This abstract was part of the 2026 UCSF AI and Education Symposium
Problem
Clinical training requires more than factual knowledge—it demands the ability to synthesize information, reason through uncertainty, and apply physiology, pharmacology, diagnostics, and guidelines to real patient scenarios. Students often struggle to transition from memorizing content to thinking like clinicians, especially when faced with nuanced exam questions or patient workups in early rotations. Existing AI tools prioritize content retrieval rather than reasoning. They can summarize information but cannot reliably evaluate clinical thinking, trace decisions to underlying mechanisms, or scaffold learners through the conceptual steps that connect basic sciences to diagnostics and guideline-driven management. There is currently no scalable, discipline-adaptive tool that provides formative, structured clinical reasoning feedback to learners across health professions.
How AI Addresses the Problem
MedFlow Atlas is an AI-powered educational platform that teaches clinical reasoning through a bidirectional approach. It connects a learner’s free-text or multiple-choice responses to the mechanisms, diagnostic frameworks, and evidence-based guidelines that shape real clinical decision-making. Unlike traditional AI assistants, MedFlow contextualizes answers by mapping reasoning in both directions:
Downstream: Links decisions to foundational sciences (anatomy, physiology, pharmacology, pathophysiology).
- Upstream: Aligns each step with current clinical guidelines, primary literature, and professional standards.
MedFlow produces consistent, structured outputs using a standardized reasoning framework across all disease states. For any clinical topic, MedFlow can generate:
Mechanistic pathophysiology maps
Diagnostic workflows (labs, imaging, rule-in/rule-out logic)
Guideline-basedtreatment flowcharts
Monitoring and follow-up parameters
Interprofessional adaptations (pharmacy, medicine, nursing, dentistry, physical therapy)
This ensures learners not only know what to do, but why options differ based on mechanism, patient factors, and evidence strength.
Prototype & Feasibility
Our working prototype already delivers disease-specific clinical vignettes across multiple topics. It parses student responses, identifies key clinical concepts, and generates feedback on gaps in reasoning, misapplied concepts, or missing guideline elements. Students can select learning modes—multiple-choice, free-response, rapid-fire, light review, or in-depth exploration.
Prototype features include:
Automated analysis of learner responses against recent guidelines
Structured feedback written in clinical language
A Gantt-style linear visual that highlights decision steps
Ability to upload class notes or select specific sources for alignment
Because MedFlow is built on modular reasoning templates, extending it to new disease states—such as asthma, HFrEF, diabetes, PE, or even toxicology—is immediately scalable.
Evaluation of Efficacy
We will evaluate MedFlow through a mixed-methods approach:
Performance Outcomes: Compare exam performance of students using MedFlow against those using traditional study methods.
Expert-Blinded Review: Faculty reviewers score reasoning quality using standardized rubrics across pre- and post-intervention responses.
Learner Feedback: Students self-report perceived efficiency, confidence, time saved, and usability.
Consistency Monitoring: Faculty verify that guideline-based outputs remain up-to-date, leveraging MedFlow’s editable guideline layer.
MedFlow Atlas is designed as a living, transparent system that evolves with guidelines and user-selected sources. Its framework demonstrates how a single, explainable AI can meaningfully strengthen clinical reasoning, bridge basic and clinical sciences, and scale across multiple health professions.
Contact
Maria Charles, [email protected]
Kim Madlangbayan, [email protected]