Paper Type

Short

Paper Number

PACIS2026-1789

Description

Online health consultation faces challenges of physician scarcity and potential prescription bias. This study proposes a Role-Aware Debiased Curriculum Learning (RADCL) framework for dialogue-based medication recommendation to address the limitation of traditional curriculum learning in distinguishing heterogeneous sources of sample difficulty arising from doctor-side diagnostic complexity and patient-side symptom ambiguity. Following the design science methodology, RADCL integrates three modules: a dialogue encoding module with BERT and Graph Attention Networks for structural understanding, a role-aware difficulty measurement module introducing attention-based metrics to quantify doctor and patient contributions, and an iterative curriculum learning module dynamically adjusting sample weights. Experiments on the DIALMED dataset with five-fold cross-validation demonstrate that RADCL outperforms state-of-the-art baselines, achieving 4.04% and 2.91% relative improvements in Jaccard and F1 scores, respectively. This work advances curriculum learning in healthcare NLP and provides a robust solution for online medication consultation.

Comments

14-Healthcare

Share

COinS
 
Jul 5th, 12:00 AM

A Role-Aware Debiased Curriculum Learning Approach for Dialogue-based Medication Recommendation

Online health consultation faces challenges of physician scarcity and potential prescription bias. This study proposes a Role-Aware Debiased Curriculum Learning (RADCL) framework for dialogue-based medication recommendation to address the limitation of traditional curriculum learning in distinguishing heterogeneous sources of sample difficulty arising from doctor-side diagnostic complexity and patient-side symptom ambiguity. Following the design science methodology, RADCL integrates three modules: a dialogue encoding module with BERT and Graph Attention Networks for structural understanding, a role-aware difficulty measurement module introducing attention-based metrics to quantify doctor and patient contributions, and an iterative curriculum learning module dynamically adjusting sample weights. Experiments on the DIALMED dataset with five-fold cross-validation demonstrate that RADCL outperforms state-of-the-art baselines, achieving 4.04% and 2.91% relative improvements in Jaccard and F1 scores, respectively. This work advances curriculum learning in healthcare NLP and provides a robust solution for online medication consultation.