Construction of a transfer learning-based depression detection model for female breast cancer patients: text sentiment analysis.

BACKGROUND

Social networks have become a vital space for breast cancer (BC) patients to share deeply personal emotions they might avoid expressing in real life. However, the unstructured and vast nature of these textual expressions poses challenges for manual analysis. To address this, our research team employed transfer learning methods to efficiently process and analyze large-scale text data for depression detection.

OBJECTIVE

This study seeks to address the emotional struggles faced by women with BC, who often grapple with depression but lack accessible mental health support. This study aims to develop a transfer learning-based model to enable timely, non-invasive identification of depression through patients' self-expressed texts, thereby offering a pathway to early intervention.

METHODS

A mixed-methods framework integrated qualitative content analysis with deep learning. We recruited 300 BC patients (inpatients and online users). Depression status was assessed via the Self-rating Depression Scale (SDS), followed by collection and preprocessing of their self-expressed texts. Texts were manually annotated for depression scores/status, and formed a corpus. Content analysis was used to explore linguistic features. A BERT-based model pre-trained on a Weibo depression corpus was fine-tuned using clinical texts. Performance was evaluated via five-fold cross-validation, adversarial testing (word replacement, misspelling, deletion), and ablation studies. Model performance was evaluated using accuracy, precision, recall, and F1-score. The model was validated by robustness analysis and ablation studies.

RESULTS

Participants were grouped into depressive (n = 88) and non-depressive (n = 212) cohorts, while financial burden (P = 0.025) and advanced cancer stage (P = 0.038) correlated with depression. Content analysis revealed significant differences in negative life attitudes (P < 0.05). The transfer learning model achieved 86.67% accuracy (F1-score = 0.79). The model demonstrated robustness to semantic noise but required spelling correction for clinical deployment.

CONCLUSION

This study established a culturally adapted detection framework. By combining social media pre-training and clinical fine-tuning, the model enables scalable, non-invasive depression screening, bridging cultural barriers to emotional disclosure. Future work should expand demographic diversity and integrate multimodal data for enhanced clinical utility.