PhD Dissertation Proposal: Zhonghai Yao, From Records to Journeys: Rethinking AI for Patient-Centric Healthcare
Content
Speaker
Abstract
Consumer AI is becoming a main interface between patients and the healthcare system. As medical records integrate with real-time health data, assistants can maintain context across sessions. This also raises a new problem: a system that is correct once can still be unsafe later if it forgets units, misreads dates, or misses a “red-flag” symptom.
This thesis studies patient education grounded in digital health records. We argue that evaluation should move from one record at one time to a longitudinal patient journey. We define the Personal Health Workspace (PHW), a user-controlled setting where a stateful assistant supports ongoing questions, corrections, and next steps. In PHW, we build EHR-grounded assistants that must both reason with evidence and explain clearly under two real pressures: multi-source inputs and long-horizon use. We extend patient education beyond text by incorporating multimodal clinical inputs and outputs, including 2D/3D medical imaging and video, thereby transforming the core interaction in patient-facing education. We also design a multi-stage training and optimization framework to help agents guide patients over progressively longer horizons, from single turns to full visits to repeated follow-ups. In addition, we develop medical DeepResearch tools that tie each response to the right evidence to improve transparency and trust.
To make safety measurable, we propose two protocolized stress-test families: multi-source tests and long-horizon tests. They target system failures that static benchmarks often miss. Our goal is to improve real outcomes: safer patient actions under record-linked guidance, reduced clinician burden in triage, and more equitable benefits across different literacy and resource levels.
Advisor
Hong Yu