⚡ Quick Summary

This systematic review evaluated the use of Large Language Models (LLMs) in virtual patient systems for medical history-taking, highlighting their potential to enhance medical education. The findings indicate that while LLMs show promise in simulating various diseases, they currently lack adequate representation of multimorbidity scenarios.

🔍 Key Details

• 📊 Dataset: 39 studies reviewed, focusing on internal medicine and mental health disorders
• ⚙️ Technology: LLMs such as GPT-3.5 and GPT-4
• 🏆 Performance Metrics: Top-k accuracy ranged from 0.45 to 0.98, with a hallucination rate of 0.31%-5%
• 🧩 Techniques Used: Role-based prompts, few-shot learning, multiagent frameworks, and knowledge graph integration

🔑 Key Takeaways

• 📈 LLMs are transforming virtual patient systems in medical education.
• 💡 Techniques like knowledge graph integration improved diagnostic accuracy.
• 🧑‍🎓 Evaluations typically involved 10-50 students and 3-10 experts.
• 🌍 Common datasets like MIMIC-III showed ICU bias and limited diversity.
• 🔍 Future research should focus on multimodal LLMs and standardized metrics.
• ⚠️ Limitations include small sample sizes and inconsistent evaluation metrics.
• 🗂️ Risk of bias was moderate across included studies.
• 🔗 Open-access datasets are essential for enhancing reproducibility and external validity.

📚 Background

The integration of Large Language Models (LLMs) in medical education represents a significant advancement, offering scalable and cost-effective alternatives to traditional standardized patients. However, the effectiveness of these models, particularly in complex scenarios involving multiple coexisting diseases, remains underexplored. This review aims to fill that gap by systematically evaluating LLM-based virtual patient systems for history-taking.

🗒️ Study

Conducted following the PRISMA guidelines, this systematic review analyzed data from nine databases, focusing on studies published between January 1, 2020, and August 18, 2025. The review specifically targeted LLMs used for history-taking tasks, excluding non-transformer models and unrelated tasks. A total of 39 studies were included, providing a comprehensive overview of the current landscape.

📈 Results

The review revealed that LLM-based virtual patient systems primarily simulated internal medicine and mental health disorders. While many studies focused on distinct single disease types, few addressed multimorbidity or rare conditions. Techniques such as role-based prompts and knowledge graph integration significantly enhanced dialogue and diagnostic accuracy, with top-k accuracy reaching 16.02%. However, the evaluations were limited by small sample sizes and inconsistent metrics, which hindered generalizability.

🌍 Impact and Implications

The findings from this review underscore the potential of LLMs to revolutionize medical education by providing realistic and immersive virtual patient interactions. However, the lack of representation for multimorbidity scenarios highlights a critical area for future research. By addressing these gaps, we can enhance the educational utility of LLMs and improve the training of future healthcare professionals.

🔮 Conclusion

This systematic review illustrates the transformative potential of LLMs in medical education, particularly in history-taking. While current systems excel in simulating various disease types, there is a pressing need for further research to incorporate multimorbidity and enhance the realism of these virtual interactions. The future of medical education could greatly benefit from the integration of advanced LLM technologies and standardized evaluation metrics.

💬 Your comments

What are your thoughts on the use of LLMs in medical education? Do you see potential for improvement in representing complex patient scenarios? Let’s discuss! 💬 Leave your comments below or connect with us on social media:

• 🐦 X
• 📘 Facebook
• 🔗 LinkedIn