Investigating the Impact of Rational Dilated Wavelet Transform on Motor Imagery EEG Decoding with Deep Learning Models

TL;DR

This study tackles the challenge of robust motor-imagery EEG decoding under non-stationarity by evaluating Rational Discrete Wavelet Transform (RDWT) as a plug-in time–frequency preprocessing step before deep learning models. It systematically tests RDWT across four backbones (EEGNet, ShallowConvNet, MBEEG_SENet, EEGTCNet) and three MI-EEG datasets (BCI-IV-2a, BCI-IV-2b, High Gamma), reporting subject-wise accuracy and Cohen’s $\kappa$. The findings show architecture- and dataset-dependent gains, with the largest improvements for temporally expressive backbones on BCI-IV-2a (e.g., EEGTCNet +4.44 pp) and smaller, but consistent gains on BCI-IV-2b and HGD, including notable subject-level improvements in challenging sessions. RDWT emerges as a low-overhead, architecture-aware preprocessing option that can enhance model robustness and agreement in non-ideal recording conditions, albeit not universally. The work provides practical guidance for integrating RDWT in MI-EEG pipelines and suggests future directions for data-driven parameter selection and adaptive activation of RDWT to maximize benefits.

Abstract

The present study investigates the impact of the Rational Discrete Wavelet Transform (RDWT), used as a plug-in preprocessing step for motor imagery electroencephalographic (EEG) decoding prior to applying deep learning classifiers. A systematic paired evaluation (with/without RDWT) is conducted on four state-of-the-art deep learning architectures: EEGNet, ShallowConvNet, MBEEG\_SENet, and EEGTCNet. This evaluation was carried out across three benchmark datasets: High Gamma, BCI-IV-2a, and BCI-IV-2b. The performance of the RDWT is reported with subject-wise averages using accuracy and Cohen's kappa, complemented by subject-level analyses to identify when RDWT is beneficial. On BCI-IV-2a, RDWT yields clear average gains for EEGTCNet (+4.44 percentage points, pp; kappa +0.059) and MBEEG\_SENet (+2.23 pp; +0.030), with smaller improvements for EEGNet (+2.08 pp; +0.027) and ShallowConvNet (+0.58 pp; +0.008). On BCI-IV-2b, the enhancements observed are modest yet consistent for EEGNet (+0.21 pp; +0.044) and EEGTCNet (+0.28 pp; +0.077). On HGD, average effects are modest to positive, with the most significant gain observed for MBEEG\_SENet (+1.65 pp; +0.022), followed by EEGNet (+0.76 pp; +0.010) and EEGTCNet (+0.54 pp; +0.008). Inspection of the subject material reveals significant enhancements in challenging recordings (e.g., non-stationary sessions), indicating that RDWT can mitigate localized noise and enhance rhythm-specific information. In conclusion, RDWT is shown to be a low-overhead, architecture-aware preprocessing technique that can yield tangible gains in accuracy and agreement for deep model families and challenging subjects.

Figures (4)

• Figure 1: Comparison of EEG signals and scalograms (Left Hand Movement, Cz, Trial 2) for subjects S1–S4. Left: time-domain signals; right: time-frequency maps. The spectral differences reflect inter-subject variability in motor-related activity.
• Figure 2: Scalograms for S1–S3 across channels Cz, C3, C4 during left-hand motor imagery. The high spectral similarity across channels suggests redundancy, supporting the choice of a single representative channel for further analysis.
• Figure 3: Scalogram visualization of the first trial for each motor imagery class (left hand, right hand, foot, tongue) in four BCI-IV-2a subjects: S3 and S7 (best classification) vs. S2 and S6 (worst). Rows: subjects; columns: classes. The absence of consistent patterns highlights the limited discriminative power of frequency-domain features.
• Figure 4: Inter-subject correlation matrices (Trial 3, Cz) for four motor imagery tasks. Matrices show pairwise maximum normalized cross-correlations between S3, S7, S2, and S6. Higher off-diagonal values indicate greater cross-subject similarity.