DepAI – AI-driven depression detection

Depression is a major global health concern, yet existing detection methods rely heavily on self-reports, which can be subjective and prone to bias. DepAI explores innovative AI-driven methodologies for early depression detection using speech and text analysis. By leveraging deep learning and privacy-preserving techniques, DepAI aims to create a more reliable and accessible solution for mental health assessment.

Key Objectives

Enhancing Depression Detection in Data-Scarce Environments

Enhancing signal quality, ensuring robustness in low-resource settings, and improving model effectiveness through multimodal learning.

Privacy-Preserving Multimodal Depression Detection

Ensuring privacy without compromising the discriminative power of depression-related linguistic and acoustic features in multimodal models.

Mitigating Gender Bias in Speech-Based Depression Detection

Enhancing fairness and reliability in depression detection models by ensuring balanced performance across genders without compromising overall accuracy.

Methodology

DepAI integrates an adaptive data selection approach with pre-trained transformer models for feature extraction, followed by classification models for depression detection. The system is designed to:

  • Select relevant speech segments dynamically to improve classification robustness.
  • Extract meaningful speech and text representations using Whisper and DepRoBERTa.
  • Implement bias-aware training and privacy-preserving techniques for ethical AI development.

Datasets

  • E-DAIC dataset (Preliminary Experiments)
  • MoodAI dataset (Proposed implementations)

Expected Outcomes

  • A reliable, multimodal AI system for early depression detection using speech and text.
  • Improved classification performance even in data-scarce conditions through adaptive selection and transfer learning.
  • A privacy-aware framework ensuring speaker de-identification while maintaining diagnostic accuracy.

Future Directions

  • Enhancing generalizability across diverse demographics and real-world conditions.
  • Exploring adversarial learning and federated learning to improve model adaptability.
  • Expanding privacy-preserving techniques to further secure sensitive mental health data.