Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·24d ago·5 min read
When Consistency Becomes Bias: Interviewer Effects in Semi-Structured Clinical Interviews
TL;DR
Research reveals that automatic depression detection models often rely on interviewer prompts rather than participant language, leading to inflated performance metrics. By focusing on participant utterances, models can better reflect genuine linguistic cues and reduce bias.
✦ Why It Matters
Focus on participant language in depression detection models to improve accuracy and reduce bias.
Key Takeaways
- Analyze decision evidence by time and speaker.
- Train models exclusively on participant language.
- Avoid reliance on interviewer scripts for performance metrics.
- Ensure models reflect genuine linguistic cues.
How It Works
The study identifies that models trained on interviewer prompts can achieve high accuracy by leveraging fixed scripts rather than participant language. By focusing on participant utterances, the models distribute decision evidence more evenly, leading to more accurate and unbiased predictions.