Go to ICLR 2025 Conference homepageMultimodal Depression Detection with Contextual Position Encoding and Latent Space Regularization
Clinical interviews are the gold standard for detecting depression, and previous work using multimodal features from participants' audio, transcribed text, and video have shown promising results. Recent approaches further improve performance by incorporating an additional textual modality—the interviewer’s prompts—during training. However, these approaches risk introducing biases, as models may over-rely on specific prompt-response pairs, which may not always be present in real-world settings. This leads to models exploiting these cues as shortcuts for detecting depression rather than learning the language and behaviors that genuinely indicate the subject's mental health, ultimately undermining consistency and objectivity. To address this, we propose a novel approach that combines Contextual Position Encoding (CoPE) and Latent Space Regularization (LSR), leveraging both subjects' responses (audio) and the interviewer's prompts (text). CoPE captures the evolving context of the interview, ensuring that the model utilizes insights from the entire conversation, preventing over-reliance on isolated or late-stage cues. This helps the model understand interactions holistically and more accurately reflect mental health indicators. LSR introduces constraints to enforce consistency in the model’s learned representations, reducing overfitting to superficial cues and guiding the model toward more generalizable patterns. By smoothing the latent space, LSR helps the model focus on meaningful, high-level representations of both audio and text. Our approach yields competitive results on the DAIC-WOZ benchmark and surpasses the state-of-the-art on the EATD benchmark. The code is released.