A systematic review (2020–2025) assessed large language model (LLM)-based virtual patient systems for history-taking, highlighting strengths in simulating single diseases but identifying limited coverage for multimorbidity and rare conditions.

Why It Matters To Your Practice

• LLM-based virtual patients offer scalable, cost-effective training alternatives to standardized patients.

• Gaps in multimorbidity simulation may limit real-world readiness for trainees.

• Biases in common datasets (e.g., MIMIC-III) can affect generalizability and equity.

• Limited sample sizes and inconsistent metrics reduce confidence in findings.

Clinical Implications

• Current systems excel at single-disease scenarios, but few address complex patient presentations common in practice.

• Knowledge graph integration improves diagnostic reasoning and accuracy.

• Advances in multimodal LLMs (e.g., speech, imaging) enhance realism for clinical encounters.

• Standardizing metrics and expanding datasets will better inform adoption for clinical teaching.

Insights

• Role-based prompts, few-shot learning, and multiagent frameworks drive performance gains.

• Hallucination rates remain low (0.31–5%) but require monitoring for patient safety.

• Usability remains high (SUS ≥80), supporting integration into curricula.

• Transparent, diverse datasets are needed to improve reproducibility and equity.

The Bottom Line

• LLM-based virtual patients are promising for medical education but must expand to include multimorbidity, standardized metrics, and open, diverse datasets for broader clinical relevance.