Paper Type
Short
Paper Number
PACIS2026-1785
Description
Dialogue-based depression detection faces a fundamental challenge: raw dialogues contain conversational noise that obscures clinical signals, while LLM-generated summaries risk hallucinating symptom claims. This paper proposes MVAF-3D, a novel mutual-verified adaptive fusion framework that integrates both views through complementary mechanisms. A mutual verification module quantifies summary claim credibility by assessing semantic alignment with original utterances, simultaneously using summaries to guide attention toward clinically salient dialogue content. Experiments on the D4 dataset demonstrate that MVAF-3D achieves an F1-score of 0.9111, outperforming the best single-view baseline with F1-score of 0.8965. The framework effectively addresses the noise-hallucination dilemma, providing a more robust approach to depression risk detection with implications for clinical screening applications.
Recommended Citation
Kuang, Junwei; Liu, Yi; Liu, Yiai; Chen, Donglei; Zhang, Xinyu; and Zhao, Shiyi, "Detecting Depression on Medical Dialogue by Adaptive Fusing LLM Summaries and Raw Texts with Mutual Verification" (2026). PACIS 2026 Proceedings. 14.
https://aisel.aisnet.org/pacis2026/ai_ml/ai_ml/14
Detecting Depression on Medical Dialogue by Adaptive Fusing LLM Summaries and Raw Texts with Mutual Verification
Dialogue-based depression detection faces a fundamental challenge: raw dialogues contain conversational noise that obscures clinical signals, while LLM-generated summaries risk hallucinating symptom claims. This paper proposes MVAF-3D, a novel mutual-verified adaptive fusion framework that integrates both views through complementary mechanisms. A mutual verification module quantifies summary claim credibility by assessing semantic alignment with original utterances, simultaneously using summaries to guide attention toward clinically salient dialogue content. Experiments on the D4 dataset demonstrate that MVAF-3D achieves an F1-score of 0.9111, outperforming the best single-view baseline with F1-score of 0.8965. The framework effectively addresses the noise-hallucination dilemma, providing a more robust approach to depression risk detection with implications for clinical screening applications.
Comments
01-AIML