Google Launches Medical AI Language Model to Optimize Healthcare Handoff Process

2023-08-31

Whenever a nurse finishes their shift at a hospital, they must quickly inform the incoming nurse about all the patients they are responsible for. "Handoff" can take various forms, including conversation, handwritten notes, and electronic medical records. "'Handoff' is a risky part of the healthcare journey because we are transferring information from one healthcare provider to another," said Michael Schlotter, Senior Vice President of Nursing Transformation and Innovation at HCA Healthcare. "We must ensure that it is done accurately and without any omissions."

HCA, based in Nashville, is one of the largest healthcare systems in the country, with 180 hospitals and approximately 37 million patients. Schlotter and his team at the company believe that applying this information transfer to generative artificial intelligence could be a great opportunity. Large language models have shown excellent performance in summarizing and organizing data. However, when HCA searched for potential vendors, Schlotter said they couldn't find any companies building solutions for this handoff problem.

HCA has partnered with Google Cloud, so they turned to Google's software suite Vertex AI, which helps customers build and deploy machine learning models. Google provides its own base model called PaLM, but the platform is model-agnostic, meaning customers can also use and build models from OpenAI's GPT-4, Facebook's Llama, Amazon's Titan, or others.

To attract more healthcare customers, Google has also developed a large language model specifically for the healthcare field. The company announced on Tuesday that it will release the latest version, Med-PaLM 2, to more customers in September. In addition to HCA, several healthcare customers including pharmaceutical giant Bayer, electronic health record company Meditech, and digital health startups Infinitus Systems and Huma have been granted early access. This push in the healthcare field comes as Microsoft and Amazon are also advancing their own AI-based initiatives, and it is currently unclear which company will take the lead.

"We are still at the forefront of the healthcare AI field," commented Gartner analyst Chirag Dekate on the prospects of the healthcare AI field.

In 2021, Google disbanded its standalone Google Health division but stated that it would continue its health-related efforts across the company. The company's latest AI solutions in the industry aim to address specific problems. For example, last year, Google released AI tools to assist healthcare organizations in reading, storing, and annotating X-rays, MRIs, and other medical images. Earlier this year, the company launched AI tools to help health insurance companies expedite the prior authorization process.

Corrado said the focus on these use cases is due to the nature of AI technology itself. He said that while large language models have received a lot of attention, it is "naive" to expect them to "perfectly complete any task immediately." He added, "In fact, these systems always need specific use cases specified."

Google has been playing catch-up with OpenAI in the field of large language models, with the latter being the startup behind the viral chatbot ChatGPT and having received a $1 billion investment from Microsoft. In 2022, Microsoft acquired Nuance Communications for $18.8 billion, making its foothold in selling new AI products to hospital customers even more important, as Nuance's medical transcription software is used by 550,000 doctors. "Nuance has a huge influence in the healthcare field," said CB Insights analyst Alex Lennox-Miller, positioning Microsoft well in using its generative AI software for management tasks in the industry.

Prior to large language models becoming mainstream, Amazon, Microsoft, and Google have been competing for cloud customers. According to research firm Gartner, Amazon's cloud revenue reached $48.1 billion in 2022, accounting for about 40% of the market share. Microsoft follows closely with a 21.5% market share, while Google ranks fourth with $9 billion in cloud revenue and a 7.5% market share, just behind Alibaba Group.

Dekate said it is not surprising that these cloud companies are now focusing on specific healthcare customers, as the healthcare industry is complex and heavily regulated. He said that if you can demonstrate use cases in more complex environments like healthcare or financial services, it will show other customers that generative AI can be widely applied.

But no one can do that yet. All the cloud companies offer building blocks to customers, Dekate said. In other words, they provide numerous ways to leverage their AI platforms to build custom applications. However, these customers want fully formed solutions.

"Amazon, Google, and Microsoft are vying for the high ground in the generative AI economy," Dekate said. "But none of them can tell a good story in its entirety."

