insights from HCA’s early tests of Google’s AI-powered nurse-handoff note. According to a company spokesperson, across all HCA Healthcare hospitals, more than 60,000 handoffs occur each day. In a recent internal survey, 89% of nurses who are testing the tech found it was “somewhat” or “very” helpful, HCA added.
“Some of these dreams that we’ve placed in health-care technologies and what they could do for us, I’m seeing them come true today,” said Canaan Stage, a leader of clinical operations at HCA Healthcare, during the presentation. “I’m giddy. That’s maybe the best word I can say.”
Another tool powered by Google’s AI that is available across four HCA Healthcare emergency room facilities is an app designed by Augmedix that can automatically transcribe patient visits and turn them into digestible reports. This could save doctors a lot of time; according to a recent study published in JAMA Internal Medicine, they spend over eight hours a week writing up their notes.
“A lot of the energy now is around generative AI trying to help doctors and nurses and care systems deal with documentation, information management, a lot of this work that many people in health care view as an overwhelming wall of drudgery,” said Corrado, the Google AI research director.
But even for this seemingly lower-risk use case for generative AI, the technology can’t be deployed without backup. Michael Schlosser, HCA Healthcare’s senior vice president of care transformation and innovation, said having humans in the loop is a must because there’s still a possibility that AI tools will miss relevant information, or even generate made-up symptoms or past patient history — what’s known in AI lingo as a hallucination. “If the AI missed something, and then the next caregiver was unaware of that, that kind of thing can become dangerous,” said Schlosser.
When HCA doctors use the technology, he said, each physician is responsible for checking the AI-generated notes carefully, before inputting the data into a patient’s electronic health record. A representative of HCA added that doctors and nurses checking the output of the AI technology results in feedback that is then used to further train the AI. “We see improvements in the technology week over week using this technique,” the representative, Harlow Sumerford, said in an email. HCA also said having humans in the loop is critical and doctors would always be responsible for approving and confirming the accuracy of the information generated by AI, regardless of how advanced the tech becomes.
Sarah West, co-executive director of the AI Now Institute, a policy research center, said Google has marketed AI tools for use in health-care settings even as the industry still lacks sufficient standards for validating the systems to make sure they don’t make critical mistakes. “Google is under a lot of pressure right now to find its place in this ecosystem,” West said in an interview. “The pressure to turn a profit is very high.”
Another important question is whether the notoriously sclerotic health industry is willing to embrace AI on a wide scale. Decades after the internet became widely accessible, some 12% of US physicians still haven’t made the transition to electronic health records. Many hospitals and doctors’ offices still rely on the fax, which can communicate with other offices no matter what record-keeping software they use. And it’s harder to hack.
‘Two different languages’
Google’s first health bet, the online health record, failed because the company hadn’t really made connections with the industry it was trying to change — the doctors, nurses and hospitals actually delivering care to patients. Patients had to track down often incomplete records from different doctors and continually update them online, which made the service too much of a hassle. Microsoft and others failed at similar attempts.
After that first flop, Page and Brin were only more motivated. In 2013, they recruited Andy Conrad, a geneticist and former top scientist at Lab Corp., to join Google’s secretive life-sciences unit, which would eventually become Verily. Google let researchers and engineers pursue audacious goals like glucose-monitoring contact lenses for diabetics, a cancer-detecting pill and an enormous personal health study called Project Baseline, which aimed to collect longitudinal health data from around 10,000 participants and pursue new research on how to decrease rates of diabetes and heart disease.
The projects grabbed headlines. But they also suffered from a clash of cultures behind the scenes, according to Obi Felten, a former Google executive who worked on the company’s early health efforts and left in 2021. The technologists wanted to move fast and see results; the doctors and scientists wanted to move slowly and methodically. “I literally felt sometimes like it was two different languages,” said Felten.
Across the company, Google’s health projects were a hodgepodge of efforts, some of which sprung up because an executive was passionate about the area, rather than because they fit an overarching strategy, according to two former employees. By 2018, Google Health had hundreds of employees and Verily over a thousand, but their portfolios were a haphazard collection of moonshots, many of which failed to get off the ground.
Google needed a more disciplined strategy. In 2018, it tapped a new executive, the former head of the health network Geisinger, David Feinberg, to come up with one. But months after Feinberg arrived, he faced a controversial news cycle. Google had been working on a search tool for the US Catholic health-care network Ascension. The Wall Street Journal detailed how some Google employees had gotten access to sensitive health data, prompting a federal inquiry that the company had to answer for. Google noted that the probe didn’t go anywhere, and said it had followed federal privacy rules throughout the project. Eventually the company still wound down the effort, though it kept the technology, integrating it into some tools it sells to health enterprise companies on its cloud platform.
