Semi-automated seizure detection using interpretable machine learning models

TL;DR

This paper introduces SeizyML, an open-source tool that uses interpretable machine learning to detect seizures from electrographic recordings, reducing manual effort and bias.

Contribution

The novel contribution is SeizyML, an open-source, semi-automated seizure detection tool using interpretable machine learning models validated on both mouse and human datasets.

Findings

Gaussian Naïve Bayes and Stochastic Gradient Descent models achieved the highest precision and F1 scores in seizure detection.

SeizyML detected all seizures in the mouse dataset with minimal training data.

The approach was successfully applied to a human EEG dataset, demonstrating its broader utility.

Abstract

Despite the vast number of seizure detection publications there are no validated open-source tools for automating seizure detection based on electrographic recordings. Researchers instead rely on manual curation of seizure detection that is highly laborious, inefficient, error prone, and heavily biased. Here we developed an open-source software called SeizyML that uses sensitive machine learning models coupled with manual validation of detected events reducing bias and promoting efficient and accurate detection of electrographic seizures. We compared the validity of four interpretable machine learning models (decision tree, gaussian naïve bayes, passive aggressive classifier, and stochastic gradient descent classifier) on an extensive electrographic seizure dataset that we collected from chronically epileptic mice. We find that the gaussian naïve bayes and stochastic gradient descent models achieved the highest precision and f1 scores, while also detecting all seizures in our mouse dataset and only require a small amount of data to train the model and achieve good performance. Further, we demonstrate the utility of this approach to detect electrographic seizures in a human EEG dataset. This approach has the potential to be a transformative research tool overcoming the analysis bottleneck that slows research progress.

Figures