[Paper Review] Topological biomarkers for real-time detection of epileptic seizures

Ximena Fernández, Diego M. Mateos|arXiv (Cornell University)|Nov 4, 2022

Topological and Geometric Data Analysis4 citations

TL;DR

This paper proposes a real-time method for detecting epileptic seizures using topological biomarkers derived from persistent homology of sliding-window embeddings of multichannel neurophysiological signals (EEG, MEG, iEEG). By analyzing the first derivative of persistence diagrams over time, the approach detects dynamic changes associated with seizure onset and termination with high accuracy across diverse clinical settings and recording types.

ABSTRACT

Real time seizure detection is a fundamental problem in computational neuroscience towards diagnosis and treatment's improvement of epileptic disease. We propose a real-time computational method for tracking and detection of epileptic seizures from raw neurophysiological recordings. Our mechanism is based on the topological analysis of the sliding-window embedding of the time series derived from simultaneously recorded channels. We extract topological biomarkers from the signals via the computation of the persistent homology of time-evolving topological spaces. Remarkably, the proposed biomarkers robustly captures the change in the brain dynamics during the ictal state. We apply our methods in different types of signals including scalp and intracranial electroencephalograms and magnetoencephalograms, in patients during interictal and ictal states, showing high accuracy in a range of clinical situations.

Motivation & Objective

• To develop a real-time computational method for automated seizure detection from raw neurophysiological recordings.
• To identify topological biomarkers that robustly capture transitions in brain dynamics during the ictal state.
• To enable detection across diverse recording modalities (scalp EEG, intracranial EEG, MEG) and clinical conditions (focal vs. generalized seizures).
• To overcome limitations of traditional preprocessing-heavy machine learning methods by enabling streaming, online analysis.
• To provide a consistent, geometry-informed framework applicable to different data acquisition methods and patient-specific dynamics.

Proposed method

• Apply Takens' delay embedding to each channel's time series to reconstruct the state space trajectory via sliding-window embeddings in R^D.
• Compute persistent homology on the time-evolving point clouds from the embeddings to generate persistence diagrams.
• Approximate the first derivative of the persistence diagram's evolution to quantify changes in topological structure over time.
• Use the derivative values as biomarkers: high values indicate seizure onset and termination phases.
• Apply the method independently per channel to detect focal vs. generalized seizure dynamics based on spatiotemporal consistency of biomarker peaks.
• Leverage algebraic topology to extract intrinsic geometric features of neural dynamics without requiring prior knowledge of underlying equations.

Experimental results

Research questions

• RQ1 Can topological features derived from persistent homology detect epileptic seizures in real time from raw neurophysiological signals?
• RQ2 How do topological biomarkers reflect changes in brain dynamics during the transition from interictal to ictal states?
• RQ3 Can the method distinguish between focal and generalized seizures using channel-wise biomarker patterns?
• RQ4 Does the approach maintain robustness across different recording types (EEG, MEG, iEEG) and patient-specific signal characteristics?
• RQ5 Can the first derivative of the persistence diagram evolution serve as a reliable indicator of seizure onset and offset?

Key findings

• The method achieves high accuracy in real-time detection of epileptic seizures across multiple types of neurophysiological recordings, including scalp EEG, intracranial EEG, and MEG.
• Topological biomarkers based on the first derivative of persistence diagrams successfully identify seizure onset and termination, with salient peaks marking these transitions.
• For focal seizures, the biomarker shows a leading peak in the channel where the seizure originates, followed by sequential activation in other channels.
• For generalized seizures, all channels exhibit synchronized peaks in the biomarker derivative at seizure onset and offset, indicating global synchronization.
• The approach is robust across different recording modalities and clinical scenarios, demonstrating consistency in detecting seizures regardless of signal type or patient-specific variability.
• The use of persistent homology on sliding-window embeddings captures intrinsic geometric changes in neural dynamics that are not detectable through conventional time-frequency or complexity-based methods.