Abstract
BACKGROUND: The emergence of chatbots in health care is fast approaching. Data on the feasibility of chatbots for chronic disease management are scarce. OBJECTIVE: This study aimed to explore the feasibility of utilizing natural language processing (NLP) for the categorization of electronic dialog data of patients with inflammatory bowel diseases (IBD) for use in the development of a chatbot. METHODS: Electronic dialog data collected between 2013 and 2018 from a care management platform (UCLA eIBD) at a tertiary referral center for IBD at the University of California, Los Angeles, were used. Part of the data was manually reviewed, and an algorithm for categorization was created. The algorithm categorized all relevant dialogs into a set number of categories using NLP. In addition, 3 independent physicians evaluated the appropriateness of the categorization. RESULTS: A total of 16,453 lines of dialog were collected and analyzed. We categorized 8324 messages from 424 patients into seven categories. As there was an overlap in these categories, their frequencies were measured independently as symptoms (2033/6193, 32.83%), medications (2397/6193, 38.70%), appointments (1518/6193, 24.51%), laboratory investigations (2106/6193, 34.01%), finance or insurance (447/6193, 7.22%), communications (2161/6193, 34.89%), procedures (617/6193, 9.96%), and miscellaneous (624/6193, 10.08%). Furthermore, in 95.0% (285/300) of cases, there were minor or no differences in categorization between the algorithm and the three independent physicians. CONCLUSIONS: With increased adaptation of electronic health technologies, chatbots could have great potential in interacting with patients, collecting data, and increasing efficiency. Our categorization showcases the feasibility of using NLP in large amounts of electronic dialog for the development of a chatbot algorithm. Chatbots could allow for the monitoring of patients beyond consultations and potentially empower and educate patients and improve clinical outcomes.