Classification of movement intention by spatially filtered electromagnetic inverse solutions

Marco Congedo, Fabien Lotte, A. LécuyerPhysics 2006 · DOI

TL;DR

This paper introduces a novel hybrid approach for classifying movement intentions from EEG signals by combining a standardized electromagnetic inverse solution (sLORETA) with a data-driven spatial filter (Common Spatial Pattern - CSP). This method achieves classification accuracy comparable to state-of-the-art techniques without requiring a trained classifier, making it faster and more suitable for real-time Brain-Computer Interface (BCI) applications.

What Problem Does This Paper Address?

Brain-Computer Interfaces (BCIs) aim to enable communication and control for individuals with severe motor impairments by translating brain activity into commands. A key challenge in BCI is accurately classifying specific intentions, such as imagining left versus right hand movements, from noisy EEG signals. Traditional methods often rely on complex, iterative algorithms for feature extraction (like Independent Component Analysis) or require extensive training of classifiers, which can be computationally expensive and time-consuming. This paper addresses the need for a more efficient and effective method to classify movement intentions by leveraging the spatial information embedded in EEG signals through advanced source localization and spatial filtering techniques.

Methods

The core of this paper lies in the novel combination of two key techniques:

  • sLORETA (Standardized Low-Resolution Electromagnetic Tomography): This is an inverse solution that estimates the location and strength of brain activity from scalp EEG measurements. It's chosen for its ability to provide a spatially resolved estimate of brain activity.
  • Common Spatial Pattern (CSP): Traditionally used for spatial filtering of EEG signals, CSP is adapted here as a data-driven beamformer. The authors use CSP to find spatial filters that maximize the separation of source power between two distinct brain activity patterns associated with left and right finger movements. This is achieved through a joint diagonalization process of covariance matrices derived from the EEG data.

The approach is "hybrid" because it uses a learning phase to design the spatial filter based on training data, but then employs an untrained classifier for real-time classification. This means the classification decision is made directly based on the filtered source power, without a separate learning step for the classifier itself. The method focuses on extracting a single "attribute" (the filtered source power) for each class (left/right intention).

Datasets

The study utilizes the BCI Competition 2003, Data Set IV. This dataset consists of EEG recordings from a non-clinical subject performing self-paced left and right finger tapping tasks.

  • Training Set: 159 epochs for left movements and 157 epochs for right movements.
  • Test Set: 49 epochs for left movements and 51 epochs for right movements. The data was acquired from 28 electrodes at a 1000 Hz sampling rate and epochs of 500 ms were extracted, ending 130 ms before the key press.

Key Insights and Ideas

  • Spatial Filtering Enhances Inverse Solutions: Simply using an EEG inverse solution like sLORETA alone is insufficient for accurate classification. The paper demonstrates that applying a data-driven spatial filter (CSP) significantly improves the ability to discriminate between movement intentions by enhancing the spatial segregation of relevant brain activity.
  • Untrained Classifiers are Viable and Faster: The research challenges the common assumption that a trained classifier is always necessary for high accuracy. By designing a powerful spatial filter, the authors show that a simple comparison of filtered source power between two regions is enough for effective classification, leading to much faster processing.
  • Movement Intention has Distinct Spatial Signatures: The study confirms that the intention to move left or right fingers generates distinct patterns of brain activity in contralateral motor cortices. The spatial filter effectively targets and amplifies these differences.
  • Focus on "Uncommonality" for Separation: The CSP method, when applied to inverse solutions, effectively finds spatial patterns that are "uncommon" or maximally different between the two classes, which is crucial for robust discrimination.

Key Results

  • Classification Accuracy: The proposed method achieved an accuracy of 83% on the test set using the last 250 ms of data. This is very close to the performance of the winners of the BCI competition 2003 for this dataset (84%).
  • Speed and Efficiency: The method is significantly faster than approaches relying on iterative ICA or trained classifiers, making it suitable for real-time BCI applications.
  • Comparison to Raw sLORETA: Using raw sLORETA without the spatial filter resulted in significantly lower classification accuracy (78%), highlighting the importance of the spatial filtering step.
  • Single Feature Sufficiency: The method achieved high accuracy using only one attribute (filtered source power) per class, suggesting that a well-defined feature can be more important than multiple, potentially redundant, features.

Limitations and Future Work

  • Head Model and Electrode Limitations: The authors acknowledge that the accuracy of source localization can be affected by the use of a spherical head model, the limited number of electrodes, and potential asymmetries in electrode placement.
  • Definition of Prototypical Covariance Structures: The quality of the spatial filter depends on the accuracy of the prototypical covariance structures (VL and VR) derived from training data. Improvements in this area, potentially using ICA, are suggested.
  • Potential for SMA Involvement: The study notes that early motor planning might involve areas like the Supplementary Motor Area (SMA), and detecting this early activation could be beneficial for even earlier detection of movement intention.
  • Further Improvements: Future work could explore more realistic head models, a denser electrode montage, and potentially incorporating features related to Bereitschaft potentials (preparatory brain activity) to further refine classification accuracy. The authors also suggest that multiple regions of interest and frequency/temporal windows could be explored.

Summary updated

Abstract

We couple standardized low-resolution electromagnetic tomography, an inverse solution for electroencephalography (EEG) and the common spatial pattern, which is here conceived as a data-driven beamformer, to classify the benchmark BCI (brain-computer interface) competition 2003, data set IV. The data set is from an experiment where a subject performed a self-paced left and right finger tapping task. Available for analysis are 314 training trials whereas 100 unlabelled test trials have to be classified. The EEG data from 28 electrodes comprise the recording of the 500 ms before the actual finger movements, hence represent uniquely the left and right finger movement intention. Despite our use of an untrained classifier, and our extraction of only one attribute per class, our method yields accuracy similar to the winners of the competition for this data set. The distinct advantages of the approach presented here are the use of an untrained classifier and the processing speed, which make the method suitable for actual BCI applications. The proposed method is favourable over existing classification methods based on an EEG inverse solution, which rely either on iterative algorithms for single-trial independent component analysis or on trained classifiers.
BibTeX
@inproceedings{congedo2006classification,
  author = {Marco Congedo and Fabien Lotte and A. Lécuyer},
  title = {{Classification of movement intention by spatially filtered electromagnetic inverse solutions}},
  booktitle = {Physics},
  year = {2006},
  doi = {10.1088/0031-9155/51/8/002},
}
Figure