Depression detection through dual-stream modeling with large language models: a fusion-based transfer learning framework integrating BERT and T5 representations

Millions of people around the world suffer from depression. While early diagnosis is essential for timely intervention, it remains a significant challenge due to limited access to clinically diagnosed data and privacy restrictions on mental health records. These limitations hinder the training of robust AI models for depression detection. To tackle this, this article proposes a parallel transfer learning framework for depression detection that integrates BERT and T5 through a fusion mechanism, combining the complementary advantages of these two large language models (LLMs). By integrating their semantic embeddings, the method captures a broader range of linguistic cues from transcribed speech. These embeddings are processed through a model with two parallel branches: a one-dimensional convolutional neural network and a dense neural network are used to construct each branch for preliminary prediction, which are then fused for final prediction. Evaluations on the E-DAIC dataset demonstrate that the proposed method outperforms baseline models, achieving a 3.0% increase in accuracy (91.3%), a 6.9% increase in precision (95.2%), and a 1.7% improvement in F1-score (90.0%). The experimental results verify the effectiveness of BERT and T5 fusion in enhancing depression detection performance and highlight the potential of transfer learning for scalable and privacy-conscious mental health applications.