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

MedGame: Storytelling Gamification Empowered by LLMs for Medical Education§

Core Methodology§

MedGame introduces a dual-engine framework consisting of a Medical Narrative Designer (MND) and a Story Director (SD). Both are instantiated as LLMs that are either pre-trained or fine-tuned. The MND takes a static clinical case (e.g., patient history, symptoms, diagnosis) and generates a structured narrative with decision nodes and branching storylines. Formally, given a case $C = (\text{history}, \text{findings}, \text{diagnosis}, \dots)$, the MND produces a directed graph $G = (V, E)$ where $V$ are narrative states and $E$ are transitions labeled with decision options. Each state $v \in V$ contains multimodal content (text, images, audio).

The SD then converts the graph $G$ into a dependency-aware orchestration plan $P$ that specifies the order and conditions for presenting states and handling user decisions. This plan is executed by the MedGame player platform, which renders the interactive experience.

Benchmark: MedGame Bench§

MedGame Bench comprises 5,000 clinical cases across 10 specialties. For each case, it includes a reference narrative graph and a reference orchestration plan. Two tasks are defined:

• Medical Narrative Generation (MNG): Given a clinical case, generate the graph $G$. Evaluation metrics include BLEU, ROUGE-L, and graph edit distance (GED).
• Story Direction (SDirect): Given a graph $G$, generate the plan $P$. Evaluated using plan accuracy and F1 score on dependency edges.

Fine-Tuning Experiments§

Open-source LLMs (e.g., LLaMA-2-7B, Mistral-7B) were fine-tuned on MedGame Bench using supervised fine-tuning (SFT). The training objective is the standard autoregressive language modeling loss:

$$ \mathcal{L}_{\text{SFT}} = - \sum_{t=1}^{T} \log p(y_t | y_{<t}, x) $$

where $x$ is the input prompt (clinical case) and $y$ is the output (narrative or plan).

Low-Rank Adaptation (LoRA) was used for efficient fine-tuning. A simplified PyTorch snippet:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, get_linear_schedule_with_warmup
from peft import LoraConfig, get_peft_model

model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.bfloat16)
lora_config = LoraConfig(
    r=8,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")

# Assume dataloader yields (input_ids, labels)
optimizer = torch.optim.AdamW(model.parameters(), lr=2e-4)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=100, num_training_steps=1000)

for batch in dataloader:
    outputs = model(**batch)
    loss = outputs.loss
    loss.backward()
    optimizer.step()
    scheduler.step()
    optimizer.zero_grad()

Results§

Fine-tuned Mistral-7B achieved a BLEU of 42.3 on MNG and plan accuracy of 88.5% on SDirect, closing the gap with GPT-4 (BLEU 48.1, accuracy 92.0%). A pilot student study (n=30) showed that MedGame significantly improved engagement (Likert score 4.5/5 vs 3.2/5 for text-only) and perceived usefulness (4.3/5 vs 3.5/5).

Key Takeaway§

MedGame demonstrates that LLMs can be effectively specialized for structured narrative generation in educational gamification, offering a practical approach to turn static content into interactive learning experiences.