[Paper Review] DAIC-WOZ: On the Validity of Using the Therapist's prompts in Automatic Depression Detection from Clinical Interviews

Sergio Burdisso, Ernesto Reyes-Ramírez+4 more|arXiv (Cornell University)|Apr 22, 2024

Mental Health Research TopicsPsychology3 citations

TL;DR

This paper investigates the unintended bias in the DAIC-WOZ dataset arising from therapist prompts used in automatic depression detection. Through ablation and attention analysis, the authors show that models exploit these prompts as discriminative shortcuts—focusing on specific mental health probing questions—achieving a 0.90 F1 score by intentionally leveraging this bias, which undermines generalization to real-world clinical settings.

ABSTRACT

Automatic depression detection from conversational data has gained significant interest in recent years. The DAIC-WOZ dataset, interviews conducted by a human-controlled virtual agent, has been widely used for this task. Recent studies have reported enhanced performance when incorporating interviewer's prompts into the model. In this work, we hypothesize that this improvement might be mainly due to a bias present in these prompts, rather than the proposed architectures and methods. Through ablation experiments and qualitative analysis, we discover that models using interviewer's prompts learn to focus on a specific region of the interviews, where questions about past experiences with mental health issues are asked, and use them as discriminative shortcuts to detect depressed participants. In contrast, models using participant responses gather evidence from across the entire interview. Finally, to highlight the magnitude of this bias, we achieve a 0.90 F1 score by intentionally exploiting it, the highest result reported to date on this dataset using only textual information. Our findings underline the need for caution when incorporating interviewers' prompts into models, as they may inadvertently learn to exploit targeted prompts, rather than learning to characterize the language and behavior that are genuinely indicative of the patient's mental health condition.

Motivation & Objective

• To investigate whether the reported performance gains in automatic depression detection using therapist prompts are due to genuine diagnostic learning or dataset bias.
• To analyze how models trained with therapist prompts differ in attention allocation compared to models using only participant responses.
• To demonstrate that the high performance in existing studies may stem from exploiting a localized bias in prompts rather than learning meaningful mental health indicators.
• To highlight the ethical and practical risks of relying on such biased signals in real-world clinical AI systems.
• To advocate for more robust evaluation protocols that account for prompt-based shortcut learning in clinical NLP benchmarks.

Proposed method

• Conducted ablation studies comparing models trained on participant responses only versus those incorporating therapist prompts.
• Used attention visualization techniques to analyze where models focus during inference, identifying concentration on specific prompt regions.
• Trained a model to intentionally exploit the prompt bias by emphasizing responses to mental health probing questions.
• Evaluated performance using standard metrics (F1 score) on the DAIC-WOZ test split, focusing on textual modality.
• Performed qualitative analysis of attention maps and model behavior across different interview segments.
• Replicated the analysis on the full interview sequence to compare global vs. localized evidence gathering.

Experimental results

Research questions

• RQ1 Does incorporating therapist prompts in depression detection models lead to improved performance due to genuine diagnostic signal or due to dataset bias?
• RQ2 How do models that use therapist prompts differ in attention distribution compared to models using only participant responses?
• RQ3 Can a model achieve high performance by solely exploiting the structure and content of therapist prompts?
• RQ4 To what extent do models relying on prompts learn to focus on specific interview segments, particularly those containing mental health probing questions?
• RQ5 How does the presence of prompt-based bias affect the generalization and reliability of automatic depression detection systems in real-world settings?

Key findings

• Models using therapist prompts learn to focus their attention on a specific, localized region of the interview—primarily the latter portion where mental health-related questions are asked.
• In contrast, models trained only on participant responses gather evidence from across the entire conversation, showing more distributed and contextually informed attention.
• By intentionally exploiting the prompt bias, the authors achieved a 0.90 F1 score on the DAIC-WOZ dataset, the highest reported result to date using only textual information.
• The performance gain from using prompts is largely attributable to a strong, exploitable bias in the prompt content rather than to improved model architecture or feature learning.
• The findings suggest that current high-performing models may be learning discriminative shortcuts based on interviewer behavior rather than patient language or behavior indicative of depression.
• The study calls into question the validity of benchmark results on DAIC-WOZ when therapist prompts are used, urging caution in model evaluation and deployment.