Semi-automated seizure detection using interpretable machine learning models

Numerous methods have been proposed for seizure detection automation, yet the tools to harness these methods and apply them in practice are limited. Here we compare four interpretable and widely-used machine learning models (decision tree, gaussian naive bayes, passive aggressive classifier, stochastic gradient descent classifier) on an extensive electrographic seizure dataset collected from chronically epileptic mice. We find that the gaussian naive bayes model achieved the highest precision and f1 score, while also detecting all seizures in our dataset and only requires a small amount of data to train the model and achieve good performance. We use this model to create an open-source python application SeizyML that couples model performance with manual curation allowing for efficient and accurate detection of electrographic seizures. Author SummarySeizure detection based on electrographic recordings is critical for epilepsy diagnosis and research. However, the current gold standard for seizure detection is manual curation, which is biased, costly, incredibly laborious, and requires extensive training and expertise, prohibiting advances in epilepsy diagnosis and research. Here we demonstrate that fast, simple, and interpretable machine learning (ML) models are sufficient to detect all seizures in an extensive dataset collected from a well-established mouse model of epilepsy. Importantly, we created an open-source python application, SeizyML, that integrates the most precise model tested here with human curation of the detected events. This semi-automated approach greatly enhances efficiency and precision of seizure detection while also being transparent. We believe that the adoption of semi-automated and transparent technologies is indispensable for understanding ML model predictions, improving their reliability, and fostering trust between ML models and neurodiagnostic professionals.