As AI chatbot trend grows, health systems choose to shape it

Assessing the types of patient-facing AI chatbots

It'd be a tall order to describe every type of AI chatbot that health systems have deployed. There's a wide spectrum in how developers and their health system partners are deploying generative AI, and the guardrails they choose to put in place.

But, generally speaking, health systems' current chatbot designs include the following:

• Home-grown systems that reflect commercial products, essentially building a version of ChatGPT for the health system.
• Generative AI tools that help patients more deeply engage in secure direct messaging, such as tools that help patients craft patient portal messages, translate medical jargon or generate plain-language discharge summaries.
• Off-the-shelf products embedded in the health system's existing software suite, such as an AI bot built by an EHR vendor.

This list is by no means exhaustive, but the varied approaches to conversational AI are noteworthy. Health systems face evolving patient and organizational needs, ranging from high costs to challenging patient navigation.

Many organizations agree that AI chatbots could be a key tool in those efforts. But the question is, how should they implement them? Health systems nationwide have taken different tacks, marking a smattering of AI confidence versus AI caution.

Hartford Healthcare goes all-in on PatientGPT

Hartford Healthcare emerged as one of the industry's earliest leaders in how health systems could adapt to emerging commercial AI products. The system launched PatientGPT in March in partnership with K Health, a clinical AI company.

"Why go outside of your trusted healthcare system and ask a question to which you may get an answer that may not lead to an endpoint that's valuable?" posited Padmanabhan Premkumar, M.D., president of Hartford Healthcare Medical Group. "You may just have more questions after that."

That's because most of the commercial products on the market -- ChatGPT, Gemini or Claude, for example -- aren't linked back into the patient's medical record. Even ChatGPT Health or Claude for Healthcare, which both say they can take in users' medical records to interpret data and provide contextualized insight, can't plug into the health system.

That's what makes PatientGPT different, Premkumar said.

The technology lives right inside Hartford Healthcare's patient portal, creating an access point for patients who need answers to their medical questions and to connect to care when necessary. For example, if the patient chats with PatientGPT about side effects to their new medication, the bot can determine what is or isn't normal and book an appointment in case their medication needs titrating.

PatientGPT represents Hartford Healthcare's confident step into offering patient-facing AI. After all, these tools are the way of the future, especially in an industry that so often leaves patients unsupported, Premkumar said.

"It's fundamentally changing healthcare. When you actually think about healthcare today, it's complicated, it's convoluted," he explained. "PatientGPT allows for patients to go in and ask questions, but more importantly, for us to guide them along the patient care journey because we know who they are."

But because the tool autonomously answers patients' medical questions, it operates much like the chatbots offered on the commercial market that have raised the eyebrows of some patient safety advocates.

Being tied to a health system and tethered within the EHR and patient portal provides some additional safeguards that might make the tool more effective. If nothing else, it adds a layer of HIPAA compliance.

But more AI-skeptical health systems still might not be willing to make the jump into a tool that mirrors PatientGPT. Taking a more iterative approach might be a better fit for health systems that prefer to assess AI progress.

VUMC's AI supports portal messaging

Like Hartford Healthcare, Vanderbilt University Medical Center quickly adapted to the industry's rapid embrace of AI models tailored for patient use.

But VUMC's technology doesn't take the same approach.

Rather than building an AI model that could chat with patients, analyze patient medical records, dispense medical advice and book any necessary treatment, VUMC's tool leverages secure direct messaging within the patient portal.

The tool sits within VUMC's branded version of MyChart and works as a virtual assistant to help patients craft a better message to their care team. The bot probes the patient about symptoms or treatment updates and helps them consider exactly what information they want to share with their provider. The bot can then draft a message for patients to send via the patient portal, but it won't give medical advice.

The tool is completely opt-in, and patients can stop answering the bot's questions at any time or decline the draft it produced.

According to Chetan Aher, M.D., associate chief medical officer for Adult Ambulatory Clinics at VUMC, the health system wanted to preserve the direct connection patients have with their clinicians.

