An ensemble deep-learning approach for single-trial EEG classification of vibration intensity

Area of Science:

• Neuroscience
• Machine Learning
• Human-Computer Interaction

Background:

• Single-trial electroencephalography (EEG) classification is crucial for understanding cognitive states during haptic feedback.
• Convolutional Neural Networks (CNNs) excel at extracting features from EEG data for cognitive function classification.
• Perception of vibration intensity is a key aspect of human-computer interaction requiring accurate cognitive assessment.

Purpose of the Study:

• To propose a novel CNN ensemble model, SE NexFusion, for classifying vibration intensity from single-trial EEG.
• To outperform existing state-of-the-art EEG models in vibration intensity classification.
• To evaluate the model's ability to capture cognitive experiences related to haptic feedback.

Main Methods:

• Developed SE NexFusion, a CNN ensemble model leveraging complementary learning from EEGNex and TCNet Fusion.
• Employed multi-branch feature encoders with squeeze-and-excitation units for enhanced feature extraction and recalibration.
• Utilized raw single-trial EEG data, minimizing the need for complex preprocessing.

Main Results:

• Achieved average accuracies of 60.7% (leave-one-subject-out) and 61.6% (within-subject) for three-class vibration intensity classification.
• Demonstrated superior performance compared to several state-of-the-art EEG models.
• Validated model robustness using Shapley values to identify influential spatio-temporal features.

Conclusions:

• SE NexFusion significantly outperforms benchmarked EEG models in classifying vibration intensity.
• Explainability analysis confirms the model's focus on relevant neural correlates of vibration intensity.
• The model offers a robust method for evaluating cognitive states associated with haptic feedback.