Kaiser Permanente uses AI to redirect 'simple' patient messages from physician inboxes

Artificial intelligence categorization can help stem the flood of patient messages that would otherwise demand physicians’ expensive time, Kaiser Permanente researchers report.

In a recently published JAMA Network Open research letter, members of the system’s research division and medical group outlined a strategy that used real-time natural language processing (NLP) algorithms to attach category labels to messages and then direct them to an appropriate respondent.

The approach, they wrote, allowed 31.9% of the more than 4.7 million patient messages reviewed by program staff to be resolved before reaching the inbox of a specific physician. Instead, these messages were handed by a “regional team” made up of medical assistants or teleservice representatives, pharmacists and other doctors.

“Physicians in a modern health care system are managing in-person patient care alongside virtual care such as video visits and secure messaging,” Kristine Lee, M.D., associate executive director of virtual medicine and technology at The Permanente Medical Group and the study’s senior author, said. “While doctors welcome expanded ways to connect with their patients, they also need assistance with the changing workload. Our program found that AI tools may offer tools that could help when paired with robust workflows.”

Kaiser Permanente’s Desktop Medicine Program trained its NLP on roughly 20,000 patient messages that were annotated by triage nurses and physicians. It was integrated into the system’s electronic health record in May 2022 with an initial focus on adult and family medicine. It received bimonthly updates and was expanded to tackle pediatric messages in October 2022.

The researchers’ study looked at the 4.7 million messages received through the system from April to August 2023. Among these, 77.6% were given at least one label, with medication-related messages (32.8%) being the most common. Nearly 3 in 10 messages were assigned two labels by the system, for which the combination of skin condition and medication labels were most frequent (41.1%).

Across the five-month period, more than 1.5 million messages were flagged for and resolved by the regional team. The messages that were diverted from physician inboxes included “simple” questions on pharmacy hours or medication refills, lead author Vincent Liu, M.D., said in a release.

Beyond limiting physician burden, the researchers noted that the system’s categorization of message types could be useful for monitoring and acting upon care trends, such as a week-to-week jump in specific infectious diseases.

They also wrote that the program still has room for improvement, either in addressing multilabeled or non-labeled content, or by including more recently released large language models like GPT-4.

“We certainly have not solved the problem of high volumes of inbox messages for our physicians, but it is gratifying to see how we could use technology to resolve these messages in real-time,” said co-author Jennifer Manickam, M.D., chair of adult and family medicine technology leads at The Permanente Medical Group, said in a statement. “This study lays the foundation for more applications of AI to enhance the way that we care for our patients through all channels.”

Several provider organizations hoping to either reduce or better justify the time physicians spend answering patient messages have attached a price tag to queries that require several minutes to answer. These have been shown to somewhat change patients’ care-seeking behaviors and more often than not are entirely handled by the patients' insurance.

Others, like Ochsner Health, have opted to explore generative AI as a potential answer. Last fall, the system announced a pilot that would use the technology to draft responses to routine patient requests, which clinicians would then review and edit before sending.