Kaiser Permanente implements AI to redirect routine patient messages
Researchers at Kaiser Permanente have demonstrated how artificial intelligence (AI) can be deployed to manage the overwhelming number of patient messages, potentially liberating physicians from routine queries that consume their valuable time. This innovative strategy, detailed in a recent publication in JAMA Network Open, employs real-time natural language processing (NLP) algorithms to categorise patient communications, thereby routing them efficiently to the most suitable responders.
The study, led by Kristine Lee, M.D., associate executive director of virtual medicine and technology at The Permanente Medical Group, reveals that using AI to label and redirect messages allowed for 31.9% of over 4.7 million patient messages to be addressed before they could reach an individual physician’s inbox. Instead, these were handled by a regional team consisting of medical assistants, teleservice representatives, pharmacists, and other doctors.
Lee explains, “Physicians today balance in-person care with virtual consultations, including video visits and secure messaging. While they appreciate these diverse interaction modes, they also require support to manage the increasing workload effectively. Our findings suggest that AI, coupled with strong workflow systems, could significantly aid this process.”
The Desktop Medicine Program at Kaiser Permanente developed the NLP system by training it on approximately 20,000 patient messages previously annotated by triage nurses and physicians. Implemented into the electronic health record system in May 2022, the NLP initially focused on adult and family medicine and was updated bi-monthly. By October 2022, the program expanded to include paediatric communications.
During the period from April to August 2023, the study examined the handling of 4.7 million messages, with 77.6% being categorised with at least one label. Messages related to medication were the most frequent, accounting for 32.8%, with many messages receiving dual labels, such as those combining skin conditions and medication issues (41.1%).
Over the five months, more than 1.5 million messages were successfully resolved by the regional team. These included straightforward inquiries concerning pharmacy hours or medication refills, which were diverted from physicians’ inboxes, thereby easing their workload, as noted by Vincent Liu, M.D., lead author of the study.
The researchers also highlighted the potential of message categorisation in identifying and responding to care trends, such as increases in specific infectious diseases. They acknowledged that the system could still be improved, particularly in handling messages with multiple labels or those that remain unlabelled. Future enhancements might include incorporating more advanced AI models, such as GPT-4.
“The high volume of messages our physicians receive hasn’t been completely addressed yet, but it’s rewarding to see how technology can help us manage these communications more effectively in real time,” said Jennifer Manickam, M.D., chair of adult and family medicine technology leads at The Permanente Medical Group.
Other healthcare providers have also experimented with AI to optimise the management of patient messages. For example, Ochsner Health has initiated a pilot programme using generative AI to compose draft responses to routine patient requests, which clinicians then review and personalise before dispatch.
As Kaiser Permanente and other organisations continue to explore AI’s potential, the landscape of physician-patient communication is set for significant transformation, promising to enhance the efficiency and quality of healthcare delivery.