Due to the strict regulations and serious consequences in the healthcare industry, use cases of generative AI need to start small. For HCA, this means that currently only one hospital, UCF Lake Nona Hospital, is testing the handoff tool as a proof of concept. Schlotter said AI retrieves patient data from the past 12 hours, including lab results, medications, significant events, and generates a handoff summary that also includes recommendations for the incoming nurse to consider in the next 12 hours.

Although using Google's Vertex AI software, HCA has been experimenting with different base models, including PaLM and Med-PaLM. "We are currently conducting a lot of positive conflict testing to determine where the general models are better, and medical training models play a more important role in providing more accurate and better results," Schlotter said. "I think both aspects will play an important role in the future we are trying to create."

Using multiple models to solve complex problems and calling it "comprehensive" artificial intelligence is an interesting challenge for cloud providers, Dekate said. He said they offer their own in-house models while also collaborating with other companies to provide a "promise of choice." Dekate expects that we will start to see Google, Microsoft, and Amazon offering services to help customers evaluate different models. Schlotter said that so far, HCA has taken a manual approach to the way they evaluate models, by having doctors and nurses compare the model's output with the practices of the human team.

Corrado said these concerns are why Google is experimenting with professional LLM models trained on narrow datasets. He said that without adjusting the models for specific use cases like healthcare, there is a "risk of having a multi-purpose tool that is neither the best knife nor the best screwdriver. Nor is it the best toothpick. We believe that in these high-value environments, it is best to have domain adaptation, understanding of use cases, and the same rigorous quality assessment and version control as real products."

Another challenge for most large language models is that they are not continuously learning. They usually have a cutoff date for training data. For example, the free version of ChatGPT was trained on data up until September 2021. But as medical knowledge advances, doctors using these tools need to know how fresh the data they are using is. Corrado said Google is still deciding what the cutoff date will be but will communicate it to customers. "We can't rely on these systems to know all the knowledge of medical practice," Corrado said.

In the future hospital, Schlotter envisions an "AI assistant for the nursing team" that he believes will have "amazing capabilities" in alleviating their daily workload. HCA is also working with Google and publicly traded environmental AI company Augmedix to automate medical notes in the emergency room. Schlotter said about 75 doctors in HCA's four hospitals are using this technology. He said the "ultimate goal" for doctors is that they can focus on providing care to patients, and "documentation will be done by itself." They started in the emergency room because it is one of the most complex environments where the technology can prove its effectiveness.

When using Augmedix's tool, doctors directly ask patients for consent to record the examination and use AI tools for documentation. As for the nurse handoff tool unrelated to patients, it falls under HCA's broader privacy authorization for using patient data for research and process improvement, Schlotter said. HCA is also exploring the use of generative AI for emergency room discharge summaries and handoffs from the emergency room to the inpatient unit. Schlotter said as HCA considers expanding the use of AI for management purposes, the company will have to consider "the appropriate way to make all patients aware that AI is part of the care process."

Using AI in the healthcare field has raised concerns about consent and privacy. Google previously partnered with hospital system Ascension to use AI to analyze millions of medical records, sparking widespread controversy. In 2019, reports about the company's "Project Nightingale" raised concerns about data privacy and security. Both Google and Ascension stated that the work complied with federal patient privacy laws.

Regarding PaLM and Med-PaLM, Google stated that these models were not trained on patient data from HCA or any other customers. "HCA's data is HCA's data, not anyone else's," said Thomas Kurian, CEO of Google Cloud, to Forbes. "You can think of our storage in the cloud as a place where versions of the models they use are trained, not shared with anyone else. None of that data is used to improve our base models."

Despite the challenges in terms of technological capabilities, privacy, and data, Schlotter is optimistic about building tools based on this technology becoming part of a doctor's standard toolkit. He said HCA has taken a slow approach to mitigate some of the burdens of daily work because he believes once doctors embrace AI, they will be able to guide how it can be best used for more complex applications.

"I hope that clinicians can fully embrace AI as a partner in their lives that can make their lives easier, and then move into some of the more controversial areas," he said.

RECENT AI NEWS

RECENT AI TOOLS