Besides scheduling doctor visits and keeping track of health records, hospital portals offer a way for patients to reach out to their providers with questions. And patients are using the online portals with increasing frequency—patient portal access has seen a 46% increase since the COVID-19 pandemic.
"We want to encourage patients to interact with us through the patient portal, but we get a lot of messages," says Tim Burdick '89, MED '02, associate professor of community and family medicine at the Geisel School of Medicine. "How do we respond to those messages in a timely fashion, and make sure that we're responding to those messages with the right priorities?"
Thanks to a collaboration between computer science researchers at the Persist Lab led by Assistant Professor Sarah Preum and a team of clinicians from Dartmouth Health working with Burdick, there are solutions on the horizon.
Their ongoing research project, called PortalPal, currently seeks to address two key issues:
- triage incoming messages from patients to expedite responses to those with more urgent symptoms;
- and ensure that providers have all the details they need to respond to a query by developing an AI-based algorithm that can instantly generate high quality follow-up questions to save time for patients as well as providers.
The work began as a class project for the Computational Healthcare course taught by Preum in 2021. Guarini School of Graduate and Advanced Studies PhD students Joseph Gatto and Parker Seegmiller used AI language models to determine the urgency of messages from publicly accessible online medical Q&A forums designed for people to ask questions about their health during the pandemic.
"It became quickly apparent that large language models have the potential to filter messages based on the urgency of symptoms and actionability, and I was eager to see if we could further refine the idea and put it to work in a real-world health care setting," says Preum.
She found an enthusiastic collaborator in Burdick, a physician who is no stranger to bringing new technology to the clinic.
In his very first job in health care, he was involved in implementing electronic health records.
"I've always believed that involving clinicians in the research process is really important to make sure that we get innovative technology developed properly," Burdick says.
The researchers sought permission to access actual patient queries sent to Dartmouth Health providers through their portal. The data was anonymized to protect patient identities and stored on secure Dartmouth servers that are not connected to any networks and only accessible to authorized research team members.
Researchers at Preum's Persist Lab began by cleaning up the data with guidance from clinicians, filtering out information that is not clinically relevant to the current message and identifying which elements from the patient's chart would be most relevant to help assess and answer the query.
Large language models were trained and adapted to sort the messages in order of urgency and identify gaps in the information provided by patients. To verify and further refine the models' performance, a large team of DH doctors and nurses agreed to annotate or label queries based on perceived priority and created follow-up questions that patients would need to answer for providers to have the most complete picture of the medical issue and be able to offer advice. Gatto will present part of this work at the annual meeting of the Association for Computational Linguistics in July.
Incomplete messages are common in portals, and they are a huge drain on providers' time, says Burdick. So, instead of a nurse having to engage in a lengthy back-and-forth with the patient, PortalPal's solution is designed to step in as soon as the message is drafted by a patient to help fill the gaps in real time.
"This way, the patient can get care much faster, and it reduces the workload of nurses and doctors," says Burdick.
The team is now working to make PortalPal’s triaging solution better so that it can make well-informed decisions about which messages need to be looked at right away and reorder the physician's inbox accordingly.
Combining the question generator for patients with the triaging tool for providers sets PortalPal apart, the researchers say—current AI assistants do one or the other.
More tools are in development. Seegmiller is currently leading a project that uses large language models to generate initial drafts of responses to queries that physicians and nurses can add detail to and refine. "Our primary goal is to develop these AI solutions in a way that keeps expert health care workers in the loop, while reducing their overall workload," says Seegmiller.
PortalPal is designed to be implemented in partnership with a health care provider, integrated into their existing patient portal, taking precautions that sensitive data is secured within the health care systems' technology firewall.
The researchers are also looking to build a variety of checks and balances into the system. For instance, having PortalPal repeat information back to patients allows them to confirm that the correct details are presented to the provider when they see the message. Nonurgent messages that are moved down on the priority queue can be flagged to catch attention if they remain unopened for a certain time. That way, messages won't slip through the cracks even in an overflowing inbox.
"We would always implement this in a setting where there is a human in the loop," says Burdick. All messages are seen and addressed by a clinician, and PortalPal makes the process more efficient at both ends.
Gatto notes that the demand for telehealth has risen without a proportional increase in the amount of administrative time that clinicians get per day, so they work longer shifts to meet the demand. A lot of studies link this phenomenon to clinician burnout, says Gatto.
Burdick agrees. "We have older, sicker patients, and it's very stressful to come to work and feel like you're not providing timely care to patients. If we can reduce the cognitive and emotional burden on clinical teams, I think we can decrease burnout and make their work more enjoyable and rewarding."