PCQC: Privileged Counterfactual Question Credit for Multi-Turn Medical Dialogue

TL;DR

The paper introduces PCQC, a method for training medical dialogue policies using privileged counterfactual question credit. PCQC learns from questions never asked by constructing their answers using privileged patient facts. This approach improves diagnostic accuracy and reduces inquiry turns.

Why It Matters

Effective medical dialogue systems can improve patient outcomes by gathering relevant information efficiently. PCQC's ability to learn from unexecuted questions can lead to more accurate diagnoses and better patient care. This technology has potential applications in clinical decision support and medical chatbots.

Key Findings

• PCQC achieves 63.10% mean diagnostic accuracy
• outperforms GRPO and ATPO by 4.38 and 4.21 percentage points
• reduces inquiry turns by 33.1% compared to GRPO