Feature-based enhanced boosting algorithm for depression detection.

Area of Science:

• Computational psychiatry
• Artificial intelligence in mental health

Background:

• Depression is a growing mental health concern impacting daily life.
• Machine learning, particularly deep learning, shows promise for early depression detection using social media.
• Existing boosting algorithms face challenges with complex features, weak learner enhancement, and large datasets.

Purpose of the Study:

• To develop a novel feature-based enhanced boosting algorithm (F-EBA) for improved depression detection.
• To enhance the performance of weak learners and handle large datasets effectively.
• To increase the accuracy and interpretability of depression detection models.

Main Methods:

• Developed a two-pipeline F-EBA model: feature engineering and classification.
• Utilized WordVec and BERT embeddings, attention mechanisms, and feature elimination for feature optimization.
• Implemented a weight maximization strategy for weak learners and an adversarial layer for data robustness.

Main Results:

• The F-EBA model achieved 95% accuracy on 46 million records, enhancing weak learner performance.
• Feature optimization significantly improved model accuracy and interpretability.
• An adversarial layer increased accuracy to approximately 97%, surpassing prior studies.
• Optimized feature sets boosted baseline classifier performance.

Conclusions:

• The F-EBA model represents a significant advancement in detecting depression from social media data.
• The proposed methods enhance accuracy, interpretability, and robustness in computational psychiatry.
• This approach offers a powerful tool for early depression identification and intervention.