A multivariate approach for drowsiness detection using empirical Fourier decomposition

Abstract

Drowsiness is a critical challenge for road safety, which can be addressed by an automatic alertness monitoring system. In this study, our goal is to investigate the multivariate characteristics of electroencephalogram (EEG) and electrooculography (EOG) signals on adaptive frequency scales for drowsiness detection. We have extended the empirical Fourier decomposition (EFD) method for multivariate signals. The proposed multivariate EFD (MEFD) is studied on synthetic and real EEG signals, and it provides the proper mode alignment of frequency bands across channels. The multi-channel EEG and EOG signals are decomposed into oscillatory modes using the proposed MEFD. From each mode, eighteen features were extracted, which include statistical, Hjorth-based, energy-based, and entropy-based features. The significance of features is assessed using the analysis of variance (ANOVA) method. We develop machine learning classifier models based on significant features. The developed subject-independent drowsiness detection framework based on multimodal channel fusion achieves an accuracy of 94.3% for three classes: drowsy, tired, and awake. The performance of the proposed MEFD-based drowsiness detection framework is compared with state-of-the-art methods. The performance of the MEFD is compared with other multivariate decomposition techniques, which shows a significant improvement in classification accuracy for drowsiness detection. These findings demonstrate that MEFD is an efficient and reliable method for analyzing multimodal physiological signals and is a promising candidate for real-time fatigue monitoring. © 2025 Elsevier B.V., All rights reserved.