Cognitively Inspired Federated Learning Framework for Interpretable and Privacy-Secured EEG Biomarker Prediction of...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Depression relapse is a significant challenge in long-term care settings.
• Predicting and preventing relapse is crucial for patient well-being and effective treatment.

Purpose of the Study:

• To develop a privacy-aware, explainable, personalized federated learning (PFL) framework for predicting depression relapse.
• To provide patient-specific, interpretable predictions using electroencephalogram (EEG) data.

Main Methods:

• Utilized a PFL framework incorporating layer-wise relevance propagation and Shapley value analysis.
• Analyzed resting-state 128-channel EEG data from 100 subjects in the Healthy Brain Network (HBN) dataset.
• Employed 10-fold cross-validation and jointly modeled multi-channel EEG features with standardized symptom scales.

Main Results:

• Achieved high performance metrics: 92% accuracy, 91% precision, 93% recall, and 90.5% F1-score.
• Generated attribution maps identifying region-anchored spectral patterns linked to relapse risk.
• Demonstrated a privacy-aware federated setup suitable for multi-site deployment.

Conclusions:

• The PFL framework offers a scalable and clinically feasible approach for trustworthy depression relapse monitoring.
• The model provides interpretable insights for early intervention and personalized patient care.
• The privacy-preserving nature facilitates broader adoption in multi-institutional settings.