Achieving Reproducibility in EEG-Based Machine Learning
Sean P. Kinahan, Pouria Saidi, Ayoub Daliri, Julie Liss, Visar Berisha
Abstract
Despite the inherent complexity of electroencephalogram (EEG) data characterized by its high dimensionality, artifactual noise, and biological variability, many machine learning (ML) studies claim impressive performance in decoding or classifying EEG signals. Recently, several studies have highlighted that flawed data analysis is a prevalent issue in the literature, leading to irreproducible results and exaggerated claims. To address this issue, we propose a framework that addresses three primary obstacles in EEG ML research: data leakage, data scarcity, and flawed model selection. We introduce the EEG ML Model Card, a standardized and transparent EEG ML model documentation tool that aims to directly address these pitfalls and enhance reproducibility and trustworthiness in EEG ML research.