Stochastic Sparse Sampling: A Variable-Length Time Series Classification Framework for Seizure Onset Zone...

Area of Science:

• Computational neuroscience
• Machine learning for healthcare
• Signal processing

Background:

• Variable-length time series classification (VTSC) is crucial in healthcare, particularly for analyzing electrophysiological recordings like EEG.
• Existing VTSC models face limitations: finite-context models risk data distortion and overfitting, while infinite-context models struggle with long-term dependencies and gradient stability.

Purpose of the Study:

• To introduce a novel VTSC framework, Stochastic Sparse Sampling (SSS), designed for accurate seizure onset zone (SOZ) localization.
• To address the challenges posed by variable-length electrophysiological data in identifying seizure-generating brain regions.

Main Methods:

• The proposed framework employs SSS to sparsely sample time series windows for local predictions.
• These local predictions are aggregated and calibrated to generate a global SOZ prediction.
• SSS facilitates post-hoc analysis by visualizing signal characteristics related to the SOZ.

Main Results:

• The SSS framework demonstrated superior performance compared to state-of-the-art baselines on the Epilepsy intracranial electroencephalography (iEEG) Multicenter Dataset.
• The method achieved better results across multiple medical centers and showed strong out-of-distribution generalization to unseen centers.

Conclusions:

• Stochastic Sparse Sampling (SSS) offers a robust and effective solution for seizure onset zone localization from variable-length electrophysiological data.
• The framework provides valuable insights and outperforms current methods, particularly in heterogeneous and out-of-distribution scenarios.