TEDNet: Cascaded CNN-transformer with dual attentions for taste EEG decoding

Background

Traditional taste evaluation methods suffer from subjective biases and limited sensor capabilities, while existing Electroencephalogram (EEG) approaches struggle to decode complex neural patterns evoked by sour, sweet, bitter, and salty stimuli due to noise sensitivity and inadequate multi-scale feature integration.

New method

To address this, we propose Taste EEG Decoding Network (TEDNet), a novel deep learning architecture integrating: 1) a Temporal Spatial Convolution Module (TSCM) capturing electrode-wise dependencies, 2) a Temporal Spatial Attention Module (TSAM) adaptively reweighting critical features, and 3) a Local Global Fusion Module (LGFM) combines the local features of taste EEG with the global ones.

Results

Evaluated on a well-controlled dataset containing 2400 EEG samples from 30 subjects, the accuracy of TEDNet is 98.92 %, the F1-score is 98.75 %, and the Kappa coefficient is 98.49 %.

Comparison with existing methods

While maintaining computational efficiency, TEDNet has surpassed the existing advanced convolution and self-attention methods.

Conclusions

This framework establishes a robust solution for objective taste perception decoding, advancing sensory evaluation in food science.