Optimal Time Window and Frequency Bandwidth Parameter Combination for Subject-Specific Motor Imagery EEG Classification

Technical Summary: Optimal Time Window and Frequency Bandwidth Parameter Combination for Subject-Specific Motor Imagery EEG Classification

Problem Statement
Motor Imagery (MI) Brain-Computer Interfaces (BCIs) rely on decoding EEG signals, specifically oscillations in the sensorimotor cortex (alpha/mu and beta waves), to interpret intended movement. While supervised machine learning techniques like Common Spatial Patterns (CSP) combined with Linear Discriminant Analysis (LDA) are standard for this task, classification performance varies significantly across subjects. A primary source of this variability is the lack of consensus on the optimal temporal (post-cue time window) and spectral (frequency bandwidth) parameters used in training protocols. Previous research has investigated these parameters in isolation, but the combinatory effect of simultaneous optimization on MI classification performance has not been systematically studied. This study addresses the gap in identifying the optimal parameter combinations for subject-specific models.

Methodology
The study utilized the PhysioNet EEG Motor Movement/Imagery Dataset, comprising data from 109 participants performing left-hand and right-hand MI tasks. The experimental protocol involved 12 runs per participant, with analysis focused on runs 4, 8, and 12.

The research employed a grid search approach to iteratively train and test subject-specific models across various parameter combinations:

• Time Windows: Five options were tested based on literature: 0–2.25 s, 0–4 s, 0.5–2.5 s, 0.5–3.5 s, and 1–3.5 s (relative to the cue).
• Frequency Bandwidths: Two sets were evaluated:
  1. Six literature-derived bands (e.g., theta-alpha, alpha, beta, broadband).
  2. A sliding window set with a 4 Hz window size, 2 Hz step size, ranging from 4 Hz to 40 Hz (yielding 23 specific bands).
• Pipeline: For each combination, features were extracted using CSP (compressing 64 channels to 4 dimensions) and classified using LDA with 10-fold cross-validation. Shrinkage regularization was applied to prevent overfitting.
• Evaluation: Performance was measured using Accuracy and Cohen's Kappa (to account for class imbalance). Statistical significance was determined using Repeated Measures ANOVA with Bonferroni correction.

Key Contributions

1. Systematic Combinatory Analysis: This work is the first to the authors' knowledge to systematically assess the simultaneous optimization of time windows and frequency bandwidths for MI classification, moving beyond isolated parameter studies.
2. Subject-Specific vs. Population Averages: The study provides a dual perspective, identifying a population-level optimal combination while simultaneously mapping the unique optimal parameters for individual subjects, highlighting the extent of inter-subject variability.
3. Statistical Validation: The use of rigorous statistical testing (ANOVA) confirms that specific time windows and frequency bands yield significantly different classification accuracies, validating the necessity of parameter tuning.

Results

• Population-Level Optima: The highest average classification accuracy across the 109 subjects was achieved with a time window of (0, 4) seconds and a frequency bandwidth of (4, 12) Hz. The lowest accuracy occurred with the (1, 3.5) s window and (36, 40) Hz bandwidth.
• Statistical Significance: Repeated measures ANOVA revealed statistically significant differences ($p < 0.001$) between different time windows and frequency bandwidths. Specifically, the (0, 4) s and (0.5, 2.5) s windows showed significantly higher accuracies than the (1, 3.5) s window. In the frequency domain, the (4, 12) Hz band significantly outperformed narrower bands like (4, 8) Hz and (6, 10) Hz.
• Individual Variability: While (0, 4) s and (4, 12) Hz were optimal on average, individual subjects exhibited diverse optimal combinations. For instance, one participant achieved perfect classification (100% accuracy) using a (14, 40) Hz bandwidth and a (0.5, 2.5) s window. The most frequent optimal windows for individuals were (0, 2.25) s and (0, 4) s, while the highest individual average accuracies were found in the (1, 3.5) s window.
• Metric Consistency: Trends in Cohen's Kappa closely mirrored those of Accuracy, confirming the robustness of the findings despite potential class imbalances in individual models.

Significance and Claims
The paper claims that while a general "optimal" combination exists for the population—specifically the (0, 4) s window and (4, 12) Hz bandwidth—relying solely on these parameters may not yield the best results for every individual. The findings underscore the efficacy of personalized models that detect subject-specific optimal temporal and spectral parameter combinations.

The authors emphasize that the variability in optimal parameters reflects inherent differences in MI-induced cortical activity across subjects. Consequently, they argue for a user-centered design in MI-BCI systems. However, the paper also modestly notes the limitations of the personalized approach, specifically the constraint of limited training data per participant (approx. 90 trials), which may restrict model performance. The study suggests that while personalized tuning is effective, future work should explore the trade-off between personalized and generalized approaches, potentially utilizing deep learning or transfer learning to leverage larger datasets while maintaining subject specificity. The authors conclude that their findings highlight the need for adaptable models that can adjust to changing optimal parameters for individual users over time.