An Educational Use Case of Post-PCI Follow-Up using a Conversational AI Platform
This abstract was part of the 2026 UCSF AI and Education Symposium
Educational Need
Cardiology training requires proficiency in systems-based practice (SBP) and quality improvement (QI), core ACGME milestones that are challenging to incorporate into busy clinical environments. As AI tools rapidly proliferate, understanding the benefits, challenges, and pitfalls of AI implementation in relation to SBP and QI has become essential. Following a structured process for AI implementation—defining the clinical need, mapping the workflow, building and validating the AI model, evaluating performance, conducting usability testing, piloting in controlled environments, and assessing real-world impact—helps safeguard patient safety, maximize value, and support reliable, ethical deployment. This project provides a concrete educational opportunity for fellows to engage in the implementation of a novel AI tool from inception to completion.
Clinical Needs Assessment
Percutaneous coronary intervention (PCI) is a highly effective revascularization strategy for acute coronary syndromes and refractory angina. Following PCI, patients must adhere consistently to dual antiplatelet therapy (DAPT) to prevent stent thrombosis, a potentially catastrophic event associated with myocardial infarction, malignant arrhythmias, and death. The risk of stent thrombosis is highest within the first 30 days after PCI, making early medication adherence crucial. More frequent communication can help improve adherence, yet clinic teams face time constraints and competing clinical duties that limit their ability to provide consistent touchpoints. This gap highlights the need for scalable, efficient approaches to support early post-PCI monitoring and adherence.
AI-Enabled Solution
AI-driven conversational agents offer an opportunity to automate routine post-discharge follow-up. A novel post-discharge follow-up protocol was developed in collaboration between the UCSF Division of Cardiology, Department of Medicine, Office of Population Health, and a commercial AI platform, Flowline Health. A non-deterministic, conversational AI voice agent is given a script to call patients who have undergone outpatient PCI and been discharged home. The voice agent confirms patient identity and procedure, assesses medication adherence to DAPT, and evaluates for complications. Depending upon the scenario, the voice agent escalates concerning responses to the appropriate level of care.
Prototype, Feasibility Evidence, and Evaluation Plan
A prototype is undergoing beta-testing by interprofessional providers at UCSF, providing iterative feedback to optimize safety and patient experience. Safety is prioritized through scope limits, limited scripts, confirming collected information, and escalation protocols. A short-term pilot program will be performed using for a subset of outpatient PCI cases. Call transcripts will be reviewed for accuracy; clinic call volume and staff experience will also be monitored. Long-term, we plan to expand this approach to all outpatient cardiac catheterization procedures and other indications once safety and efficacy are established.
We present a novel opportunity for trainees to engage in experiential learning in the implementation of an AI-based voice agent. Through this project, trainees will participate in a needs assessment, identify opportunities for AI integration, conduct usability testing, evaluate a pilot intervention, and help to evaluate the outcomes of a novel AI intervention. As AI becomes increasingly integrated into clinical care, competency in AI implementation should be an essential component of physician training.
Contact
Jessica Holtzman, [email protected]
Stephanie Rogers, [email protected]
Krishan Soni, [email protected]
Christopher Barnett, [email protected]
Salman Rahman, [email protected]