HealthCoach: An LLM-driven chatbot to help patients gain actionable insights from their EHR notes and prepare for the next visit - When Steven Keating helped diagnose his own brain cancer from MRIs and health data in 2015, he became a resolute champion for patients’ access to their data. He didn’t survive to see implementation of the Cures Act in 2021, but his perspective remains relevant: “Why, as patients, are we the last in line for our own health data?” Patients who read their electronic medical notes (open notes) report better remembering next steps, more participation in shared decision-making, and enhanced self-management of chronic diseases. They are often the only connecting thread between various healthcare encounters, sometimes at different centers. As a result, they are uniquely poised to identify “blindspots” – safety issues that patients know about but clinicians do not. Some patients only recognize blindspots by reading their notes (ie their main concern was completely missed). To harness this information, members of our team developed an online patient-clinician alignment tool (PCAT) to gather patient visit priorities, recent health history, and concerns prior to a visit. At 2 sites, patients shared actionable information, such as data from other centers or test/ referral problems, including blindspots. Although it’s been over 5 years since Keating passed, many patients still don’t know about open notes or other e-tools or cannot use them due to language or literacy gaps. Artificial Intelligence (AI) could help, but research on patient AI use is sparse, especially for patients with limited English health literacy. This is a missed opportunity because an AI tool designed for patients could be a powerful engine for understanding their own data. We propose to develop and test a “HealthCoach” (HC) chatbot in English and Spanish to help patients with varying literacy levels gain insights from their notes, use this knowledge to prepare for the next visit, and co-generate with HC a pre-note that reflects their values and priorities ahead of the visit. We will use focus groups in 2 states to understand patient priorities for HC and guide end-user design. Using deidentified databases of patient notes and PCAT data, we will iteratively develop and test the HC chatbot. We will use standardized assessments (PDQI-10) for HC-generated notes with both patient and clinician raters. Then we’ll live-test the HC chatbot in a controlled HIPAA-compliant compute research environment with English and Spanish-preferring patients who have an upcoming primary care visit. HC will simplify the last note by patient preferences, elicit patient questions, visit priorities, recent history, and concerns (through an adapted PCAT), and structure patient input into a pre-note that patients can review and approve. We will test 1) Patient Engagement/Learning; 2) Documentation Quality/Safety; and 3) Feasibility/Acceptability. The know-how for patient-centered AI use is an information vacuum. AI tools lack patients’ voice and rarely reach those who need them most. Our research will establish foundational knowledge, novel patient perspectives on AI assessments, and a freely available HealthCoach chatbot with educational materials to empower patients with their health data.
