Abstract
IntroductionElectronic patient portals, such as Epic MyChart provide a new method for patients and their healthcare team to share personal medical information and request medical advice. The volume of messages that clinicians and staff are required to respond to has significantly increased over time. Using natural language processing (NLP) and machine learning (ML), we identify the most frequent conversation topics within a large health system with over 800,000 patients, 200,000 with CV disease or risk factors for CV disease, and 140,000 yearly outpatient encounters. Our ultimate objective is to develop tailored support systems using artificial intelligence for electronic patient communication and engagement.HypothesisUsing NLP and ML, we can identify common topics from MyChart messages between patients and their cardiology teams.MethodsWe identified 75,138 individual messages from 6/2013 to 1/2019 generated by patients and only free-text medical advice requests. The conversation data was pre-processed, including removing stop words and lemmatizing words. After removing sparse terms, we extracted bag-of-words (unigram, bigram, trigram) tokens as features for each message. Using a Latent Dirichlet Allocation (LDA) model to fit the data, we used a UCI coherence score (C_v) to choose the optimal number of topics. A C_v greater ≥ 0.6 indicates that the extracted topics have adequately good interpretation.Results and ConclusionsWe identified 10 topics sent by cardiology patients with a C_v of 0.59. Other than miscellaneous (15.8%), the most frequent questions were about appointments and medications, blood pressure readings, description of results as well as expression of emotions or new symptoms. Artificial intelligence can provide insight and automatically categorize patient free-text messages. Future work should address message triaging and automated responses to ensure correct and efficient triage for urgent cardiovascular medical issues.