Overview
Epileptic seizures represent transient, pathological hypersynchronizations of large cortical networks. To design automated architectures capable of rapid detection and preemptive therapeutic deployment, signal processing pipelines must extract reliable biomarkers from continuous scalp electroencephalograms (EEGs). Scalp EEGs are non-invasive recordings captured via arrays of scalp electrodes arranged in accordance with the International 10-20 electrode placement system.
The CHB-MIT Scalp EEG Seizure Database, collected at Boston Children's Hospital, is an extensive diagnostic resource consisting of continuous EEG recordings spanning 22 pediatric patients with intractable epilepsy. Because pediatric EEGs present elevated baseline variability, rhythmic non-pathological microstructures (such as sleep spindles, vertex sharp waves, and artifactual muscular contractions) frequently mimic the spatiotemporal signatures of clinical seizures. Standard classifiers routinely experience high false-positive rates when confronted with this elevated physiological variability, making robust, multi-dimensional feature engineering an essential prerequisite for reliable machine learning performance.
By modeling temporal, spectral, and complexity distributions across multi-channel structures, researchers can identify pre-ictal anomalies before clear physical manifestations occur. The mathematical framework deployed in this simulator mirrors clinical criteria, tracking signal evolution, amplitude variances, and spectral power density offsets.
How to Use
The laboratory simulator allows you to alter signal parameters, electrode selections, and feature extraction bounds to observe the impact on high-dimensional classification in real time:
1. Selecting Datasets and Channels: Use the "CHB-MIT Dataset Recording" dropdown menu to toggle between synthetic simulation files and direct binary EDF parse fetches. The "Target Electrode Channel" selection dynamically changes the underlying electrode pair being analyzed among the 18 standard clinical channels.
2. Signal Presets: Manually trigger signal generator baselines using the control buttons: "Normal Background" models typical 1/f neural noise; "Sleep Spindle" injects rhythmic 11–14 Hz waveforms; "Muscle Artifact" replicates high-frequency EMG noise; and "Seizure Activity" models rhythmic, slowing spike-and-wave patterns.
3. Feature Sliding Windows: Adjust sliding window length $L$, overlap ratio, and temporal stacking $W$ via the sliders to alter the epoch size and history depth. You will immediately observe changes in calculated features (Hjorth metrics, line length, and spectral entropy) in the dynamic panels, as well as shifts in the SVM scatter space and decision thresholds on the feature canvas.
4. Running the Demo: Click "Start Seizure Demo" to run an automated diagnostic progression showing transition states from background to spindle, seizure, and post-ictal artifact. Adjusting any manual slider will instantly halt the demo and restore your manual settings.
Technical Details
To capture the temporal, spectral, and spatial evolution of brain activity, continuous signals are divided into structured, overlapping sliding temporal windows of length L seconds. Within each epoch, we compute a diverse set of signal representations:
1. Time-Domain Statistical Markers: The statistical distribution of signal amplitudes provides immediate insight into the baseline state of cortical networks. Basic parameters include the Mean, Standard Deviation, and Root Mean Square (RMS) voltage.
2. Hjorth Parameters of Activity, Mobility, and Complexity: Designed by Bo Hjorth, these metrics describe the statistical properties of a signal in the time domain, mapping directly to physical concepts of signal power, average frequency, and bandwidth:
$$\text{Activity} = \sigma_x^2 = \text{Var}(x(t))$$
$$\text{Mobility} = \sqrt{\frac{\text{Var}(x'(t))}{\text{Var}(x(t))}}$$
$$\text{Complexity} = \frac{\text{Mobility}(x'(t))}{\text{Mobility}(x(t))}$$
Where x'(t) is the first derivative of the signal. Seizure activity generally presents a substantial surge in overall Activity (amplitude variance) paired with a collapse in Complexity, reflecting highly repetitive, synchronized oscillatory behavior.
3. Line Length (LL): Line length measures the total vertical distance traveled by the signal over a window. It acts as a highly sensitive and computationally efficient tracker of high-frequency rhythmic spikes, which increase the cumulative trajectory length:
$$LL = \sum_{t=1}^{N-1} |x(t+1) - x(t)|$$
4. Power Spectral Density (PSD) and Spectral Entropy: Scalp EEG rhythms span classic bands: Delta ($0.5-4$ Hz), Theta ($4-8$ Hz), Alpha ($8-12$ Hz), Beta ($12-30$ Hz), and Gamma ($>30$ Hz). Normal relaxed states display prominent posterior alpha waves, while sleep is characterized by sleep spindles ($11-14$ Hz rhythmic waves). Seizures typically manifest as slow, high-amplitude rhythmic bursts in the Delta or Theta range. Spectral Entropy quantifies the chaotic distribution of power across these frequency components:
$$P_i = \frac{\text{PSD}(f_i)}{\sum_j \text{PSD}(f_j)}, \quad H_{\text{spec}} = -\sum_{i} P_i \ln(P_i)$$
A flat, white noise signal possesses high spectral entropy, whereas a localized seizure oscillation causes a reduction in entropy as energy condenses into a narrow spectral peak.
5. The Shoeb ICML 2010 Pipeline: As outlined by Ali Shoeb and John Guttag (ICML 2010), high-accuracy seizure detection demands a combination of spatial and spectral parameters. Rather than relying on a single channel, the Shoeb architecture constructs a unified multidimensional feature vector 𝒳. First, a bank of M = 8 bandpass filters spanning $0.5$ to $25$ Hz captures frequency-specific energy. To incorporate spatial geometry, this process is repeated across all N = 18 channels, constructing a single epoch vector of dimension $144$. Temporal stacking aggregates W = 3 windows for a stacked dimension of 432. Classification is resolved using an RBF SVM:
$$f(\mathcal{X}) = \sum_i \alpha_i y_i \exp\left(-\gamma \|\mathcal{X} - \mathcal{X}_i\|^2\right) + b$$
Future Directions
The next frontier of seizure monitoring involves deploying feature extraction pipelines directly onto wearable, low-power edge hardware. High-fidelity computation of Hjorth parameters and Line Length requires fewer CPU resources compared to intensive Fourier Transforms or deep neural networks, making them ideal candidates for low-power microcontroller deployment. Additionally, combining scalp EEG feature streams with secondary physiological metrics, such as Heart Rate Variability (HRV) derived from continuous Electrocardiograms (ECGs), can significantly decrease false alarm rates. During the onset of many clinical seizures, sympathetic nervous activation produces a rapid heart rate acceleration before distinct cortical spike-and-wave patterns physically manifest in frontal scalp electrodes. Incorporating cross-modal parameters into the SVM feature space unlocks highly resilient patient-specific prediction arrays.
Explore Related Biomedical Signal Environments
-
SeizureSim Visualizer
Dynamic cellular cortical fields mapping real-time micro-grid discharge states and 3D visual propagation loops.
-
EEG Cocktail Party Source Separator
Apply Independent Component Analysis (ICA) and Principal Component Analysis (PCA) to blind source EEG mixtures.
-
Neurofeedback Training Lab
Interactive brain-computer interface simulator measuring real-time focus metrics via spectral alpha/beta band ratios.
-
SleepSim Circadian Dashboard
Interactive hypnogram modeling normal sleep architecture, delta wave generation, and micro-arousal event loops.