[Paper Review] Real-time EEG-based Emotion Recognition using Discrete Wavelet Transforms on Full and Reduced Channel Signals

Josef Bajada, Francesco Borg Bonello|arXiv (Cornell University)|Oct 11, 2021

EEG and Brain-Computer Interfaces4 citations

TL;DR

This paper proposes a real-time EEG-based emotion recognition system using Discrete Wavelet Transform (DWT) features on both full 32-channel and reduced 5-channel EEG data, with baseline removal preprocessing to enhance accuracy. The SVM classifier achieved 95.32% accuracy on Valence and 95.68% on Arousal in subject-dependent settings, outperforming prior real-time benchmarks, while maintaining 92% accuracy on 5-channel data—demonstrating feasibility for low-cost consumer EEG devices.

ABSTRACT

Real-time EEG-based Emotion Recognition (EEG-ER) with consumer-grade EEG devices involves classification of emotions using a reduced number of channels. These devices typically provide only four or five channels, unlike the high number of channels (32 or more) typically used in most current state-of-the-art research. In this work we propose to use Discrete Wavelet Transforms (DWT) to extract time-frequency domain features, and we use time-windows of a few seconds to perform EEG-ER classification. This technique can be used in real-time, as opposed to post-hoc on the full session data. We also apply baseline removal preprocessing, developed in prior research, to our proposed DWT Entropy and Energy features, which improves classification accuracy significantly. We consider two different classifier architectures, a 3D Convolutional Neural Network (3D CNN) and a Support Vector Machine (SVM). We evaluate both models on subject-independent and subject dependent setups to classify the Valence and Arousal dimensions of an individual's emotional state. We test them on both the full 32-channel data provided by the DEAP dataset, and also a reduced 5-channel extract of the same dataset. The SVM model performs best on all the presented scenarios, achieving an accuracy of 95.32% on Valence and 95.68% on Arousal for the full 32-channel subject-dependent case, beating prior real-time EEG-ER subject-dependent benchmarks. On the subject-independent case an accuracy of 80.70% on Valence and 81.41% on Arousal was also obtained. Reducing the input data to 5 channels only degrades the accuracy by an average of 3.54% across all scenarios, making this model appropriate for use with more accessible low-end EEG devices.

Motivation & Objective

• To develop a real-time EEG-based emotion recognition system suitable for low-cost, consumer-grade EEG devices with limited channels (e.g., 5 channels).
• To improve classification accuracy using Discrete Wavelet Transform (DWT) features combined with baseline removal preprocessing on both 32-channel and 5-channel EEG data.
• To evaluate and compare two classifier architectures—SVM and 3D CNN—across subject-dependent and subject-independent setups for valence and arousal classification.
• To demonstrate that high accuracy can be maintained when reducing from 32 to 5 channels, enabling practical deployment on accessible EEG hardware.

Proposed method

• Discrete Wavelet Transform (DWT) is applied to extract time-frequency domain features from EEG signals, using Daubechies wavelets and decomposition levels optimized for emotion recognition.
• Baseline removal preprocessing is applied to DWT-derived features (entropy and energy), normalizing each subject’s active state relative to their rest state to improve feature discriminability.
• A 3D Convolutional Neural Network (3D CNN) is trained on DWT entropy and combined DWT entropy + energy features, while an SVM with RBF kernel is used for comparison.
• Time-windowing is applied to EEG data in 2–4 second intervals to enable real-time classification, avoiding post-hoc processing of full sessions.
• A chained SVM model is proposed, where the output of one emotion dimension (e.g., Arousal) is used as input to predict the other (e.g., Valence), leveraging conditional dependence.
• The models are evaluated on the DEAP dataset using both 32-channel and 5-channel subsets, with performance measured in subject-dependent and subject-independent setups.

Experimental results

Research questions

• RQ1 Can Discrete Wavelet Transform (DWT) features combined with baseline removal improve real-time EEG-based emotion recognition accuracy on low-channel EEG data?
• RQ2 How does model performance degrade when reducing from 32 to 5 EEG channels, and can high accuracy still be achieved?
• RQ3 Does the SVM classifier outperform the 3D CNN in both subject-dependent and subject-independent emotion recognition tasks using DWT features?
• RQ4 Can a chained SVM model that uses conditional prediction between Valence and Arousal dimensions improve overall classification accuracy?
• RQ5 Is the proposed method robust enough to achieve state-of-the-art accuracy on 5-channel EEG data, enabling deployment on consumer-grade EEG devices?

Key findings

• The SVM classifier achieved 95.32% accuracy on Valence and 95.68% on Arousal in the subject-dependent 32-channel setup, surpassing the prior state-of-the-art of 92.7% and 93.1% respectively.
• On the subject-independent setup, the SVM achieved 80.70% accuracy on Valence and 81.41% on Arousal, demonstrating strong generalization across individuals.
• Reducing the input from 32 to 5 channels degraded average accuracy by only 3.54%, with the SVM maintaining 92% accuracy on Valence and 92.1% on Arousal in the subject-dependent 5-channel case.
• Baseline removal preprocessing significantly improved classification accuracy across all scenarios, indicating its importance for normalizing individual differences in EEG signals.
• The chained SVM model, which uses conditional prediction between Arousal and Valence, achieved slightly higher accuracy than the single SVM model, suggesting value in modeling inter-dimension dependencies.
• The 3D CNN model performed worse than the SVM, indicating that simpler, kernel-based models may be more effective than deep learning for this specific DWT-based feature representation.