"We still believe firmly that when a patient logs into our portal and wants to send a message to their care team, that's really the outcome that they want -- an answer from their care team," Aher said. "They don't want to get on there and be served up some kind of answer [from a bot]."

VUMC's tool is designed to help patients get the most out of patient portal messaging, helping patients to ask more direct questions that help providers get to the heart of their medical concerns. This was the best place to start as VUMC considered how AI would transform patient engagement, Aher suggested. There's a lot to consider when implementing AI for any use case, and this first foray gives VUMC the opportunity to validate the system over time.

"We're talking about an industry that still relies heavily on fax machines," Aher stressed. "It's a difficult space to integrate really new technologies into. There's a lot of regulation, there's a lot of hesitation and there's a lot of concern about making sure that the quality of care is not affected in any negative way by something that is relatively new or relatively untested."

There may come a day when VUMC's patients want to interface directly with an AI chatbot, but until then, the health system plans to iterate.

How rural hospitals jump on the AI trend

Health systems taking the plunge into the AI waters need to consider cost. AI system development costs money, and designing a custom system won't be attainable for every organization.

"We are not a Cleveland Clinic and we are not a Mayo Clinic," said Muhammad Siddiqui, M.D., CIO of the Indianna-based Reid Health. "But our mission is just as real, maybe more urgent, because when our patients need care, they do not have a backup option."

Reid Health is a rural health system that serves Eastern Indiana and Western Ohio. The organization's patient population doesn't have many care access options, sometimes having to travel long distances to see a provider.

AI chatbots have a lot of potential to help patients overcome rural healthcare's biggest challenges.

But Reid Health had a lot of the same concerns that VUMC and even Hartford Healthcare had. Siddiqui and his team wanted to make sure their patients had access to validated AI-generated answers that came from them, not an external tool like ChatGPT or Gemini.

Unlike its larger peers, however, Reid Health didn't have the personnel or the funds for a custom-built solution. Instead, the health system explored the tools at its disposal right within its Epic EHR suite.

Earlier this year, Epic launched Ask Emmie, which leverages generative AI to answer patients' health-related questions within the context of the patient portal and health system workflows.

That's a key AI safeguard, Siddiqui said. Ask Emmie isn't just drawing on general health information.

"Emmie knows that patient very well. Emmie knows the conversation between the patient and the doctor and Emmie knows the test results from the last 12 months," he stated. "So, Emmie won't give an answer based on only one level of test results that a patient has shared with it."

Rather, it pulls from the entire patient's health history.

And, importantly, it comes right off the shelf from Epic, providing a pathway for the rural health system to jump on the emerging patient engagement trend in a more cost-effective way.

Of course, Siddiqui, his team and various committees within Reid Health validated the tool before turning it on for all patient users. But still, the health system couldn't be caught dragging its feet on AI. Even rural organizations need to stay abreast of the leading patient engagement trends.

AI keeps patients connected to the health system

Regardless of how health systems chose to implement AI chatbots, the motivation was the same: keep the patient connected to the health system, not a technology company.

"Patients are spending significant time and effort in outside tools querying their medical records and understanding what they might have and asking all sorts of questions," said Ran Shaul, co-founder and chief product officer of K Health, who worked alongside Premkumar on PatientGPT. "We had to ask ourselves: can we do it in a way that we are extending the patient care team relationship as opposed to a standalone AI tool?"

For VUMC, it was imperative for the health system itself to shape patients' relationship with AI, if for no other reason than to promote patient safety. There's risk in leaving patients to explore commercial AI tools, like ChatGPT, Claude or Gemini, and not knowing exactly what they'll tell the patient.

Siddiqui had a similar realization. As more of Reid Health's patients turned to AI, there was a big opportunity to reimagine patient engagement. "We really have two choices: Either we dismiss it, or we shape it," he said.

"We choose to shape it."

Sara Heath is an executive editor at Xtelligent Healthcare Media, where she covers patient engagement, healthcare policy and health IT.