WavTSK: An Interpretable Fuzzy Network with Learnable Wavelet-Based Feature Extraction for Motor Imagery EEG Decoding

Decoding motor imagery (MI) from electroencephalogram (EEG) signals is a cornerstone of brain-computer interface (BCI) systems. However, existing methods often face a critical trade-off between decoding accuracy and model interpretability, limiting their applicability in real-world settings. To address this challenge, we proposed the WavTSK, an end-to-end interpretable fuzzy neural network for efficient MI-EEG decoding. The WavTSK employs a deeply integrated architecture that combines a learnable wavelet-based feature extraction module with a multi-rule Takagi-Sugeno-Kang (TSK) fuzzy classifier. The feature extractor incorporates learnable wavelet filters, statistical descriptors, and an adaptive band-weighting mechanism to capture rich multi-scale time-frequency representations directly from raw EEG. The extracted representations are then processed by a TSK fuzzy reasoning layer, enabling rule-level transparency in decoding. Experiments on the four-class BCI Competition IV-2a dataset showed that WavTSK achieved mean decoding accuracies of 71.63% in cross-validation and 70.55% in the cross-session hold-out setting, consistently outperforming state-of-the-art black-box deep learning models while maintaining strong interpretability. The results highlight the potential of WavTSK as a powerful and interpretable framework for reliable EEG decoding, advancing the development of trustworthy and clinically applicable BCI technologies.