1. Overview: Wavelet Scattering & Non-Stationary Seizure Electrophysiology
Epilepsy is a complex neurological disorder characterized by recurrent, unprovoked paroxysmal synchronized discharges across neuronal cortical assemblies. Conventional scalp Electroencephalography (EEG) records extracellular macroscopic electrical potentials, yet clinical seizure detection is severely hindered by non-stationarity, biological artifacts (electromyographic, ocular, and baseline wander), and intense structural variability across patient populations.
Traditional Fourier transform techniques ($\mathcal{F}$) operate under the assumption of wide-sense stationarity, obscuring localized high-frequency transients within large window averages. Conversely, standard Continuous Wavelet Transforms (CWT) provide time-frequency localization but lack spatial deformation stability. To overcome these constraints, the Wavelet Scattering Transform (WST), pioneered by Stéphane Mallat, constructs a translation-invariant, deformation-stable representation by cascading complex wavelet filter banks with non-linear modulus operators and spatial low-pass averaging.
The scattering operator linearizes small time-warping deformations $\tau(t)$, which frequently characterize the temporal jitter and morphological dilation of 3 Hz generalized spike-and-wave discharges (GSWD) and focal paroxysmal fast activity. By extracting coefficients across multiple scattering orders ($m=0, 1, 2$), the framework preserves high-frequency transient energy while maintaining structural invariance:
$$S_J x(t, j_1) = |x * \psi_{j_1}| * \phi_J(t)$$
where $\psi_{j_1}(t)$ represents a dilated analytic bandpass wavelet, $|\cdot|$ is the non-linear complex modulus acting as a full-wave rectifier, and $\phi_J(t)$ is a scaling low-pass filter establishing spatial averaging invariance up to scale $2^J$.
2. Operational Guidelines & Telemetry Workflow
This interactive laboratory allows researchers, clinical neurophysiologists, and biomedical engineers to explore how cascading wavelet operators decouple transient epileptiform spikes from stochastic background electrocortical noise:
- Generating Synthetic Epochs: Click the "⚡ GENERATE NEW EPOCH" button to compute a new realization of non-stationary EEG containing stochastic $1/f^\alpha$ pink noise and Markov-clustered epileptic spike-wave bursts.
- Scale Parameter ($J$) Tuning: Manipulate the Scale Parameter ($J$) slider ($1 \le J \le 10$). Notice that as $J$ increases, the scaling filter $\phi_J(t)$ expands its integration window, smoothing high-frequency ripples while accentuating long-range envelope modulation and slow-wave after-discharges.
- Morphological Presets:
- Interictal Normal Rhythm: Baseline awake EEG characterized by dominant background $1/f$ pink noise with stochastic low-amplitude micro-transients.
- Focal Clustered Spikes: Simulates focal temporal lobe paroxysms exhibiting localized, high-velocity spike clusters followed by slow-wave repolarization envelopes.
- Generalized 3 Hz Spike-and-Wave: Canonical absence seizure electrophysiology exhibiting high-amplitude rhythmic spike-and-dome trajectories.
- Burst Suppression: Alternating cycles of high-voltage multi-frequency paroxysmal discharges and profound isoelectric suppression.
- Paced Demonstration Mode: Click "Start Demo" at the top of the sidebar to initiate an automated physiological sequence demonstrating baseline rest, acute focal ictal onset, scale parameter modulations, and spectral domain transitions. Interacting with any control instantly restores manual configuration authority.
3. Mathematical Derivations & Scattering Filter Architecture
The scattering transform constructs deep convolutional representations without requiring parameter learning via backpropagation. Given an input discrete EEG signal $x(t) \in \mathbf{L}^2(\mathbb{R})$, the zero-th order scattering coefficient corresponds to the global moving average:
$$S_0 x(t) = x * \phi_J(t)$$
The first-order scattering coefficients $S_1 x(t, \lambda_1)$ capture localized high-frequency envelope dynamics by convolving with dilated wavelets $\psi_{\lambda_1}(t)$ centered at frequency $\lambda_1 = 2^{-j_1 / Q_1}$, rectifying via the complex modulus, and low-pass filtering:
$$U_1 x(t, \lambda_1) = |x * \psi_{\lambda_1}(t)|$$
$$S_1 x(t, \lambda_1) = U_1 x * \phi_J(t) = |x * \psi_{\lambda_1}| * \phi_J(t)$$
Discrete Fourier Transform (DFT) & Spectral Energy
To monitor energy redistribution across frequency bins $k$, the simulator evaluates the discrete Fourier transform of the signal and its scattering envelope:
$$X[k] = \sum_{n=0}^{N-1} x[n] e^{-i 2\pi k n / N}, \quad k = 0, 1, \dots, N-1$$
The Seizure Spike Index ($\Gamma_{\text{ictal}}$) is computed as the normalized kurtosis of the first-order scattering envelope:
$$\Gamma_{\text{ictal}} = \frac{\frac{1}{N}\sum_{n=1}^{N} (S_1 x[n] - \mu_S)^4}{\left(\frac{1}{N}\sum_{n=1}^{N} (S_1 x[n] - \mu_S)^2\right)^2} - 3$$
High positive kurtosis values ($\Gamma_{\text{ictal}} > 2.5$) indicate sparse, heavy-tailed paroxysmal spike bursts characteristic of ictal transitions, whereas background noise yields near-zero Gaussian excess kurtosis.
4. Future Directions: Multi-Channel EEG & Clinical Deep Learning
The translation of scattering representations into embedded closed-loop neuro-stimulators (e.g., Responsive Neurostimulation, RNS) is advancing across several fronts:
- Joint Time-Frequency Scattering (JTFS): Cascading frequency-domain convolutions across the time-frequency plane $(\lambda_1, t)$ to capture continuous spectro-temporal modulations and chirping patterns during seizure propagation.
- Low-Power Edge Hardware Implementation: Deploying integer-quantized scattering filter banks directly onto ultra-low-power neuromorphic microcontrollers for real-time seizure forecasting with sub-millijoule energy footprints.
- Hybrid Scattering-Attention Networks: Coupling frozen, physics-grounded second-order scattering coefficients ($S_2 x$) as fixed feature extractors for Transformer backbones to eliminate overfitting on small pediatric clinical cohorts.
Explore Related Neurotech Environments
Open Access License: This interactive educational module is released under
CC BY-NC 4.0 (Attribution-NonCommercial)
for non-commercial research, academic study, and clinical education.
Commercial & Enterprise Licensing: For white-labeling, proprietary LMS/course embedding, hardware dashboard telemetry integration, or custom feature engineering, secure a commercial license at
BioniCloud.com or contact
Dr. Yuri Beno.