Data-Centric AI for EEG-Based Emotion Recognition: Noise Filtering and Augmentation Strategies

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Neuroscience

Background:

• Biomedical research faces data scarcity and high costs, hindering machine learning (ML) model development.
• Existing ML models for emotion recognition often rely on complex architectures.
• Electroencephalogram (EEG) data presents unique challenges for analysis.

Purpose of the Study:

• To introduce a data-centric AI framework for EEG-based emotion recognition.
• To demonstrate the impact of data quality improvements (noise filtering, augmentation) on model performance.
• To compare the proposed data-centric approach against complex models.

Main Methods:

• Implemented participant-guided noise filtering on EEG data.
• Applied systematic data augmentation techniques.
• Evaluated performance across binary, four-quadrant, and discrete emotion classification tasks using the SEED-VII dataset.

Main Results:

• Data-centric strategies significantly improved accuracy and F1 scores across all classification settings.
• The approach achieved competitive or superior performance compared to more complex models.
• Noise filtering and data augmentation proved effective for enhancing EEG emotion recognition.

Conclusions:

• Prioritizing data quality over model complexity offers a robust and generalizable pathway for biomedical AI.
• The proposed framework provides a practical and reproducible method for advancing EEG-based emotion recognition.
• Data-centric AI is a viable strategy for overcoming data limitations in biomedical research.