Scholarlink

Major Depressive Disorder (MDD) is a major global health challenge and is commonly assessed through semi-structured interviews and questionnaires, which are clinically validated but partly subjective. Affective Computing has therefore explored digital biomarkers from multimodal behavioural data, including acoustic, visual, and linguistic cues. This paper evaluates a hierarchical deep learning framework for depression detection that combines a frozen DistilBERT encoder for text, Bi-LSTMs for temporal modelling of audio and video feature streams, and a Transformer-based fusion module to capture inter-modal dependencies. Using the Extended DAIC-WOZ corpus ( \(N{=}275\) ), we conduct a systematic ablation study across unimodal, bimodal, and multimodal variants. While the multimodal configuration attains strong performance (F1=0.992), text-enhanced variants approach near-ceiling results, suggesting potential shortcut learning and/or leakage effects in interview-based settings. Overall, the results highlight both the promise of multimodal fusion and the need for stricter leakage controls and more clinically grounded linguistic inputs to improve generalisability and interpretability for computer-aided screening.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multimodal Depression Detection Using Hierarchical Fusion of DistilBERT and Bi-LSTM Encoders: A Deep Learning Approach on the DAIC-WOZ Corpus

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Francisco Ortiz-Vidal,
  • Pedro Salcedo-Lagos,
  • Juan M. Górriz Sáez

摘要

Major Depressive Disorder (MDD) is a major global health challenge and is commonly assessed through semi-structured interviews and questionnaires, which are clinically validated but partly subjective. Affective Computing has therefore explored digital biomarkers from multimodal behavioural data, including acoustic, visual, and linguistic cues. This paper evaluates a hierarchical deep learning framework for depression detection that combines a frozen DistilBERT encoder for text, Bi-LSTMs for temporal modelling of audio and video feature streams, and a Transformer-based fusion module to capture inter-modal dependencies. Using the Extended DAIC-WOZ corpus ( \(N{=}275\) ), we conduct a systematic ablation study across unimodal, bimodal, and multimodal variants. While the multimodal configuration attains strong performance (F1=0.992), text-enhanced variants approach near-ceiling results, suggesting potential shortcut learning and/or leakage effects in interview-based settings. Overall, the results highlight both the promise of multimodal fusion and the need for stricter leakage controls and more clinically grounded linguistic inputs to improve generalisability and interpretability for computer-aided screening.