BCIwiki (bciwiki.com) — Researchers at Universitas Budi Luhur in Indonesia used the UCI Machine Learning Repository's EEG-Eye-State dataset to detect open and closed eye states with high accuracy, applying interquartile range (IQR) clipping for outlier handling combined with a random forest classifier, achieving 92.49% accuracy. The study was published in Sinkron on July 27, 2026.
The team applied IQR clipping (bounds [Q1 − 1.5×IQR, Q3 + 1.5×IQR]) to 14,980 samples across 14 EEG sensor features, handling 12,737 outlier instances without discarding any samples, followed by StandardScaler normalization and an 80:20 stratified train-test split. The random forest model, using 100 estimators and balanced class weights, was validated with stratified 10-fold cross-validation, achieving a balanced accuracy of 92.14%, ROC-AUC of 0.9791, PR-AUC of 0.9759, macro F1-score of 0.9236, Matthews correlation coefficient of 0.8486, and Cohen's kappa of 0.8474. Cross-validation confirmed stability with a mean accuracy of 92.86% ± 0.36% and ROC-AUC of 0.9809 ± 0.0020.
Feature importance analysis identified sensors O1 (11.81%), P7 (10.59%), and F7 (10.15%) as the most dominant contributors, consistent with neuroanatomical regions involved in visual processing. The researchers concluded that combining IQR clipping with random forest produces a stable, accurate, and neuroanatomically interpretable model for EEG-based eye state classification, offering a foundation for real-world BCI and driver drowsiness detection systems.