Towards correlation-based time window selection method for motor imagery BCIs

Jiankui Feng, Erwei Yin, Jing Jin, Rami Saab, Ian Daly, Xingyu Wang, Dewen Hu, Andrzej CichockiNeural Networks 2018 · DOI

TL;DR

This paper introduces a novel "correlation-based time window selection" (CTWS) algorithm for motor imagery (MI)-based brain-computer interfaces (BCIs). By dynamically adjusting the time window used for feature extraction based on trial-to-trial variations, CTWS significantly improves classification accuracy compared to traditional fixed-window approaches, particularly for stroke patients.

What Problem Does This Paper Address?

Motor imagery (MI) BCIs translate imagined movements into control signals. A crucial step is selecting the optimal time window within the EEG signal to extract relevant features. However, the precise timing of an MI event can vary significantly between trials for a given individual. Relying on a fixed time window, often determined by the cue onset, can lead to the inclusion of irrelevant or noisy data, thereby reducing classification accuracy. Existing methods for time window selection are often manual, time-consuming, and may not achieve consistent performance. This paper addresses the challenge of automatically and effectively adapting the feature extraction window to the inherent variability of MI signals.

Methods

The core of this work is the Correlation-based Time Window Selection (CTWS) algorithm. Here's a breakdown of its key components:

  • Optimized Reference Signal Generation: For each class (e.g., imagining left vs. right hand movement), the algorithm first averages the EEG signals from all training trials to create initial reference signals. These signals are then iteratively refined by analyzing correlations with new time windows and evaluating classification performance. This process aims to find the most representative signal for each MI class.
  • Dynamic Time Window Adjustment: The CTWS algorithm then uses these optimized reference signals to adjust the starting point of the feature extraction time window for each individual trial. This is achieved by calculating the correlation between the trial's EEG data and the reference signals. The time window that yields the highest correlation is selected. This process is applied to both training and testing data.
  • Feature Extraction and Classification: Standard feature extraction methods like Common Spatial Pattern (CSP) are used to extract features from the dynamically selected time windows. These features are then fed into a classifier, such as a Support Vector Machine (SVM), to perform the classification. The paper emphasizes that the feature extraction and classification algorithms are substitutable, meaning other methods could be integrated.
  • Cross-Validation: A 10-fold cross-validation scheme is employed to robustly evaluate the performance of the CTWS algorithm across participants.

Datasets

The CTWS algorithm was evaluated on two datasets:

  • Dataset 1 (BCI Competition IV Dataset I): This dataset consists of EEG recordings from seven participants (four healthy, three simulated) performing two-class motor imagery (left hand vs. right hand, or left hand vs. feet). It was sampled at 100 Hz.
  • Dataset 2 (Stroke Patients): This dataset was collected by the authors from seven stroke patients performing two-class motor imagery (left hand vs. right hand). It was sampled at 256 Hz and used 16 electrodes. This dataset is notable for representing a more challenging BCI application with a patient population.

Key Insights and Ideas

  • Trial-to-Trial Variability is Key: The most significant insight is that the timing of motor imagery events is not fixed, even within the same individual. This variability is a major bottleneck for traditional fixed-window approaches.
  • Correlation as a Dynamic Guide: Using correlation analysis to dynamically adjust the time window is an effective way to capture the most relevant signal segments for each trial, leading to improved feature extraction.
  • Universality of CTWS: The CTWS algorithm's effectiveness was demonstrated not only with CSP but also when combined with a more recent algorithm, Sub-Alpha-Beta Log-Det Divergences (Sub-ABLD), highlighting its potential as a general framework for MI feature extraction.
  • Patient Data Benefits: The CTWS algorithm shows promise for clinical applications, as it significantly improved performance on the stroke patient dataset, although the gains were less pronounced than in healthy individuals.

Key Results

  • Improved Accuracy with CSP: CTWS+CSP achieved a significant average improvement in classification accuracy of 16.72% for healthy participants (dataset 1) and 5.24% for stroke patients (dataset 2) compared to using CSP directly.
  • Improved Accuracy with Sub-ABLD: When combined with Sub-ABLD, CTWS+Sub-ABLD showed an average accuracy increase of 7.36% for dataset 1 and 9.29% for dataset 2.
  • Feature Separation: Figure 4 and 5 visually demonstrate that features extracted using CTWS+CSP (or CTWS+Sub-ABLD) are better separated between classes, indicating enhanced discriminative power.
  • Starting Time Variation: Figure 3 clearly illustrates that the optimal starting time for feature extraction varies considerably from trial to trial for each participant.

Limitations and Future Work

  • Channel Selection for Reference Signals: The study used channels C3 and C4 for reference signal optimization due to computational load. However, the authors acknowledge that individual differences in MI feature spatial distribution might mean these are not always the optimal channels. Future work could involve developing channel selection approaches for reference signal optimization.
  • Reduced Benefit in Stroke Patients: While CTWS improved performance for stroke patients, the benefit was less pronounced than for healthy individuals. The authors suggest this might be due to greater disturbance in stroke patients' MI features, which could limit the information available for classification. Further research could explore how to better handle these complex signal characteristics.

Summary updated

BibTeX
@article{feng2018towards,
  author = {Jiankui Feng and Erwei Yin and Jing Jin and Rami Saab and Ian Daly and Xingyu Wang and Dewen Hu and Andrzej Cichocki},
  title = {{Towards correlation-based time window selection method for motor imagery BCIs}},
  journal = {Neural Networks},
  year = {2018},
  doi = {10.1016/j.neunet.2018.02.011},
}
Figure