Along with many other large health systems, HCA Healthcare is exploring ways to reduce its clinicians’ administrative burden through the use of generative AI.

The health system, which is the largest for-profit hospital chain in the country, began a partnership with Google Cloud in 2021, initially focused on data privacy and security. This year, the two organizations have expanded their collaboration by exploring ways to integrate Google’s generative AI into HCA’s workflows. The most developed project under this initiative is a pilot that began in February in which HCA’s emergency department physicians are testing a voice-enabled medical dictation tool.

About 75 doctors across four of HCA’s emergency departments are using Augmedix’s app on their phones to produce clinical notes from the conversations they have with patients. Augmedix’s app listens to the interaction and then combines its natural language processing capabilities with Google’s generative AI models to convert the data into medical notes that physicians review and finalize before transferring to the EHR.

Augmedix is a medical scribing company that went public in 2021. HCA has a small ownership stake in the company.

Michael Schlosser, HCA’s senior vice president of care transformation and innovation, noted in an interview that the hospital industry has been trying to use natural language processing models to alleviate clinician’s documentation-related burnout for years now. 

“It has been somewhat slow going because you’re basically looking at specific NLP models to be able to pull out specific chunks of that data and then turn it into structured documentation — which is computationally very hard to do. Generative AI has really massively accelerated this work — in particular with the parts of the notes that are more narrative, like the history of present illness or the medical decision making in which the physician is telling the patient the story of how they’re thinking about the diagnosis and treatment,” Schlosser explained.

By adding Google’s generative AI models to Augmedix’s app, the tool can do a much better job of understanding narrative speaking and breaking it down into a structured clinical note, he said.

When HCA was introducing the tool to its emergency department physicians, the health system recommended that they verbalize their decision making process more when they’re speaking with patients, Schlosser pointed out. Not only does this give the AI models more to work with so they can produce a more detailed clinical note, it is also a good practice to adopt because it facilitates better communication with patients and gives them a better understanding of what’s going on with their health, he declared.

The version of the tool that HCA’s physicians are currently using still has a human in the loop as a medical scribe who assists the AI in completing the note, Schlosser noted. He added that HCA has a data sharing agreement with Augmedix in which the health system is sharing data from the emergency room encounters included in its pilot. This way, Augmedix and Google can use that data to train their AI models to be more accurate.

“We think we’ll pass 50% automation — where the AI is creating 50% of the note completely by itself without the need of human help — by the end of this year, and then we’ll continue to work on it. We want to get it even better than that. We think somewhere in the 75-80% range is probably where it becomes really functional for the ER doctors — where they have to do very little work after the AI has completed its job to have a really high quality, complete note,” Schlosser said.

To measure the success of its pilot, HCA is collecting feedback from its physicians about how well the tool fits into their workflow, its ease of use, and whether or not it is saving them time. The health system is also measuring the percentage of emergency department notes that are finalized within 24 hours of the patient being seen — the goal is for that measure to be 100%, Schlosser declared.

In addition to this pilot, HCA plans to build a system using one of Google’s large language models to automatically generate reports for nurses when they hand off their patients to a different nurse. The health system is also looking to adopt Med-PaLM 2 — Google’s generative AI tool designed specifically for healthcare providers — in the future. The tool is currently being tested by Mayo Clinic and other health systems.

Healthcare stakeholders know that AI systems have promised to revolutionize healthcare in the past and failed, such as IBM Watson Health. However, Schlosser believes that Google’s AI can be different from the failed projects that have come before it.

“A lot of the early work in AI focused on training systems to do tasks like interpret medical images, discover clinical patterns in large datasets to find new clinical insights or treatment recommendations, or try and exceed human decision making for complex conditions. While these are worthy pursuits, they did not focus on addressing the friction points of what those delivering care day in and day out would see as the greatest opportunities in healthcare delivery,” he explained.

HCA’s projects related to Google’s generative AI center on use cases focused on improving the process of care delivery, and these efforts involve solutions that are clinician-driven, Schlosser said. By designing their projects this way, HCA can increase the likelihood of broad adoption, he noted.

Photo: Elena Lukyanova, Getty Images