Bias Identification in Deep Learning-Based Psychological Hotline Models*

This study examines gender and emotion-related biases in deep learning-based suicide risk prediction models, leveraging psychological hotline voice data to quantify these biases. Using 1,275 Mandarin hotline recordings from Beijing Huilongguan Hospital, we trained six deep learning models. Voice features were extracted using Whisper2, and counterfactual datasets (varying in gender and emotion) were generated through ChatTTS and CosyVoice. Biases were assessed using metrics such as Mean Error, F1-score, and Accuracy. The results revealed significant systematic biases.the findings highlight the impact of sociocultural factors and data imbalances on model bias.