A multimodal depression recognition method based on EEG-fNIRS-SDS

Traditional depression diagnosis relies heavily on clinical interviews and subjective questionnaires, leading to concerns about subjectivity and limited accuracy. While recent advances in multimodal physiological signal analysis (using data such as EEG, fNIRS, and eye-tracking) have shown promise, existing fusion approaches often suffer from inflexible weight allocation, feature redundancy, and inadequate cross-modal interaction.

We propose a novel Mutual Information Maximization-based Weighted Multimodal Fusion Network (MI-WNet). This network is designed to synergistically analyze EEG, fNIRS, and clinical scale data. Its core innovation lies in leveraging mutual information maximization to automate and optimize modality weighting, effectively filter redundant features, and enhance meaningful cross-modal interactions.

Experimental results demonstrate that the proposed MI-WNet model achieves a high depression recognition accuracy of 95.52% ± 0.42%.

The model's performance shows a significant improvement over baseline models that rely on single modalities (EEG, fNIRS, or clinical data alone). This underscores the superiority of our adaptive fusion strategy over traditional methods with fixed weighting schemes.

The MI-WNet framework presents a robust and effective solution for automated depression recognition. By addressing key limitations in existing multimodal fusion techniques, it significantly enhances diagnostic accuracy and paves the way for more objective and data-driven clinical decision support tools.