Why $4.6 Billion Health Records Giant Epic Is Betting Big On Generative AI

Led by billionaire founder and CEO Judy Faulkner, Epic is building a suite of 60 tools it hopes to sell to thousands of U.S. hospitals.

Medical diagnosis and procedure codes are so numerous and varied that Debbie Beall, manager of coding at Houston Methodist in Texas, needs a 49-person team to translate the medical notes written by the system’s 1,600 clinicians into the codes needed to bill insurers.

There is a medical code for every imaginable scenario – from “burn due to water-skis on fire” to “spacecraft collision injuring occupant” — and their specificity determines how much the insurance companies pay. Each team member processes anywhere from 70 to 250 claims per day, depending on the complexity, she said. That’s why Beall is so excited about the possibility of using artificial intelligence to speed up the job.

“There’s no way I’m ever going to replace coders completely with an AI system,” Beall told Forbes. But for run-of-the-mill procedures performed multiple times a day in a hospital, like X-rays and EKGs? “Yes, an AI engine can do that.”

Beall was one of the first dozen or so people to test a prototype of an AI-powered medical coding tool from electronic health records giant Epic Systems, which had $4.6 billion in revenue in 2022. Based on GPT-4, the large language model that powers the viral chatbot ChatGPT, Epic’s coding assistant prototype ingests and summarizes clinician notes and then tees up the “most likely” diagnosis codes and procedures codes, along with suggestions of “other potential codes,” according to mock ups viewed by Forbes that did not include real patient information.

To address GPT-4’s tendency to hallucinate, or make up information that might sound legit but is entirely false, the tool also links back to the information that led to a certain code. For example, if a surgeon writes “liver failure” in the original note, the tool shows that it used these words as the basis for the suggested code “K72.90 hepatic failure.”

Epic told Forbes it expects to release the tool in May, but that version will only be able to suggest codes and requires careful human oversight. “We have not moved to fully automate – at least not yet,” said Ryan Krause, a vice president at Epic who oversees revenue-related software. Ultimately, it will be up to individual hospital customers to decide when and what codes they might be comfortable doing without human intervention in the future. “But it’s really exciting to see that now there is a technology that allows that easy summarization and makes the coders much faster.”

The U.S. spends $4.5 trillion each year on healthcare, so it makes sense that medical coding is where AI will have the biggest impact in healthcare in the near-term, Zak Kohane, the editor-in-chief of NEJM AI and chair of biomedical informatics at Harvard Medical School told Forbes. “It’s billions of dollars at stake,” he said. “It has relatively low patient risk.”

The coding tool is one of more than 60 generative AI applications that the Verona, Wisconsin-based company is developing. “We’re really trying to decrease the burden on the clinicians,” Epic founder and CEO Judy Faulkner told the Forbes Health Summit in December 2023. One example she gave was a new generative AI tool to help doctors draft responses to patient portal messages, currently being used by around 90 customers. “So far our clinicians and our patients have found that they liked the AI’s response better than the human beings’ response, because the AI was more empathetic,” Faulkner said.

Epic is moving quickly so as not to lose its incumbent advantage to the dozens of healthcare startups building a range of workflow tools from summarizing medical records to coding. But because its software is already used by more than 2,700 U.S. hospitals out of around 6,100 total, Epic has a distinct advantage.

“Part of the problem and the strength of Epic is it has captured a large part of the C-suite of these hospitals,” said Kohane. “It’s almost a Stockholm Syndrome,” he said, referring to the fact that hospital CEOs spend tens of millions – if not hundreds of millions – of dollars to install Epic, meaning they are more likely to stick with their original investment.

Whenever Epic signs a new customer, a baroque wedding march plays across the company’s 1,100 acre Verona, Wisconsin headquarters. “It’s a very long relationship for many of our customers,” Faulkner told Forbes in a 2021 cover story.

“They have a better shot than most at getting that sale,” said Kohane, but it’s also not a foregone conclusion that Epic will win when it comes to generative AI, which will ultimately come down to performance and price.

At the HIMSS healthtech conference in Orlando last week, Epic had a poster advertising existing and upcoming generative AI tools, including drafting patient discharge notes, writing appeal letters for insurance company denials and automating scheduling follow-up with patients. “We recognize that there is both a staffing shortage across healthcare, as well as real challenges from a financial perspective,” said Seth Hain, senior vice president of research and development at Epic. “Generative AI can help alleviate some of those challenges when we focus it on having the most impact directly in workflow.”

Epic has directly partnered with two companies – Microsoft and startup Abridge – to co-develop automated medical scribes that record patient exams and help draft the doctor notes in the electronic health record. Epic said 73 customers are currently using the so-called ambient AI tools. It has a longstanding partnership with Microsoft, which invested $10 billion in OpenAI, the startup that developed the GPT models.

But it’s unclear how those partners play into Epic’s overall generative AI strategy since it’s also building in-house tools. When asked to explain Epic’s approach, Hain wouldn’t reveal specifics: “Our focus is on providing an efficient and accurate experience in workflow, and through our design, evaluation, and development process we determine which approach is best to pursue.”

Any time an Epic function is using generative AI, there is a clear indication to the user with the words “AI” and a symbol next to whatever has been AI-generated, said Hain. For users, there are citations so they can quickly review the reference in the patient record. On the backend, Hain said Epic’s engineers have a system for evaluating and auditing any use of generative AI to be able to “monitor trust and safety at a system-wide scale.”

While the current applications are primarily being built with GPT-4, Hain said Epic is building all of these tools “to be able to support multiple different types of models over time.” That means they could potentially swap in other competing large language models, including open source or commercial models, as well as small language models. Like many organizations, Epic is experimenting with different commercial and open source large language models to figure out which ones perform the best for a given task and what kind of additional fine-tuning or prompting is needed.

And there is fine tuning yet to be done. Beall, who helped test Epic’s prototype, said that while the coding tool correctly coded a simple hernia repair surgery, it didn’t generate the right codes for the general office visit. “The computer got the wrong level of service, because the doctor did not document the severity,” said Beall.

This is why human calibration is so important. Epic’s prototype didn’t register that there was missing information, something Beall would have noted immediately. Beall gave the example of a liver transplant patient who is running a fever. “Anytime a transplant patient gets a fever, that means there’s something majorly wrong,” she said. This would mean this patient was “critical,” which is the highest level of severity possible, but for the system to recognize this, the doctor would have to write “critical” in the note. In this case, the tool didn’t understand the connection between fever and the severity of the patient’s condition.

Part of rolling out the coding assistant at Houston Methodist will involve training alongside human coders who will give real-time feedback, she said. Her hope is that ultimately the tool can autosuggest codes for minor procedures and simple surgeries, like knee replacements, hernia repairs and tonsillectomies. That way her team of coders can focus on more challenging cases – transplant patients, genetic abnormalities, cutting edge surgeries. “It does take a human touch,” said Beall.

While Epic has so far focused on using generative AI in back office functions, it has also been working on a patient-facing application that wouldn’t require human review. Krause told Forbes a tool that would help explain the patient’s bill, including their deductible and outstanding balance, could be rolled out by November. “We feel like that’s a fairly benign place to start. It’s not about healthcare at that point, but it’s really about their billing,” he said. “That’s not going to harm a patient in any way.”