EEG Signal Processing

NeuroLab employs a state-of-the-art EEG signal processing pipeline designed to handle noisy real-world data and extract meaningful features for mental state classification.

Pipeline Overview

The EEG processing pipeline consists of five primary stages:

graph LR
    A[Raw Data] --> B[Validation]
    B --> C[Preprocessing]
    C --> D[Epoching]
    D --> E[Feature Extraction]
    E --> F[ML Classification]

1. Signal Quality Validation

Before processing, raw signals undergo a rigorous validation step to ensure data integrity:

• Missing Data Detection: Identifying and handling null or infinite values.
• Signal-to-Noise Ratio (SNR): Calculating the ratio of signal power to noise power.
• Artifact Detection: Monitoring peak-to-peak amplitude and zero-crossing rates to identify high-amplitude noise or hardware artifacts.
• Entropy Analysis: Measuring signal complexity to distinguish between physiological data and flat/repetitive noise.

2. Preprocessing & Artifact Removal

EEG signals are extremely low-amplitude (microvolts) and easily contaminated by environmental and physiological noise (EOG, EMG).

Artifact Cleaning

The system uses automated artifact removal techniques to clean the signal without losing neural information. This includes removing eye blinks, muscle activity, and electrode movement artifacts.

Filtering

We apply standard digital filters to isolate neural activity:

• Band-pass Filter (0.5Hz - 45Hz): Removes low-frequency drifts and high-frequency noise.
• Notch Filter (50Hz/60Hz): Specifically targets electrical line noise.

3. Epoching

Continuous EEG data is segmented into fixed-length windows called epochs.

• Window Size: Typically 257 samples (~1.028 seconds at 250Hz).
• Overlap: Configurable overlap (e.g., 50%) to ensure continuity in real-time classification.

4. Feature Extraction

NeuroLab extracts a comprehensive set of 930+ features per epoch to provide a high-dimensional representation of brain activity.

Frequency Domain Features (Spectral)

Spectral analysis is the cornerstone of EEG processing. We use Welch's method to compute Power Spectral Density (PSD) and extract power from canonical bands:

| Band | Frequency | Associated Mental States | | :--- | :--- | :--- | | Delta (δ) | 0.5 - 4 Hz | Deep sleep, unconsciousness | | Theta (θ) | 4 - 8 Hz | Drowsiness, meditation, light sleep | | Alpha (α) | 8 - 13 Hz | Relaxed wakefulness, closed eyes | | Beta (β) | 13 - 30 Hz | Alertness, active thinking, focus | | Gamma (γ) | 30 - 45 Hz | High-level cognitive processing, peak focus |

Ratios: We also calculate band power ratios such as Alpha/Beta (Relaxation Index) and Beta/Theta (Engagement Index).

Time Domain Features

• Hjorth Parameters: Activity, Mobility, and Complexity.
• Statistical Moments: Mean, variance, skewness, and kurtosis.
• Waveform Characteristics: Zero-crossing rate, RMS, and peak-to-peak amplitude.

Non-linear & Complex Features

• Entropy Metrics: Sample entropy, permutation entropy, and spectral entropy.
• Phase Synchrony: Measuring coherence and phase synchronization between different brain regions.
• Wavelet Transform: Multi-resolution analysis to capture transient neural events.

5. Feature Engineering for ML

The extracted features are further prepared for the machine learning models:

• Outlier Removal: Using Isolation Forest to prune anomalous data points.
• Scaling: Robust or Standard scaling to normalize feature ranges.
• Class Balancing: Applying SMOTE or ADASYN to handle imbalance in mental state datasets.
• Feature Selection: Selecting the most discriminative features using F-classif and Mutual Information.

---

Next Steps

Continue with the mental state model or the AI service reference: