Boosted Convolutional Neural Networks for Motor Imagery EEG Decoding with Multiwavelet-based Time-Frequency Conditional Granger Causality Analysis

Decoding EEG signals of different mental states is a challenging task for brain-computer interfaces (BCIs) due to nonstationarity of perceptual decision processes. This paper presents a novel boosted convolutional neural networks (ConvNets) decoding scheme for motor imagery (MI) EEG signals assisted by the multiwavelet-based time-frequency (TF) causality analysis. Specifically, multiwavelet basis functions are first combined with Geweke spectral measure to obtain high-resolution TF-conditional Granger causality (CGC) representations, where a regularized orthogonal forward regression (ROFR) algorithm is adopted to detect a parsimonious model with good generalization performance. The causality images for network input preserving time, frequency and location information of connectivity are then designed based on the TF-CGC distributions of alpha band multichannel EEG signals. Further constructed boosted ConvNets by using spatio-temporal convolutions as well as advances in deep learning including cropping and boosting methods, to extract discriminative causality features and classify MI tasks. Our proposed approach outperforms the competition winner algorithm with 12.15% increase in average accuracy and 74.02% decrease in associated inter subject standard deviation for the same binary classification on BCI competition-IV dataset-IIa. Experiment results indicate that the boosted ConvNets with causality images works well in decoding MI-EEG signals and provides a promising framework for developing MI-BCI systems.

💡 Research Summary

**
The paper proposes a novel framework for decoding motor‑imagery (MI) electroencephalogram (EEG) signals by integrating high‑resolution time‑frequency conditional Granger causality (TF‑CGC) with a boosted convolutional neural network (ConvNet). The authors first model the non‑stationary multichannel EEG using time‑varying autoregressive with exogenous inputs (TVARX) representations built on multi‑wavelet basis functions. Multi‑wavelets provide a flexible, smooth approximation of time‑varying coefficients, while the regularized orthogonal forward regression (ROFR) algorithm selects a parsimonious subset of basis functions, preventing over‑fitting and ensuring good generalization.

With the identified TVARX model, the Geweke spectral causality measure is extended to a conditional, time‑frequency domain, yielding TF‑CGC maps that capture directed interactions between any pair of channels while conditioning on the remaining channels. The authors focus on the alpha band (8‑13 Hz), where motor‑related rhythms are most prominent, and convert the TF‑CGC values into “causality images”. In these images, the horizontal axis encodes time, the vertical axis encodes frequency, and the color (or channel‑wise stacking) encodes spatial location, thereby preserving the three‑dimensional information (time, frequency, space) in a 2‑D representation suitable for convolutional processing.

The second major contribution is a boosted ConvNet architecture that consumes the causality images. The network combines spatial 2‑D convolutions (to learn inter‑electrode patterns) with temporal 1‑D convolutions (to capture the evolution of connectivity over the trial). To mitigate the limited training data typical of BCI studies, two deep‑learning tricks are employed: (i) cropping, which extracts multiple overlapping sub‑windows from each trial, effectively augmenting the dataset; and (ii) boosting, which trains a series of weak learners on different crops and aggregates their predictions, thereby improving robustness and reducing variance. The overall model remains relatively shallow, keeping the number of trainable parameters modest while still exploiting multi‑scale features.

The method is evaluated on the publicly available BCI Competition‑IV dataset IIa (four subjects, binary left‑ vs. right‑hand MI). Compared with the competition winner (average accuracy ≈ 71 % and inter‑subject standard deviation ≈ 0.12), the proposed boosted ConvNet achieves an average accuracy of 84 % (a 12.15 % absolute gain) and reduces the standard deviation to 0.03 (a 74 % reduction). Additional baselines—including common spatial patterns (CSP), filter‑bank CSP (FBCSP), directed transfer function (DTF), and partial Granger causality (PGC)—are also outperformed, confirming that TF‑CGC‑derived images provide more discriminative and subject‑invariant features than traditional power‑based or static connectivity measures.

Key strengths of the work are: (1) a rigorous, model‑based extraction of dynamic, frequency‑specific directed connectivity using multi‑wavelet TVARX and ROFR; (2) an innovative transformation of these connectivity patterns into images that can be processed by deep learning; (3) the use of data‑augmentation (cropping) and ensemble (boosting) techniques to alleviate the small‑sample problem common in BCI; and (4) a substantial empirical gain on a benchmark dataset, demonstrating both higher accuracy and lower inter‑subject variability.

Limitations include the focus on a single frequency band (alpha), the lack of experiments on multi‑class MI tasks, and the absence of a detailed analysis of computational load for real‑time deployment. Future directions suggested by the authors involve extending the approach to multiple frequency bands, integrating other connectivity metrics (e.g., phase‑locking value, transfer entropy), and developing online adaptive versions of the model to enable practical, real‑time MI‑BCI applications. Overall, the paper presents a compelling combination of advanced signal‑processing theory and modern deep‑learning practice, offering a promising pathway toward more reliable and interpretable motor‑imagery brain‑computer interfaces.