Differentially Private Multimodal Laplacian Dropout (DP-MLD) for EEG representative learning.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Multimodal electroencephalogram (EEG) learning shows potential for disease detection.
• Privacy concerns in clinical studies necessitate robust protection methods like differential privacy (DP).
• DP has not been extensively applied to complex multimodal EEG data.

Purpose of the Study:

• To propose a novel Differentially Private Multimodal Laplacian Dropout (DP-MLD) scheme for privacy-preserving multimodal EEG learning.
• To develop a multimodal representative learning model integrating language models and vision transformers with cross-attention.
• To introduce an adaptive feature-level Laplacian dropout for dynamic privacy budget optimization.

Main Methods:

• Utilized language models for EEG data (text) and vision transformers for other modalities (images).
• Implemented cross-attention mechanisms for effective feature extraction and integration.
• Designed an adaptive feature-level Laplacian dropout for dynamic randomness allocation within privacy budgets.

Main Results:

• Achieved an approximate 4% improvement in classification accuracy on a Freezing of Gait (FoG) dataset for Parkinson's Disease (PD).
• Demonstrated state-of-the-art performance in multimodal EEG learning under differential privacy.
• Validated the effectiveness of the DP-MLD scheme on an open-source multimodal dataset.

Conclusions:

• The proposed DP-MLD scheme effectively balances privacy protection and performance in multimodal EEG analysis.
• This method offers a promising solution for privacy-preserving disease detection using complex neuroimaging data.
• DP-MLD advances the field of secure and accurate analysis of multimodal clinical data.