Validation and interpretation of a multimodal drowsiness detection system using explainable machine learning.

Area of Science:

• Neuroscience
• Machine Learning
• Road Safety Engineering

Background:

• Drowsy driving is a significant road safety hazard.
• Existing detection systems often function as "black boxes," lacking robustness and explainability.
• There is a need for trustworthy machine learning models in drowsy driving detection.

Purpose of the Study:

• To rigorously validate drowsy driving detection systems using multiple techniques.
• To enhance model trustworthiness through explainable machine learning.
• To interpret the physiological signals contributing to drowsiness detection.

Main Methods:

• Simulated driving task with physiological signal recording (EEG, EOG, ECG).
• Application of subject-dependent and independent validation techniques.
• Utilized K-nearest neighbours, support vector machines, and random forest classifiers.
• Employed SHapley Additive exPlanation (SHAP) and partial dependency analysis (PDA) for model interpretation.

Main Results:

• Subject-independent validation (leave one participant out) proved most effective.
• Random forest classifier achieved 70.3% sensitivity, 82.2% specificity, and 80.1% accuracy.
• Explainable methods identified key physiological features for drowsiness detection.

Conclusions:

• The study provides a robust validation and explainable approach for trustworthy drowsiness detection systems.
• Explainable machine learning enhances the reliability and interpretability of in-vehicle systems.
• This approach promises improved road safety with potentially lower system costs.