Comparative Evaluation of Deep Learning and Transformer-Based Models for Depression Detection Using Clinical Interview Transcripts

The article is a comparative essay on the traditional deep learning structures, i.e., Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and Bidirectional Gated Recurrent Unit (BiGRU) as opposed to a transformer-based model ( DistilBERT + Logistic Regression ) regarding the automated detection of depression on the clinical interview transcripts. The controlled and standardized experiments are conducted using the Distress Analysis Interview Corpus- Wizard of Oz (DAIC-WOZ) data to be able to fairly compare the results. Findings indicate that DistilBERT + Logistic Regression is better than all deep learning baselines (F1 = 0.9506), whereas classical models are also competitive and less costly to run. The comparison of the transformer advantage with a significance test (McNemar test) shows that the transformer advantage cannot be explained by the chance (p < 0.05). The article describes the trade-offs between predictive accuracy and computational efficiency, ethical issues, and implications of the use of AI-based decision-support in digital mental health information systems.