EEG signals similarity based on compression

Michal Prilepok, Jan Platos, Václav SnáselDATESO 2014

TL;DR

This paper proposes a novel method for detecting index finger movements from EEG signals by leveraging Lempel-Ziv (LZ) complexity for similarity measurement. By converting smoothed EEG data into a sequence of "turtle graphics" commands, the authors demonstrate a successful approach to distinguishing between movement and non-movement states, achieving detection rates up to 58.78%.

What Problem Does This Paper Address?

Electroencephalography (EEG) signals, representing the brain's electrical activity, hold immense potential for applications like Brain-Computer Interfaces (BCIs). However, the complexity and noisy nature of EEG data make it challenging to reliably detect specific brain states or actions. This paper tackles the problem of accurately identifying when a person is performing a specific motor task – in this case, pressing a button with their index finger – based solely on their EEG signals. The goal is to develop a robust method that can differentiate between periods of movement and periods of rest, which is a fundamental step towards more sophisticated BCI control.

Methods

The authors employ a multi-step approach to process and analyze the EEG data:

  1. Filtering: Raw EEG signals are first band-pass filtered (0.5 Hz to 60 Hz) to remove unwanted low and high frequencies, as well as noise.
  2. Movement Segmentation: The data is then segmented to extract periods corresponding to the index finger movement, including a small window before and after the actual movement.
  3. Polynomial Curve Fitting: To further smooth the data and capture the underlying trend, a 15th-order polynomial curve fitting is applied using a least squares approach. This step aims to reduce noise while preserving the essential signal characteristics.
  4. Turtle Graphics Conversion: The smoothed EEG data is then transformed into a sequence of "turtle graphics" commands. This is a novel step where numerical data points are converted into a symbolic representation of movement commands (e.g., forward, turn left/right). This conversion is done using only two quadrants of the turtle graphics space to represent the time-series nature of the data.
  5. Lempel-Ziv (LZ) Complexity for Similarity: The core of the similarity measurement lies in applying LZ complexity to the generated turtle graphics command sequences. LZ complexity quantifies the complexity of a finite sequence by measuring the number of distinct substrings and their recurrence rate. By comparing the LZ complexity of sequences from different EEG trials, the authors can quantify their similarity. A similarity metric (SM) is calculated based on the number of common string sequences between two lists of LZ sequences.
  6. Classification: Based on the calculated similarity scores, a classification is performed. Training data is used to establish thresholds (Tmin and Tmax) for similarity. A test trial is then classified as "with movement" or "without movement" based on whether its similarity to the training data falls within these thresholds. The final classification is determined by the ratio of movement trials to the total number of selected trials in a group.

Datasets

The experiments were conducted using EEG data recorded in the authors' laboratory.

  • Data Source: Data was collected from four different subjects performing the same task: pressing a button with their left index finger.
  • Recording Details: 7 channels of EEG data were recorded at a sampling rate of 256 Hz.
  • Data Split: The pre-processed data contained 4606 trials in total (2303 with finger movement, 2303 without). This data was divided into seven groups, one for each sensor. For each sensor group, 75% of the data was used for training and 25% for testing. The training part for a single sensor contained 492 trials (246 with movement, 246 without), and the testing part contained 166 trials (83 with movement, 83 without).

Key Insights and Ideas

  • Bridging EEG and Symbolic Representation: The conversion of smoothed EEG data into "turtle graphics" commands is a creative approach to represent complex time-series data in a more structured, symbolic format suitable for sequence analysis.
  • LZ Complexity as a Robust Similarity Measure: Applying LZ complexity to these symbolic sequences allows for a non-parametric and computationally efficient way to measure the similarity between different EEG trials, even in the presence of noise.
  • Movement Detection via Sequence Similarity: The core idea that distinct mental tasks (like movement vs. rest) manifest as different patterns in EEG, which can then be captured by the complexity of their symbolic representations, is a powerful concept.
  • Sensor-Specific Performance: The results highlight that the effectiveness of the detection method can vary across different EEG sensors, suggesting that sensor placement and signal quality play a significant role.

Key Results

The proposed method achieved a successful detection rate for index finger movement ranging from 56.02% to 58.78% across different sensors.

  • Best Performance: The best results were observed on sensors S5 (58.78%) and S2, S4 (58.43%).
  • Worst Performance: The lowest detection rate was for sensor S7 (56.02%).
  • Movement vs. No Movement Detection:
    • Detection rates for trials with movement varied between 36.14% (S6) and 72.28% (S7).
  • Detection rates for trials without movement varied between 39.75% (S7) and 77.10% (S6).
  • Thresholds: The study found that minThreshold values were generally around 0.30 and maxThreshold values around 0.50.

Limitations and Future Work

  • Limited Task Scope: The current study focused on a single, relatively simple motor task (index finger movement). The authors acknowledge that the method's effectiveness on more complex tasks or a wider range of mental states needs further investigation.
  • Sensor Variability: The performance varied significantly across different sensors, indicating a need for further research into sensor selection or adaptive methods that can account for this variability.
  • Real-World Application: While the method shows promise, its application to real-world, unconstrained scenarios with more diverse and potentially less controlled EEG data is a natural next step.
  • Generalizability: Future work could explore the application of this approach to other types of EEG signals and BCI applications, potentially with different symbolic conversion methods or LZ complexity variations.

Summary updated

BibTeX
@inproceedings{prilepok2014eeg,
  author = {Michal Prilepok and Jan Platos and Václav Snásel},
  title = {{EEG signals similarity based on compression}},
  booktitle = {DATESO},
  year = {2014},
}
Figure