Federated learning and deep learning framework for MRI image and speech signal-based multi-modal depression detection

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Adolescent Psychiatry

Background:

• Adolescent depression is a growing concern, leading to severe educational, social, and life-threatening risks.
• Early identification and treatment are crucial for managing adolescent mental health.
• Traditional deep learning models struggle with large datasets, necessitating efficient solutions.

Purpose of the Study:

• To develop and evaluate a federated learning (FL) framework for detecting depression in adolescents.
• To integrate multi-modal data, including speech signals and Magnetic Resonance Imaging (MRI), for enhanced detection accuracy.
• To propose an optimized deep learning model for depression identification within the FL framework.

Main Methods:

• A federated learning (FL) framework with local and global modules was employed for depression detection.
• The local module utilized an Exponential African Pelican Optimization-based Deep Convolutional Neural Network (ExpAPO-DCNN).
• Speech signals and MRI data were individually pre-processed, features extracted, and then fused using the overlap coefficient.

Main Results:

• The ExpAPO-DCNN model achieved a high accuracy of 98.00%.
• Performance metrics included a Loss of 0.023, RMSE of 0.058, and MSE of 0.240.
• The model demonstrated strong diagnostic capabilities with a True Negative Rate (TNR) of 97.90% and a True Positive Rate (TPR) of 96.30%.

Conclusions:

• The proposed FL framework effectively addresses data size challenges in deep learning for adolescent depression detection.
• The integration of speech and MRI data with the ExpAPO-DCNN model shows significant promise for accurate and early diagnosis.
• This approach offers a scalable and privacy-preserving solution for mental health monitoring in adolescents.