MedGame: Turning Medical Records into Interactive Narrative Games with LLM Directors
The Problem: Static Records vs. Dynamic Reasoning
A medical student preparing for clinical rotations receives a case: a 55-year-old male with 3 hours of chest pain radiating to the left arm, diaphoresis, hypertension history, and ST-segment elevation on ECG. Traditional study means read, memorize, answer. But real clinical reasoning unfolds step by step—what to ask, what to examine, when to order tests, when to adjust treatment. Medical records are static; reasoning is dynamic. MedGame targets exactly this mismatch.
Core Idea: A Two-Engine Architecture
MedGame separates concerns into two engines:
Engine 1 — Medical Narrative Designer. Takes a static case summary and outputs a structured story graph of state nodes and decision points. Each node is a clinical scene ("ER initial consult," "lab results return," "sudden deterioration"); each decision point branches (immediate thrombolysis vs. angiography first vs. conservative treatment). The model generates functional descriptions only ("consultation room," "resuscitation bay"), not images.
Engine 2 — Story Director. Expands each node into concrete multimodal generation tasks: scene images, character images, dialogue, vital-sign panels. Crucially, these tasks are organized as a directed acyclic graph (DAG)—some tasks must wait on predecessors (the resuscitation scene depends on the deterioration node), others run in parallel. Multimodal generation becomes a dependency-aware scheduling problem rather than a linear pipeline.
The separation has engineering value: narrative design requires medical knowledge, story direction requires visual orchestration—optimizing each with different models or fine-tuning data beats one giant model doing everything.
Data and Scale
- 5,000 clinical cases across eight specialties: cardiology, endocrinology, gastroenterology, hematology/oncology, nephrology, neurology, pulmonary/critical care, rheumatology
- 4,000 training / 1,000 test
- Each case is converted into a branching narrative game where the player (medical student) makes choices at decision points
- Claude-Sonnet-4.5: 99.5%
- Gemini-3-Pro: 100.0%
- Qwen3-32B: 79.4%
- Gemma-3-27B / MedGemma-27B: below 60%
- Claude-Sonnet-4.5 / Gemini-3-Pro: above 99%
- Open-source models: 56.50%–80.30%
- The two-engine decoupling is a sound engineering decision
- DAG scheduling turns multimodal generation from a linear pipeline into a dependency graph with clear efficiency gains
- 5,000 cases across eight specialties is a solid data scale
- Evaluation relies mainly on GPT-5.2 as judge; LLM-as-judge bias remains an open issue despite expert calibration
- Case source representativeness is limited
- No reported learning outcomes—whether playing these games actually improves clinical reasoning versus reading records is unanswered
- The large open-source gap means the system is effectively gated behind commercial frontier models
- Paper: https://arxiv.org/abs/2607.21570
- Code: https://github.com/med-air/MedGame
Results: Commercial Models Dominate
Evaluation has two layers: automatic structural validation (JSON compliance, business-logic consistency) and GPT-5.2-as-judge content quality scoring.
Strict Validation:
Story Director task (Sample-wise Validation):
Fine-tuning narrows the gap—Qwen3.5-27B improved Strict Validation from 79.4% to 99.1%—but requires separate narrative and direction fine-tuning data.
Honest Assessment
Strengths:
Limitations:
A notable detail: the multimodal rendering is claimed to add "perceptual value beyond pure text interaction," but the paper never quantifies it. Is gamified learning actually more effective, or just flashier?
Conceptual Takeaway
MedGame's essence is converting static knowledge artifacts into dynamic decision trajectories. The idea generalizes beyond medicine—to legal cases, historical events, engineering troubleshooting—anywhere the reasoning process matters as much as the conclusion. The two-engine architecture (one for narrative structure, one for multimodal orchestration) is a transferable template.