Source-linked AI summary
Feature Learning from Incomplete EEG with Denoising Autoencoder
Junhua Li, Zbigniew Struzik, Liqing Zhang, Andrzej Cichocki
TL;DR
The paper addresses incomplete EEG segments in BCI systems by combining the Lomb-Scargle periodogram with learning methods including DAE. The proposed approach achieves comparable classification performance even after substantial data removal and is considered acceptable for BCI applications.
Problem
Incomplete EEG can prevent a BCI system from generating corresponding outputs during affected periods.
Method
The method combines the Lomb-Scargle periodogram with DAE or SVM to distinguish incomplete EEG segments.
Results
Comparable classification performance is achieved even when most data points in a segment have been removed, supporting the method's acceptability for BCI applications.
Takeaways & Limitations
The proposed method provides an alternative solution for a frequent problem in BCI systems involving incomplete segments.
Takeaways & Limitations
The method appears relatively sensitive to block-form data removal, and DAE does not significantly exceed SVM in EEG classification.
Abstract
from arXiv · showhide
An alternative pathway for the human brain to communicate with the outside world is by means of a brain computer interface (BCI). A BCI can decode electroencephalogram (EEG) signals of brain activities, and then send a command or an intent to an external interactive device, such as a wheelchair. The effectiveness of the BCI depends on the performance in decoding the EEG. Usually, the EEG is contaminated by different kinds of artefacts (e.g., electromyogram (EMG), background activity), which leads to a low decoding performance. A number of filtering methods can be utilized to remove or weaken the effects of artefacts, but they generally fail when the EEG contains extreme artefacts. In such cases, the most common approach is to discard the whole data segment containing extreme artefacts. This causes the fatal drawback that the BCI cannot output decoding results during that time. In order to solve this problem, we employ the Lomb-Scargle periodogram to estimate the spectral power from incomplete EEG (after removing only parts contaminated by artefacts), and Denoising Autoencoder (DAE) for learning. The proposed method is evaluated with motor imagery EEG data. The results show that our method can successfully decode incomplete EEG to good effect.
1. Introduction
BCIs translate brain activity into continuous outputs, but extreme EEG artefacts can force removal of entire segments and interrupt decoding. The proposed approach retains unaffected samples, estimates spectral power from incomplete EEG with the Lomb-Scargle periodogram, and uses DAE or SVM to classify motor imagery.
- The BCI must translate brain activity continuously, yet discarding EEG segments contaminated by extreme noise prevents output during those periods.
- The method retains unaffected portions of artefact-contaminated EEG instead of removing the entire affected segment.
- Spectral power features support distinguishing motor imagery classes such as left-hand and right-hand imagery.
- The Lomb-Scargle periodogram estimates spectral power when EEG data are incomplete or unevenly spaced, where the Fourier transform fails.
- The proposed system combines Lomb-Scargle spectral estimation with DAE or SVM classification for incomplete motor-imagery EEG.
- The results show that the proposed method is suitable for decoding incomplete EEG in a BCI system.
2. Methodology
The method estimates spectral features from incomplete EEG with the Lomb-Scargle periodogram, then uses DAE initialization or an RBF-kernel SVM for motor-imagery classification. A three-layer neural network is pretrained by reconstructing corrupted features and subsequently fine-tuned supervised.
- 2.2. DAE-based neural network: The normalized features feed either a DAE-initialized neural network or an SVM with an RBF kernel for classifying motor imageries.The DAE mapping weights initialize the neural network before supervised fine-tuning.
- 2.1. Lomb-Scargle Periodogram: Lomb-Scargle spectral estimation converts incomplete EEG into features for classification.The method repeats spectral estimation across channels and segments, then forms subband-power features.
- 2.1. Lomb-Scargle Periodogram: Four subband powers from 8–30 Hz are averaged per channel and concatenated into a feature vector.The subbands are 8–12, 13–17, 18–22, and 23–27 Hz.
- 2.2. DAE-based neural network: The constructed classifier uses three neural-network layers with DAE initialization, while deeper networks might improve performance through more extensive feature learning.The paper leaves the effect of deeper networks unresolved.
- 2.2. DAE-based neural network: The DAE corrupts Lomb-Scargle power features, maps them through a 120-unit sigmoid hidden layer, and reconstructs the uncorrupted features.Training minimizes average squared reconstruction error over the training set.
3. Evaluation Data
Evaluation uses simulated signals and two-class motor-imagery EEG to test spectral estimation and classification with removed data points. The experiments report that spectral power remains estimable for unevenly spaced data and classification accuracy does not dramatically decrease as more segment data are removed.
- 3. Evaluation Data: The evaluation combines simulated data with real two-class motor-imagery EEG.The simulated data test spectral estimation, while real EEG tests classification under data removal.
- 3. Evaluation Data: Spectral power can be correctly estimated after simulated signals become unevenly spaced through data-point removal.The simulated signals were generated by mixing two sinusoidal signals.
- 3. Evaluation Data: Classification accuracy does not dramatically decrease as the percentage of removed data within a segment increases.This result supports processing incomplete data in a BCI system.
4. Results
The method preserved spectral components and classification accuracy despite substantial incomplete EEG, although block removal and DAE–SVM differences introduced important conditions and limitations.
- Simulated data: The 3 Hz and 6 Hz spectral components were well estimated even with up to 80% of data points removed.The simulated signals combined 3 Hz and 6 Hz sinusoids, and spectral power was compared across removal proportions from 10% to 80%.
- Real motor imagery data: Accuracies generally did not decrease dramatically across sessions and subjects under data-point and block-point removal.Trial accuracies varied more than sliding-time-window accuracies because trial labels depended on the majority of window classifications.
- Real motor imagery data: Classification accuracy could remain comparable even when 80% of data were removed.The result was observed across data-removal conditions, with high accuracies retained from 10% to 80% removal in several sessions.
- Real motor imagery data: Block removal was more challenging for subject 1 at 80% removal, where accuracy was substantially worse than at 70%.The authors state that the method appears relatively sensitive to block-form data removal.
- Comparison between DAE and SVM: DAE achieved higher overall average accuracy than SVM for both data-point and block-point removal.DAE was better in some sessions and worse in others; its advantage was smaller under block-point removal.
- Comparison between DAE and SVM: DAE was comparable to SVM, but whether it can significantly exceed SVM for EEG classification remained unclear.The authors suggest deeper networks and better parameter tuning might improve DAE performance.
5. Conclusion
The proposed method combines the Lomb-Scargle periodogram with SVM or DAE to classify incomplete EEG segments. It maintains acceptable classification performance after data removal, enabling contaminated segments to remain usable for BCI commands instead of being discarded.
- Combining the Lomb-Scargle periodogram with SVM or DAE enables classification of incomplete EEG segments.Incomplete segments are those from which part of the data was removed because of noise contamination.
- Classification accuracy is not dramatically decreased when different percentages of data are removed.
- The method provides an alternative to discarding contaminated segments, allowing them to remain usable for outputting BCI commands after only the noisy portion is removed.
- The proposed method achieves comparable classification performance even when most data points in a segment have been removed.