EEG-Based Mental Tasks Recognition via a Deep Learning-Driven Anomaly Detector

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Accurate recognition of mental tasks from electroencephalography (EEG) signals is crucial for brain-computer interfaces.
• Existing methods often struggle with the complexity and variability of EEG data.
• Unsupervised learning offers a promising avenue for robust EEG signal analysis without extensive labeled data.

Purpose of the Study:

• To introduce an unsupervised deep learning scheme for enhanced mental task recognition using EEG signals.
• To develop a robust artifact removal and feature extraction process for EEG data.
• To propose a novel classification method combining deep belief networks (DBN) and Isolation Forest (iF) for improved discrimination.

Main Methods:

• EEG signals were preprocessed using the Multichannel Wiener filter for artifact removal.
• Quadratic Time-Frequency Distribution (QTFD) was applied to extract discriminative time-frequency features.
• A deep belief network (DBN)-driven Isolation Forest (iF) scheme was developed for one-vs.-rest classification.
• The DBN learns data representations without distribution assumptions, while iF performs discrimination.

Main Results:

• The proposed DBN-based iF scheme demonstrated superior performance in discriminating between five mental tasks.
• The method achieved better classification accuracy compared to DBN-based Elliptical Envelope and Local Outlier Factor.
• The combination of QTFD features and the DBN-iF model significantly improved EEG-based mental task recognition.

Conclusions:

• The unsupervised DBN-driven iF scheme offers a powerful and effective approach for mental task recognition from EEG signals.
• The method's robustness to artifacts and ability to capture spectral variations contribute to its high performance.
• This work advances the field of brain-computer interfaces by providing a more accurate and reliable EEG classification technique.