Automatic subject-specific spatiotemporal feature selection for subject-independent affective BCI
Badar Almarri1,2, Sanguthevar Rajasekaran1, Chun-Hsi Huang3
1Dept. of Computer Science and Engineering, University of Connecticut, Storrs, CT, United States of America.
Plos One
|August 26, 2021
Summary
This study introduces a subject-independent framework for emotion recognition using electroencephalogram (EEG) brain-computer interfaces (BCI). It enhances accuracy by learning from individual differences, outperforming existing methods.
Area of Science:
- Neuroscience
- Computer Science
- Machine Learning
Background:
- Electroencephalogram (EEG) based brain-computer interfaces (BCI) face challenges in emotion recognition due to high dimensionality and limited temporal resolution.
- Conventional methods often struggle with inter-subject variability, either by removing data or making models subject-dependent.
- Effective emotion recognition requires robust preprocessing and feature extraction from both temporal (time-series) and spatial (electrode channels) data.
Purpose of the Study:
- To propose a subject-independent emotion recognition framework for EEG-BCI systems.
- To mitigate the significant subject-to-subject variability inherent in EEG emotion recognition.
- To enhance the accuracy and generalizability of emotion recognition models across different individuals.
Main Methods:
- Utilized an unsupervised feature selection algorithm to reduce the feature space from time-series EEG signals.
- Developed a subject-specific unsupervised learning algorithm for online learning of inter-channel co-activation patterns.
- Implemented a nested cross-validation strategy with a Support Vector Machine (SVM) for training and testing on benchmark datasets (DEAP, MAHNOB-HCI, DREAMER).
Main Results:
- Achieved enhanced performance in classifying human emotions based on valence and arousal, with accuracy improvements of 16%-27% compared to other subject-independent studies.
- Demonstrated superior performance over existing state-of-the-art subject-independent algorithms in EEG-based emotion recognition.
- Validated the framework's effectiveness on multiple real-world EEG datasets.
Conclusions:
- The proposed framework effectively addresses subject-to-subject variability in EEG emotion recognition, leading to significant performance gains.
- The developed unsupervised learning approach for spatial features offers an online analysis solution for real-time affection recognition.
- This research advances the field of BCI by providing a more robust and accurate method for subject-independent emotion recognition.
