EEG representation using multi-instance framework on the manifold of symmetric positive definite matrices
Khadijeh Sadatnejad1, Mohammad Rahmati1, Reza Rostami2,3
1Computer Engineering and Information Technology Department, Amirkabir University of Technology, Hafez Ave., Tehran, Iran.
Journal of Neural Engineering
|March 8, 2019
Summary
This study introduces a novel electroencephalogram (EEG) representation using a multi-instance framework to enhance brain signal analysis. The new method improves classification accuracy for cultural neuroscience and mental disorder diagnosis.
Area of Science:
- Neuroscience
- Signal Processing
- Machine Learning
Background:
- Generalization and robustness of electroencephalogram (EEG)-based systems are critical for practical applications.
- Existing EEG analysis methods may not fully capture the complex, non-stationary nature of brain signals.
Purpose of the Study:
- To propose a new EEG representation using a multi-instance (MI) framework to model the non-stationarity of EEG signals.
- To enhance the realistic view of brain functionality by considering signal segments as bags of concepts.
Main Methods:
- Utilized a multi-instance (MI) framework to represent EEG signals as bags of concepts, derived from spatial covariance matrices of homogeneous segments.
- Employed Riemannian framework for adaptive segmentation to determine boundaries of homogeneous segments due to the nonlinear geometry of covariance matrices.
- Applied bag-level discriminative information for classification, describing each subject as a bag of covariance matrices.
Main Results:
- Evaluated the approach in cultural neuroscience for classifying Iranian versus Swiss subjects, identifying potential distinguishing brain activity patterns.
- Assessed performance in EEG-based mental disorder diagnosis, including attention deficit hyperactivity disorder (ADHD)/bipolar mood disorder (BMD), Schizophrenia/normal, and Major Depression Disorder/normal.
- Experimental results demonstrated the superiority of the proposed approach over existing methods.
Conclusions:
- The proposed EEG representation offers superior performance due to robust covariance descriptors, effective Riemannian geometry, and consideration of non-stationary brain activity.
- The method automatically handles artifacts and leverages bag-level discriminative information for improved classification.
- This approach provides a more realistic and robust analysis of brain electrical activity for various applications.
