•End-to-end neural network model for classifying motor imagery EEG signals.•Using 1-D CNN layers to learn temporal and spatial filters for feature extraction.•Application of transfer learning to ...calibrate the model for individual subjects.•Analysis of the temporal and spatial filters learned by the model.
Goal: To develop and implement a Deep Learning (DL) approach for an electroencephalogram (EEG) based Motor Imagery (MI) Brain-Computer Interface (BCI) system that could potentially be used to improve the current stroke rehabilitation strategies.
Method: The DL model is using Convolutional Neural Network (CNN) layers for learning generalized features and dimension reduction, while a conventional Fully Connected (FC) layer is used for classification. Together they build a unified end-to-end model that can be applied to raw EEG signals. This previously proposed model was applied to a new set of data to validate its robustness against data variations. Furthermore, it was extended by subject-specific adaptation. Lastly, an analysis of the learned filters provides insights into how such a model derives a classification decision.
Results: The selected global classifier reached 80.38%, 69.82%, and 58.58% mean accuracies for datasets with two, three, and four classes, respectively, validated using 5-fold crossvalidation. As a novel approach in this context, transfer learning was used to adapt the global classifier to single individuals improving the overall mean accuracy to 86.49%, 79.25%, and 68.51%, respectively. The global models were trained on 3s segments of EEG data from different subjects than they were tested on, which proved the generalization performance of the model.
Conclusion: The results are comparable with the reported accuracy values in related studies and the presented model outperforms the results in the literature on the same underlying data. Given that the model can learn features from data without having to use specialized feature extraction methods, DL should be considered as an alternative to established EEG classification methods, if enough data is available.
Reliable signal classification is essential for using an electroencephalogram (EEG) based Brain-Computer Interface (BCI) in motor imagery (MI) training. While deep learning (DL) is used in many areas ...with great success, only a limited number of works investigate its potential in this domain. This study presents a DL approach, which could improve or replace current state-of-the-art methods. Here, an end-to-end convolutional neural network (CNN) model is presented, which can be applied to raw EEG signals. It consists of a temporal and spatial convolution layer for feature extraction and a fully connected (FC) layer for classification. The global models were trained on 3s segments of EEG data. Training a subject-independent global classifier reaches 80.10%, 69.72%, and 59.71 % mean accuracy for a dataset with two, three, and four classes, respectively, validated in 5-fold cross-validation. Retraining the global classifier with data from single individuals improves the overall mean accuracy to 86.13%, 79.05%, and 68.93%, respectively. The results are superior to the results reported in the literature on the same data. Generally, the reported accuracy values are comparable with related studies, which shows that the model delivers competitive results. As raw signals are used as input, no pre-processing is needed, which qualifies DL methods as a promising alternative to established EEG classification methods.
We study the statistical characteristics of radio environment noise, in particular focusing on spurious received power arising from adjacent channel interference, harmonics and wideband man-made ...interference. Unlike the majority of existing studies, our measurements enable a comprehensive study on frequencies ranging from 80MHz to 2.64 GHz instead of being focused on the properties of an application-specific narrow band. We also apply a comprehensive statistical methodology to study how closely the measured noise can be modeled using a traditional white Gaussian stochastic process. In particular, we do not only focus on the parametric modeling of the marginal power distribution, but seek deviations from the whiteness of the noise through detailed time domain characterization, and apply a rigorous hypothesis testing on deviations from normality. Our results show that while the measured noise is often close to being white and nearly Gaussian, these properties are not as universal as often assumed in the literature. In particular, impulsive noise causes deviations both from normality of the marginals, as well as from the whiteness of the spectrum at levels that depend on the environment.