Modular Multi-Classifier Framework for Robust Seizure Detection from Superficial EEG Signals

Seizure detection from EEG signals remains a major clinical and technical challenge due to the variability across seizure types and the frequent presence of noise and artifacts in superficial (non-invasive) recordings. Standard approaches often rely on a single classifier trained on all types of seizures, leading to poor generalization and sensitivity, particularly for rare or atypical seizure patterns.

This project proposes a multi-classifier architecture in which separate models are trained to detect specific types or subtypes of seizures. In parallel, a dedicated artifact classifier is trained to identify and isolate signal segments contaminated by muscle activity, eye movements, or environmental noise. The output of the artifact model is used to subtract or mask noise, yielding a cleaner EEG signal that is passed to the ensemble of seizure classifiers.

Multiple classifiers can be assigned to the same seizure subtype, allowing redundant detection. Final decisions are made using a majority voting scheme, which flags a seizure when at least one or more classifiers consistently detect an event. This flexible architecture allows better handling of seizure heterogeneity and improves robustness in real-world, noisy EEG environments.

The project aims to explore this modular architecture on open EEG datasets (e.g., TUH EEG Seizure Corpus) and evaluate improvements in sensitivity, specificity, and robustness across seizure categories.

 

Methodology

• - Train a dedicated artifact detection model to identify and mask/remove contaminated segments.

• - Define a set of seizure-specific classifiers, each trained on a subset of seizure types (e.g., focal vs. generalized, tonic-clonic vs. absence).

• - Implement a majority voting or confidence-weighted decision system to combine outputs from the seizure classifiers.

• - Evaluate the system’s performance (sensitivity, specificity, false positive rate) on benchmark datasets.

 

Expected Outcomes

• - A modular, extensible EEG classification framework for seizure detection.

• - Insight into which seizure types benefit from individualized classifiers.

• - Demonstrated benefit of artifact subtraction and redundant classification via voting.

• - A scalable foundation for patient-specific tuning or clinical translation.

 

Available Material

• - Public EEG seizure datasets

• - Neuromorphic chip

 

Contacts

Elisa Donati (elisa@ini.uzh.ch)