When Covid swept through the world, it looked like Alphabet finally had an opportunity to make a large-scale impact in health care. In March 2020, then-President Donald Trump surprised Americans — and Google — by promising nationwide testing from the tech company. Verily launched a testing project shortly thereafter, and California eventually spent $62.5 million contracting with the Alphabet unit to run testing sites in 28 counties. But seven months later, two of the state’s most populous counties cut ties with Verily, citing privacy concerns. Within a year, California ended the contract completely. Verily’s tests ended up accounting for less than 1% of all Covid tests in 2020.
A Google spokesperson said the company’s Covid efforts extended beyond its virus-testing contract with California. Verily partnered with RiteAid, the US Department of Health and Human Services, the Centers for Disease Control and Prevention, and others, the spokesperson said, and provided software that enabled over 6 million people across 16 states to get tested. As testing shifted from laboratory-based PCR testing to at-home rapid testing, Verily switched focus to using wastewater to estimate population-level pathogen incidence, the spokesperson added.
Not long after, Google changed strategy again. The company had built up its health division during the pandemic, and now it decided to disband it. Employees working on health products were redistributed to other units at the company. As part of the reorganization, Feinberg also left Google.
Though key divisions at Google have employees working on health tools within their own units, the only dedicated health team at the company is now led by DeSalvo, a physician who was appointed chief health officer in 2019. She spearheads a small team of clinicians that leads health strategy for Google and counsels other parts of Google on health-related issues.
In an interview, DeSalvo said that because people naturally ask Google questions about their health, it’s the company’s responsibility to meet them with high-quality information. Every time someone searches Google to identify their weird skin rash or figure out whether they just have the flu or some sort of rare and terrible cancer, they are involving Google in their health.“The reality is, even if we didn’t want to think about health, we’re getting asked about it all the time,” DeSalvo said.
Missing details
One morning in February at the University of Central Florida’s Lake Nona hospital, which is part of HCA Healthcare, emergency room physician Alex Stinard walked into the room of a 56-year-old patient who complained of lower back and leg pain. Stinard fired up the Google AI-powered medical transcription app, Augmedix Go, then placed his smartphone at the edge of the bed and hit record. He talked with the patient for seven minutes, asking her where it hurt, and placing his hands on different parts of her body.
After the consultation, he thumbed through the Augmedix Go summary of the visit on his phone. It generated text more quickly than Stinard’s usual note-taking process, and was more organized, too.
But like Loomis, Stinard noted that the AI had missed some things. It didn’t pick up on the fact that she had no tenderness to her middle back and neck, and that she had reported normal leg strength, vital details for an accurate diagnosis.
A new challenge for Google in the era of generative AI is applying the technology in the service of delivering more high-quality information to people. The results so far have been mixed. In May, Google announced AI Overviews for search written by generative AI, the same underlying technology it uses for its health-care apps. The launch didn’t go well. Some AI Overviews offered embarrassing suggestions, advising people to eat rocks or to put glue on pizza, for example. In response, Google implemented new guardrails, in part to prevent the feature from inadvertently presenting satirical content as fact.
Google’s generative AI tools in health are trained to be far more focused on delivering medically accurate answers. According to a Google spokesperson, “on health queries, we always had strong guardrails in place for the quality of content we display and since launch, we’ve added additional triggering refinements to enhance our quality protections.” But the episode shows how easily AI — without the help of human judgment — can veer down the wrong path. That’s why many checks and balances are still in place for Alphabet’s AI health tech. For instance, with Augmedix Go, practitioners like Stinard must double-check the AI’s work.
“Overall, I think Google has struggled to gain trust amongst consumers and also physicians, that they should have access to health data and that they will put the appropriate safeguards in place,” said Leventhal, the health-care analyst.
“We understand that health is human,” Google said in a statement. “AI is not a replacement for doctors and nurses, for human judgment, the ability to understand context, the emotional connection established at the bedside or understanding the challenges patients face in their zip codes.”
But for Alphabet to truly transform health care — for its AI to do more than play secretary to practitioners — the company will have to gain more trust.
Rob Bart, chief medical information officer at the University of Pittsburgh Medical Center, said the conversation on generative AI’s role in medicine falls into two buckets. “There’s clinical operational efficiency and effectiveness. That’s the place most organizations are currently comfortable,” Bart said. “Then there’s clinical diagnostic medical decision-making. We’re not so comfortable with that yet.”
--With assistance from Julia Love and Ilena Peng.
©2024 Bloomberg L.P.