A Spatial Feature-Engineering Framework for EEG-Based Eye State Classification using Within-Sample Inter-Channel Descriptors and Extra Trees

Document Type : Original Article

Authors

1 Faculty of Engineering & Technology, University of Mazandaran, Babolsar, Iran

2 University of Mazandaran, Babolsar, Iran.

Abstract
This is a spatial feature engineering framework for EEG based eye state classification, with no spectral transforms and no dimensionality reduction. Each observation is represented by within-sample, inter-channel descriptors: per sample statistics, spatial z-scores, hemispheric asymmetry differences, region level aggregates and pairwise absolute channel differences. All of the descriptors are derived from a single time point and standardisation is fit only on the training folds. A class balanced Extra Trees ensemble is used to deal with the very mild 1.23:1 imbalance without any synthetic oversampling. On the UCI EEG Eye State dataset, using stratified 10 fold cross validation we achieve 97.54% (±0.36%) accuracy, 97.24% F1 and 99.74% AUC, compared to 95.11% and 94.44% for the raw channel Extra Trees baseline (Wilcoxon signed rank test, p=0.00195). Under a more stringent temporal block protocol the accuracy is 82.38%, which quantifies the gap between the benchmark performance and the ability to generalise to unseen segments. On the same folds it beats six re-implemented approaches from the literature under the temporal block protocol, with the smallest drop between the two. An ablation study shows the gain is largely attributable to the inter electrode descriptors, in particular the pairwise differences which take the temporal block F1 from 74.24% to 78.96%. The representation is interpretable and reproducible. Given that the benchmark is on a single subject its utility for monitoring applications remains to be seen on multi subject recordings.

Keywords

Subjects

Volume 3, Issue 2 - Serial Number 2
January 2026
Pages 153-163

  • Receive Date 14 October 2026
  • Revise Date 12 November 2026
  • Accept Date 22 November 2026
  • First Publish Date 22 November 2026
  • Publish Date 01 January 2026