MedGame: Storytelling Gamification Empowered by Large Language Models for Medical Education

Synthesis: MedGame transforms static clinical cases into structured, executable storytelling games for medical education, moving beyond the localized question-answering and single-turn feedback that characterize most Large Language Models (LLMs) medical-training systems. It uses a dual-engine design: a Medical Narrative Designer synthesizes case-grounded clinical storylines with states and decision nodes, while a Story Director converts them into dependency-aware Multimodal AI orchestration plans rendered by an interactive platform. The authors release MedGame Bench, a 5,000-case Benchmark and evaluation protocol for Medical Narrative Generation and Story Direction; task-specific fine-tuning substantially improves Open Source LLMs and narrows the gap with commercial models. A pilot student study finds learners perceive MedGame as more engaging and useful than text-only alternatives, extending decision-centered, immersive approaches seen in Multimodal Dialogue in STEM Education and Intelligent Tutoring. Its use of Generative AI for scenario authoring connects to A Semi-Automated System for Generating Dialogue-Based TTS Lessons Using Large Language Models: An Exploratory Study of Educational Potential, and the engagement gains speak to engagement metrics as a design target in professional training.

What this means for practice

• Instructors. Pair every generated case with expert review before it reaches learners: fine-tuning lifted open-source models to near-perfect structural validity, but the authors name medical accuracy as a bottleneck that task-specific tuning does not remove.
• Designers. Adapt a compact open-weight model to your own case corpus rather than prompting a frontier model — LoRA adaptation (rank 32, alpha 64) on the benchmark's 4,000-case training split moved a 32B model's persona-utilization score from 5.41 to 8.50 — and prefer a clinically pretrained base model.
• Designers. Plan the rendering dependencies explicitly, since image, audio and video generators sit outside the framework evaluated here and the study does not test how their quality affects the resulting stories.
• Instructors. Pilot presentation formats with real learners before scaling: in an eight-student paired comparison, multimodal MedGame storylines scored 4.19 versus 3.79 for text-only on a 1-5 perception scale (p = 0.0039), with a mild cognitive-load trade-off.
• Researchers. Evaluate authoring systems on generation criteria first and defer engagement claims: this study's learner-facing evidence is perceived engagement and usefulness from eight students, not retention or clinical reasoning.

Limitations

• The learner-facing evaluation is a pilot with 8 senior medical students, each rating 5 clinical cases on a 1-5 Likert scale; the authors state they could not conduct longitudinal tracking of knowledge retention, clinical reasoning or skill transfer.
• Image, audio and video generation modules are treated as external tools, so the reported results cover only the narrative-design and story-direction capabilities of the LLMs, not the multimodal stack that learners actually see.
• Benchmark contamination was checked but not eliminated: 409 of the 1,000 test cases (40.9%) have a high-similarity patient variant in the 4,000-case training set, and the comparison relies on LLM-as-a-judge scores.
• Medical accuracy remains dependent on domain pretraining and expert revision rather than on the fine-tuning demonstrated here.

Citation

Wu, Zhou, Ma, Chen, Gao, Lin, Wu, Gou, Liu, Lau & Dou (2026). MedGame: Storytelling Gamification Empowered by Large Language Models for Medical Education. arXiv preprint (cs.CL).