Identifying Common Topics in Patient Portal Messages with Unsupervised Natural Language Processing.

PURPOSE/OBJECTIVE(S): Patient portal messaging is an increasingly important form of communication between patients and medical providers. This has become particularly relevant in oncology, where patients undergo intense longitudinal treatments that require frequent communication regarding symptoms, appointments, and diagnostic results. The rise in the volume of these messages has significantly increased the workload of medical providers and consequent physician burn-out. Natural language processing (NLP), particularly transformer-based models, may offer an automated approach to characterize the content of patient messages and improve message triage and routing. In this study, we employed a state-of-the-art language model (Bidirectional Encoder Representations from Transformers; BERT) to identify data-derived categories of representative topics from real-world data thereby providing basic information to build an appropriate routing system.

MATERIALS/METHODS: Patient-generated portal messages sent to a messaging pool for a single institution radiation oncology department from 2014 to 2023 were extracted. BERTopic, an NLP-based topic modeling technique based on BERT was optimized for topic modeling of patient messages. Uniform Manifold Approximation and Projection (UMAP) was used to reduce dimensionality and visualize topic relationships across messages. The BERTopic-identified topic categories were subsequently labeled manually by one of the physician investigators. Differences of number of messages over time were assessed using t-tests.

RESULTS: A total of 47,492 messages were retrieved. The average number of messages per month from a single patient ranged from 1 to 18 (median 1.67, interquartile range 1.0-2.4). The total volume of patient messages showed a ten-fold increase over the study period, with 101 messages per month sent in 2014 and 999 messages per month in 2022 (p<0.001). BERTopic initially identified 35 topics whose relationships and degrees of overlap were visualized by UMAP. Due to physician-identified similarities, these topics were reduced into 13 categories. The most frequent topic category was messages about laboratory tests or imaging studies: 24.3%, followed by messages expressing appreciation: 18.9%, scheduling discussions: 15.6%, symptom-related messages: 11%, and treatment-related messages: 10.7%.

CONCLUSION: Patient portal messages sent to a single institution radiation oncology department have increased dramatically in volume since implementation, corresponding to a broader national trend. NLP successfully identified common subject themes across patient messages, many of which are related to scheduling. This presents potential opportunities to apply NLP to automate message routing in the future.