Designing an Embedded Feature Selection Algorithm for a Drowsiness Detector Model Based on Electroencephalogram Data.

Area of Science:

• Neuroscience
• Computer Science
• Automotive Engineering

Background:

• Driver fatigue is a major safety concern, hindering the adoption of autonomous vehicles.
• Effective drowsiness detection requires efficient processing of complex physiological data like electroencephalogram (EEG).
• High-dimensional feature sets pose challenges for machine learning model performance and interpretability.

Purpose of the Study:

• To develop an embedded feature selection (FS) algorithm for enhancing neural network-based driver drowsiness detection.
• To reduce the dimensionality of EEG-based drowsiness indicators for improved model efficiency.
• To create a foundational component for robust driver monitoring systems in automated driving.

Main Methods:

• An embedded feature selection algorithm using a 'Feature Prune Layer' integrated into a neural network architecture.
• Iterative adjustment and deletion of feature weights based on importance until a target feature subset is achieved.
• Validation using electroencephalogram (EEG) data, a key physiological indicator of drowsiness.

Main Results:

• The proposed FS algorithm achieved a 95% reduction in the original feature set with only a 1% loss in precision.
• Selecting the top 10% and 20% of features resulted in precision increases of 1.5% and 2.7%, respectively.
• Outperformed Principal Component Analysis (PCA) and Chi-squared test by achieving 24.3% and 3.2% higher precision, respectively, at 95% feature reduction.

Conclusions:

• The embedded FS algorithm effectively reduces feature dimensionality for drowsiness detection without compromising accuracy.
• This method offers a significant improvement over existing feature selection techniques for EEG-based drowsiness monitoring.
• The developed algorithm serves as a valuable building block for safer and more reliable driver assistance and autonomous driving systems.