Wearable, Real-time Drowsiness Detection based on EEG-PPG Sensor Fusion at the Edge

• Published in: Biomedical Circuits and Systems Conference pp. 1 - 5
• Main Authors: Frey, Sebastian, Rapa, Pierangelo Maria, Amidei, Andrea, Benatti, Simone, Guermandi, Marco, Kartsch, Victor, Cossettini, Andrea, Benini, Luca
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.10.2024
• Subjects: Accuracy; Biomedical monitoring; Convolutional neural networks; drowsiness detection; Electroencephalography; embedded system; Image edge detection; Industries; Power demand; Real-time systems; Sensor fusion; ultra-low-power design; Wearable devices; wearable EEG; wearable PPG
• ISSN: 2766-4465
• DOI: 10.1109/BioCAS61083.2024.10798380

Drowsiness and fatigue pose significant risks across various industries, causing 15-20% of severe road crashes in the driving industry alone. Wearable devices are a promising approach for detecting drowsiness since they do not require modifications to existing vehicles. Common wearable approaches are based on PPG signals, and EEG-based solutions are also gaining more popularity. However, poor signal quality, user discomfort, and lack of computational capabilities for low-power processing at the edge hinder the development of standalone wearable drowsiness detection systems. This paper introduces a novel drowsiness detection system based on BioGAP, a compact acquisition and processing platform for heterogeneous biosignals powered by the GAP9 parallel ultra-low-power System-onChip. BioGAP is integrated into a comfortable, non-stigmatizing headband and acquires data from 8 fully-dry EEG channels and a PPG sensor on the earlobe. Five subjects performed drowsiness experiments wearing the headband while engaged in a professional car driving simulator. A lightweight Convolutional Neural Network (CNN) enables sensor fusion between EEG and PPG, achieving an average accuracy of 91.1% in detecting drowsy states (+\mathbf{6 \%} increase compared to using a single modality only). The network is lightweight (21.6 k parameters) and is deployed on GAP9, demonstrating an inference time of 10.17 ms, energy per inference of 0.36 mJ, and average system power consumption of only 19.6 mW, thereby enabling continuous operation for 14 h in realistic driving conditions when powered by a small 75 mAh battery.