Scholarlink

<p>This study focuses on improving the accuracy of depression level identification in current prediction methods. It addresses model misclassification issues resulting from inadequate structured modeling of subjects’ psychological states. The study proposes a deep learning depression prediction model integrating psychological features and Psychology-Guided Network (PsyGuide-Net) to enhance the model’s accuracy and practicality in identifying depressive tendencies. The study builds on the DAIC-WOZ and AVEC 2014 datasets to construct multimodal input samples containing speech spectrograms, text semantic vectors, and standardized psychological features. It uses clinical Patient Health Questionnaire-8 (PHQ-8) scores as supervision signals for training and evaluation. Under 5-fold cross-validation, the proposed model achieves an accuracy of 0.848, F1-score of 0.778, and AUC-ROC of 0.898 in classification tasks. For PHQ-8 score prediction, this model reaches a Pearson correlation coefficient of 0.732, demonstrating a stable capability to judge depression severity. Additionally, the model has an average inference time of 106 milliseconds, a robustness test performance retention rate exceeding 95%, and an F1-score of 0.878 in the extremely severe depression group. The results show that while maintaining high inference speed, PsyGuide-Net balances sensitive identification of high-risk populations and has strong generalization ability and scene adaptability. Furthermore, the psychological feature gating mechanism provides an interpretable basis for the prediction results. This enables the model to reveal the role of personality and emotional variables in identifying depression, enhancing its application potential in clinical screening and intervention scenarios. This study offers a more reliable and valuable intelligent depression detection tool for psychological counselors, campus psychological service institutions, and telemedicine platforms.</p>

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

Depression prediction model based on deep learning and psychological feature extraction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月16日

查看全文
  • Weiwei Su

摘要

This study focuses on improving the accuracy of depression level identification in current prediction methods. It addresses model misclassification issues resulting from inadequate structured modeling of subjects’ psychological states. The study proposes a deep learning depression prediction model integrating psychological features and Psychology-Guided Network (PsyGuide-Net) to enhance the model’s accuracy and practicality in identifying depressive tendencies. The study builds on the DAIC-WOZ and AVEC 2014 datasets to construct multimodal input samples containing speech spectrograms, text semantic vectors, and standardized psychological features. It uses clinical Patient Health Questionnaire-8 (PHQ-8) scores as supervision signals for training and evaluation. Under 5-fold cross-validation, the proposed model achieves an accuracy of 0.848, F1-score of 0.778, and AUC-ROC of 0.898 in classification tasks. For PHQ-8 score prediction, this model reaches a Pearson correlation coefficient of 0.732, demonstrating a stable capability to judge depression severity. Additionally, the model has an average inference time of 106 milliseconds, a robustness test performance retention rate exceeding 95%, and an F1-score of 0.878 in the extremely severe depression group. The results show that while maintaining high inference speed, PsyGuide-Net balances sensitive identification of high-risk populations and has strong generalization ability and scene adaptability. Furthermore, the psychological feature gating mechanism provides an interpretable basis for the prediction results. This enables the model to reveal the role of personality and emotional variables in identifying depression, enhancing its application potential in clinical screening and intervention scenarios. This study offers a more reliable and valuable intelligent depression detection tool for psychological counselors, campus psychological service institutions, and telemedicine platforms.