Ensembles of BERT for Depression Classification.

Area of Science:

• Computational psychiatry
• Natural Language Processing (NLP) in mental health
• Machine learning for clinical diagnostics

Background:

• Depression is a prevalent global mental health disorder with increasing incidence.
• Early depression detection is crucial for effective treatment and improved prognosis.
• Current methods for depression detection face challenges, necessitating advanced approaches.

Purpose of the Study:

• To evaluate the efficacy of ensemble methods using Bidirectional Encoder Representations from Transformers (BERT) variants for depression detection.
• To compare the performance of individual BERT variants against various ensemble strategies.
• To assess the impact of ensemble composition (strategies, components, layers) on classification accuracy.

Main Methods:

• Utilized transcripts from responses to 12 clinical interview questions.
• Implemented and compared three individual BERT variants.
• Developed and tested four distinct ensembles of BERT variants, varying ensemble strategies, component numbers, and architectural layer combinations.

Main Results:

• Ensemble models demonstrated increased mean F1 scores compared to individual BERT variants.
• Ensemble approaches significantly improved the robustness of depression classification across clinical interview data.
• Specific ensemble configurations showed superior performance in identifying depression from textual responses.

Conclusions:

• Ensembles of BERT variants offer a promising advancement for detecting depression using text data.
• This research supports the development of novel healthcare applications for mental health screening and diagnosis.
• The findings underscore the potential of advanced NLP techniques in clinical settings for mental health assessment.