With so many patient messages coming into a portal, it’s possible to miss important details, especially if a message were long and convoluted. But by using AI as a tool, physicians get an assist in message analysis.
"AI can help humans read messages more closely by highlighting essential pieces of information that are important for the doctor to review,” said Jason Hill, MD, MS, clinical innovation officer for Ochsner Health in New Orleans.
For example, “sometimes a message will be paragraphs long, and buried in the middle might be concerning information about a patient experiencing concerning symptoms. AI will flag those vital pieces of information,” said Dr. Hill.
Ochsner Health is a member of the AMA Health System Program, which provides enterprise solutions to equip leadership, physicians and care teams with resources to help drive the future of medicine. Ochsner Health collaborated with EHR vendor Epic to launch Microsoft Azure’s OpenAI's GPT-4 language model last fall.
“Epic has entered a partnership with Microsoft in order to leverage the development of generative AI in health care,” said Dr. Hill. “We are very lucky to be chosen as one of the pilot sites for this partnership.”
AI is often called artificial intelligence, but the AMA House of Delegates uses the term augmented intelligence to focus on AI’s assistive role, emphasizing that its design enhances human intelligence rather than replaces it. The AMA has developed new advocacy principles (PDF) that address the development, deployment and use of health care AI.
In an interview with the AMA, Dr. Hill explained how AI can help physicians review and process patient messages, and how it has improved key performance metrics at Ochsner Health.
AI’s not squeezed for time
Time-squeezed physicians may not always “write the most in depth, emotive messages back to our patients. And sometimes we're a little terse when we don’t mean to be,” noted Dr. Hill.
For generative AI that moves at computer speed, time is not an issue.
GPT-4 and other large language models unlock language and allow AI to understand that language. In analyzing patient messages, it understands what the patient is asking, and it can assist a physician in composition. It can also leverage the patient’s chart in the EHR.
AI has the capacity to quickly process a patient’s entire record—including their medications, their health conditions and which physicians they’re seeing. This data helps inform auto-generated response prompts, which assists the doctor in two ways.
“One, it helps us to make sure we're drafting a good message to our patient. And two, it allows us to access the entire health record,” which can be very difficult to take in wholly because of how information is scattered over reams of data and notes, explained Dr. Hill.
It can manage high data volume
If a patient is on a medication that’s causing nausea, for example, GPT-4 can look this up and can flag a prescribed medicine that might have an interaction with another the patient is already taking.
Sorting through and making sense of those reams of health care data and finding the important bits “helps me as a doctor,” said Dr. Hill. “Instead of my having to run through the chart and look for pockets of information, it is all in front of me and I can make some decisions about it.”
Additionally, if a patient sends a patient portal message that is extremely lengthy and detailed, AI can help organize and manage the issues.
All patient messages and AI responses live within the secure protected environment of the health record, Dr. Hill noted, explaining that “the data doesn't travel anywhere and it’s in one space.”
Someone always reviews
Stay up to date on AI
Follow the latest news on AI and its applications and effects for health care—delivered to your inbox.
A physician or member of the health care team always reviews the message for its validity and cross-checks it to make sure that everything looks OK. Making sure a human is in the loop is critical, Dr. Hill emphasized.
“Questions that involve medical decision-making beyond the level of a nurse, or that require a decision of whether or not a visit is necessary, will go to the physician,” said Dr. Hill. “We have found that when the team’s staff are given the tools via the chart integration with the AI draft response, the vast majority of patient comments are able to be handled by the team.”
Patients who receive AI-generated messages get a footer message that says that part of the message was auto-generated, but edited and approved by their clinician.
“AI itself is a great tool, but it is not perfect. There are times when it does make mistakes,” Dr. Hill said. “That is why you must make sure that you have a human reviewing these messages.”
For example, “generative AI can ‘hallucinate’ … it can provide false answers to a variety of questions,” he explained, noting “the more complicated the question and the vaguer the answer, the more likely the AI will hallucinate.”
Ochsner Health has taken many steps to mitigate this, most importantly requiring that a care team member review the message prior to sending. Ochsner also has set up its prompt engineering to prevent AI from making medical recommendations such as changing medications, Dr. Hill said. “We are also engaging in a slow roll out of this technology only across primary care currently, as we learn the limitations of this technology.”
“However, this technology is improving at a very rapid rate, and we except these issues to become more and more rare,” he added.
Training is offered
Ochsner Health has a module that is required for any physician’s team that is implementing this technology. It outlines the limitations of the AI in its present state—such as hallucinations—and what is expected of the physician and their team.
“Most importantly, it is imperative that the team members review the AI-generated response prior to hitting send, and several process steps have been placed to ensure proper verification, including measuring the time from the opening of the message until it is sent,” Dr. Hill explained.
Documentation burden is reduced
Through this AI model, Ochsner has seen a significant reduction in physician time and notes, as well as a significant increase in first-response resolution—that is, when a patient sends a message, and the team resolves the problem with one response and without forwarding or moving it around in the system.
It’s tempting to want to have a “zero in-basket,” which can be easily accomplished by forwarding everything to someone else, but the problem is that doctors are the top of the approval pyramid, so everybody was forwarding the responses to them, Dr. Hill said. “Instead of trying to create multiple forwarding pathways, we try to find ways to help the health care team to resolve that message as quickly as possible.”
“If it needs to be forwarded to the doctor, that's great and we'll manage it,” he said. “But we want to decrease that level of forwarding and allow patients to have self-service so that you can get your question answered as expeditiously as possible. And it doesn't have to go through four people to get that answer.”
“We know that the AI is impacting physician well-being for the better. A lot of the physicians we've talked to have had a significant increase in their capability interacting with the EHR, which is where we saw a lot of consternation around documentation burden and trying to deal with these extra messages,” said Dr. Hill.
Primary care physicians are currently the chief users of this model, although Dr. Hill would like to expand it to other specialties like ob-gyn and pediatrics. About 30 Ochsner Health clinics in southeastern Louisiana are using GPT-4 as part of the pilot.
Dr. Hill says the eventual goal is to roll out the AI message assistant systemwide.
Learn more with the AMA about the emerging landscape of augmented intelligence in health care (PDF).
