Robust Unsupervised Transient Detection With Invariant Representation based on the Scattering Network

Randall Balestriero, Behnaam AazhangarXiv 2016 · DOI · arXiv

Abstract

We present a sparse and invariant representation with low asymptotic complexity for robust unsupervised transient and onset zone detection in noisy environments. This unsupervised approach is based on wavelet transforms and leverages the scattering network from Mallat et al. by deriving frequency invariance. This frequency invariance is a key concept to enforce robust representations of transients in presence of possible frequency shifts and perturbations occurring in the original signal. Implementation details as well as complexity analysis are provided in addition of the theoretical framework and the invariance properties. In this work, our primary application consists of predicting the onset of seizure in epileptic patients from subdural recordings as well as detecting inter-ictal spikes.

Introduction

The analysis of intra-cranial EEG (iEEG) is critical for understanding epilepsy, yet it remains a significant challenge due to the high-dimensional and non-stationary nature of the signals. Biological markers such as inter-ictal spikes or high-frequency oscillations (HFOs)—collectively termed "transients"—are vital for seizure prediction but are often obscured by significant background noise and artifacts. Traditional supervised deep learning methods are frequently hampered by the lack of large, accurately labeled datasets. This work introduces an unsupervised framework designed to produce a parsimonious and invariant representation of these transients without requiring labeled training data.

Methodology

The proposed framework extends the Scattering Network architecture, which utilizes predefined wavelet filters to achieve translation invariance and stability against small deformations. To adapt this to the complexities of iEEG, two additional processing stages are introduced to generate the final representation, $Lx$:

  1. Nonlinear Thresholding ($\rho$): Second-layer scattering coefficients ($S_2x$) are processed using a threshold derived from the signal's median. This median serves as a robust centroid for "ambient noise." A polynomial nonlinearity is applied to values exceeding this threshold, effectively suppressing background noise while amplifying high-energy transients, resulting in a representation with at least 50% sparsity.
  2. Frequency Invariance ($l$): Standard scattering networks provide time invariance, but biological signals often exhibit frequency shifts. The authors introduce local dimensionality reduction across the frequency dimension ($\lambda_1$) using either Principal Component Analysis (PCA) or Max-Pooling. This ensures the detector remains stable despite frequency perturbations.
EEG Example 3 illustrating inter-ictal transients and movement artifacts that the unsupervised framework must distinguish.

Experimental Setup

The study utilizes subdural iEEG recordings from epileptic patients sampled at 1000 Hz. These recordings capture various physiological states, including pre-ictal, ictal, and post-ictal phases, alongside inter-ictal spikes. A significant challenge addressed by the model is the presence of non-neural artifacts, such as those caused by patient movement, which must be distinguished from actual neural transients.

Results and Discussion

The $Lx$ representation provides a 2D map (time vs. scale) that facilitates the qualitative visualization of neural state transitions. In real-time simulations using 1-minute windows, the transition from inter-ictal to pre-ictal states becomes visually distinct as energy shifts within the frequency-invariant map.

EEG Example 1 showing a clear ictal transition around the 11-minute mark across three channels.

Furthermore, the authors demonstrated that the first eigenvalues from the PCA step serve as effective state descriptors. These eigenvalues cluster into biologically meaningful groups, such as inter-ictal activity versus ictal states. When combined with a k-Nearest Neighbors (k-NN) clustering algorithm, the method successfully isolated individual spikes with high precision, achieving zero false negatives and zero false positives in the provided test cases.

Conclusion

This unsupervised approach offers a robust alternative to supervised models in data-scarce medical environments. By leveraging fixed wavelets and frequency invariance, the system adapts to patient-specific "ambient" energy levels. Future improvements include the implementation of a moving median to better handle long-term non-stationarity and exploring non-linear dimensionality reduction techniques like Mixture of Probabilistic PCA to further refine the representation of complex brain dynamics.

Summary updated

Abstract

We present a sparse and invariant representation with low asymptotic complexity for robust unsupervised transient and onset zone detection in noisy environments. This unsupervised approach is based on wavelet transforms and leverages the scattering network from Mallat et al. by deriving frequency invariance. This frequency invariance is a key concept to enforce robust representations of transients in presence of possible frequency shifts and perturbations occurring in the original signal. Implementation details as well as complexity analysis are provided in addition of the theoretical framework and the invariance properties. In this work, our primary application consists of predicting the onset of seizure in epileptic patients from subdural recordings as well as detecting inter-ictal spikes.
BibTeX
@misc{balestriero2016robust,
  author = {Randall Balestriero and Behnaam Aazhang},
  title = {{Robust Unsupervised Transient Detection With Invariant Representation based on the Scattering Network}},
  year = {2016},
  eprint = {1611.07850},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1611.07850},
}
Figure