Interpretable Machine Learning for Seizure Detection
Summary Here we describe the research behind the development of SeizyML, an open-source tool designed to automate the detection of electrographic seizures using interpretable machine learning models. It supports manual validation to reduce bias and was tested across a large dataset of chronically epileptic mice.
Key findings:
- Interpretable machine learning models can achieve 100% seizure detection accuracy in mice, suggesting their potential for automating seizure detection in research.
- Per-file normalization and model type were the biggest contributors to seizure detection performance.
- We compared four models and found Gaussian Naïve Bayes to be the most accurate and resilient, requiring minimal training data.
Authors:
Pantelis Antonoudiou, Trina Basu & Jamie L. Maguire