RatioWaveNet: A learnable wavelet-based framework for robust and interpretable electroencephalogram motor imagery classification

Decoding non-stationary electroencephalogram (EEG) signals remains a primary bottleneck in developing reliable Brain–Computer Interfaces (BCIs). Traditional deep learning approaches often struggle with complex spatiotemporal dynamics, low signal-to-noise ratios, and severe inter-subject variability. To address these challenges, we propose RatioWaveNet, a novel deep learning framework that integrates a Rational Dilated Wavelet Transform (RDWT) with a hybrid multi-window attention and Temporal Convolutional Network (TCN) architecture. Unlike conventional fixed-dilation transforms, our approach leverages learnable rational dilation factors to dynamically optimize time–frequency resolution for specific motor imagery rhythms. The pipeline combines dilated convolutional layers for spatial filtering, a multi-window mechanism merging multi-head and convolutional block attention for spatiotemporal dependency modeling, and a hierarchical TCN for sequential decoding. Comprehensive experiments across three public EEG benchmarks demonstrate that RatioWaveNet achieves statistically significant performance gains, yielding a 3%–7% accuracy improvement (p<0.01) over current state-of-the-art methods. Furthermore, rigorous cross-subject evaluations confirm enhanced model robustness and stable decision-making under severe subject shifts. Beyond raw accuracy improvements, RatioWaveNet offers a highly interpretable and mathematically grounded framework. The RDWT front-end provides physiologically meaningful feature representations — corroborated by scalogram visualizations — while the attention-TCN core ensures noise-resilient decoding. These advancements establish RatioWaveNet as a highly effective and transparent solution for high-fidelity EEG classification. The complete codebase and pretrained models are publicly released to promote reproducibility.