Deep learning with ensemble-based hybrid AI model for bipolar and unipolar depression detection using demographic and...

Area of Science:

• Artificial Intelligence
• Computational Psychiatry
• Machine Learning in Mental Health

Background:

• Depression, encompassing Bipolar and Unipolar types, is a prevalent mental health condition.
• Current diagnostic methods for depression are subjective, risking bias and underreporting.
• Machine learning (ML) and deep learning (DL) present automated solutions for depression detection using behavioral and demographic data.

Purpose of the Study:

• To develop a hybrid AI framework for classifying Bipolar and Unipolar Depression.
• To integrate structured demographic data with synthetic actigraph time-series data for enhanced depression detection.
• To improve the accuracy and interpretability of depression classification models.

Main Methods:

• A hybrid AI framework combining XGBoost for demographic data and a deep convolutional neural network (CNN) for time-series data.
• Utilized stratified k-fold cross-validation and hyperparameter tuning for robust model training.
• Employed SHAP and Grad-CAM for enhanced model explainability, identifying key predictive features and temporal patterns.

Main Results:

• The hybrid model achieved strong performance in accuracy, sensitivity, and specificity for depression classification.
• Integration of temporal and static features significantly improved the prediction of Bipolar and Unipolar Depression.
• Interpretability techniques successfully highlighted crucial features and time-related patterns influencing model predictions.

Conclusions:

• Introduced a robust and interpretable AI framework for depression classification using synthetic multimodal data.
• The developed model serves as a methodological foundation for future research involving real-world clinical datasets.
• While not yet clinically validated, the framework demonstrates the potential of AI in objective depression assessment.