TL;DR
This paper introduces a novel classification-based approach for assessing human attention using single-channel EEG signals. By combining Fast Fourier Transform (FFT) for denoising, Empirical Mode Decomposition (EMD) for trend removal, and Singular Spectrum Analysis (SSA) for enhanced feature extraction, the method achieves higher classification accuracy than traditional approaches, offering a more objective way to diagnose attention-related disorders.
What Problem Does This Paper Address?
Accurately assessing human attention is crucial for diagnosing conditions like Attention Deficit Hyperactivity Disorder (ADHD) and autism. Current methods often rely on subjective assessments from medical professionals, leading to inconsistencies. While EEG offers a more objective measure of brain activity, extracting meaningful attention-related signals from noisy, single-channel EEG data, especially during various everyday activities, remains a challenge. Existing EEG-based attention assessment methods are often limited to specific tasks like playing electronic games and may not generalize well to other activities. Furthermore, traditional signal processing techniques can introduce distortions, and the limited number of features extracted from brain waves can hinder classification accuracy.
Methods
The proposed method tackles these challenges through a multi-stage signal processing pipeline:
- FFT-Based Denoising: The raw single-channel EEG signals are first processed using FFT. High-frequency components above 50 Hz, considered noise, are discarded. This step is crucial for cleaning the signal without introducing phase distortion, unlike traditional filtering methods.
- Empirical Mode Decomposition (EMD) for Detrending: EMD is employed to remove the underlying trend from the denoised EEG signals. This is important because eye and muscular movements can introduce drifts in EEG data, which EMD can effectively isolate and remove by discarding the last few intrinsic mode functions (IMFs) that represent these trends.
- Singular Spectrum Analysis (SSA) for Feature Enhancement: To overcome the limitation of a small number of features, SSA is applied. SSA decomposes each detrended EEG signal and its constituent brain waves into multiple components. This significantly increases the number of extractable features, providing a richer representation of the signal.
- Feature Extraction: From the SSA components, a comprehensive set of 11 physical quantities are extracted. These include approximate entropy, mean, variance, interquartile range, mean absolute deviation, range, skewness, kurtosis, and L1, L2, and L∞ norms. These features are then combined to form a feature vector.
- Classification: Three common machine learning classifiers are used for attention assessment: Random Forest, Support Vector Machine (SVM), and Back-Propagation (BP) Neural Network. The classification accuracy achieved by these models is then used as the "attention score."
Datasets
The study utilized a newly built dataset of single-channel EEG recordings from 12 participants (9 males, 3 females) performing seven different activities (drawing, eating, doing computer exercises, playing electronic games, reading, playing toys, and watching television). Each activity was performed twice: once under a "concentration state" (quiet environment, no interruptions) and once under an "immersion state" (with disturbances). The data was acquired using a NeuroSky TGAM1 chip at a sampling frequency of 512 Hz. The authors note that a public database for this specific type of activity recognition using single-channel EEG at 512 Hz was not readily available, necessitating the creation of this new dataset.
Key Insights and Ideas
- Classification for Regression: The paper cleverly reframes attention assessment, typically a regression problem, as a classification task. This bypasses the need for subjective reference values from medical officers, making the assessment more objective.
- Synergistic Signal Processing: The combination of FFT, EMD, and SSA is a key innovation. FFT cleans the signal, EMD removes trends that could interfere with feature extraction, and SSA dramatically expands the feature space, leading to improved classification performance.
- Feature Richness from SSA: The ability of SSA to decompose signals into numerous components is crucial for extracting a wide array of features, which are then shown to be statistically significant for attention assessment.
- Random Forest Superiority: The Random Forest classifier consistently outperformed SVM and BP neural networks across various scenarios, attributed to its inherent randomness in feature selection, which mitigates overfitting.
Key Results
- The proposed method, incorporating FFT, EMD, and SSA, achieved higher classification accuracies compared to traditional methods (FFT-based denoising and classification without EMD/SSA) across all activities and both concentration and immersion states.
- The highest recorded classification accuracy in the concentration state using the proposed method reached 100%.
- The average classification accuracies for the proposed method were consistently higher than traditional methods, demonstrating its effectiveness. For instance, in the concentration state, the average accuracies for BP, RF, and SVM were 82.04%, 84.99%, and 80.11%, respectively, compared to 74.85%, 79.80%, and 78.46% for traditional methods.
- The sorting order of activities based on attention levels (e.g., reading and computer exercises requiring more attention than eating) was consistent between the proposed and traditional methods, indicating reliability.
Limitations and Future Work
- Hardware Constraint: The study used only a single-channel EEG device, limiting the amount of information that could be extracted.
- Limited Data: The current dataset is relatively small.
- Future Work: The authors suggest incorporating other signal types like image sequences and motion signals from cameras and motion sensors to enhance attention assessment. They also plan to acquire more extensive EEG data to improve the reliability of the classification accuracies. The study also explored optimal parameters for the classifiers (e.g., number of neurons in BP networks, leaf nodes and trees in Random Forest, kernel parameters in SVM) and found that the optimal parameters can vary between individuals and even between recording sessions.
Summary updated