DrowsyDG-Phys, a novel domain generalization framework using physiological signals, achieves 78.5% accuracy on the domain generalization protocol and 88.4% accuracy on the cross-subject protocol for driver drowsiness detection, outperforming baseline methods.
Key Findings
Results
DrowsyDG-Phys achieved 78.5% accuracy on the domain generalization protocol and 88.4% accuracy on the cross-subject protocol.
The model was evaluated using a multi-source domain generalization benchmark across three existing datasets and one self-collected dataset.
The domain generalization protocol tests the model's ability to generalize across datasets with different recording conditions and populations.
The cross-subject protocol evaluates generalization across individual participants within datasets.
DrowsyDG-Phys outperformed all baseline methods on both evaluation protocols.
Methods
The DrowsyDG-Phys framework uses three physiological signals — electrocardiogram (ECG), electrodermal activity (EDA), and respiration — measurable by in-vehicle or wearable sensors.
These signals were chosen because they can be unobtrusively collected in real-world driving contexts.
The framework extracts both time domain and frequency domain features from these signals via a backbone network.
Traditional physiological-signal-based methods relied on manually extracted features processed through machine learning algorithms, which lacked flexibility and robustness.
Methods
The framework integrates three novel loss functions to improve generalization and robustness.
The first is a prior knowledge-based contrastive regularization loss designed to enhance model robustness.
The second is a feature centralization loss to promote generalization across heterogeneous data distributions.
The third is a novel loss function specifically designed to align drowsiness assessment criteria across datasets.
These loss functions were introduced to address domain shift problems that limit the generalization of existing deep learning models.
Methods
A self-collected dataset of 60 participants in a simulated SAE Level-3 driving scenario was included in model evaluation.
The scenario involved conditional automated driving (SAE Level 3), where the driver may need to take over control from the automated system.
This dataset was combined with three existing publicly available datasets to form a multi-source domain generalization benchmark.
The inclusion of a Level-3 automated driving scenario addresses the relevance of drowsiness monitoring in contemporary vehicle automation contexts.
Background
Domain shift was identified as a key limitation of existing physiological signal-based drowsiness detection models, motivating the domain generalization approach.
Recent deep learning approaches improved detection accuracy through automated feature extraction but remained limited in generalization due to domain shifts.
Domain shifts arise from differences in recording equipment, participant populations, experimental protocols, and environmental conditions across datasets.
The DG framework was specifically designed to address these cross-domain variability challenges.
Results
DrowsyDG-Phys demonstrated improved generalization and robustness compared to baseline physiological signal-based drowsiness monitoring methods.
Experimental results confirmed that DrowsyDG-Phys outperformed all baseline methods tested.
The model's performance improvements were observed on both the domain generalization and cross-subject evaluation protocols.
The backbone network was designed for explicit time and frequency domain feature learning to support robust signal representation.
What This Means
This research suggests that a new artificial intelligence framework called DrowsyDG-Phys can detect driver drowsiness from physiological signals — specifically heart activity (ECG), skin conductance (EDA), and breathing patterns — with meaningful accuracy even when tested on data from entirely different studies or participants than those used during training. The system was built to tackle a common problem in this field: models that work well in a lab setting often fail when applied to real-world conditions or new groups of people, a challenge known as 'domain shift.' By combining specialized feature extraction with three new mathematical techniques that help the model focus on the most generalizable patterns, DrowsyDG-Phys reached 78.5% accuracy when tested across different datasets and 88.4% accuracy when tested across different individuals within a dataset.
The study is particularly relevant to modern cars with partial automation — sometimes called Level 3 or 'conditional automated' vehicles — where a driver may be less actively engaged but still needs to be alert enough to take over control when needed. The researchers tested their model on data from 60 participants in a simulated Level-3 driving environment, in addition to three existing public datasets, giving the evaluation a broader and more realistic foundation than studies relying on a single dataset.
This research suggests that physiological signals collected through wearable or in-vehicle sensors could form the basis of robust drowsiness monitoring systems that work reliably across different people and conditions — potentially contributing to safer roads by providing earlier and more consistent warnings before a drowsy driver causes an accident. The open availability of the benchmark and evaluation protocol may also help future researchers improve upon these results.
Wang J, Li W, Wang Z, Ayas S, Donmez B, He D, et al.. (2026). DrowsyDG-Phys: Generalizable driver drowsiness estimation in conditional automated vehicles using physiological signals.. Accident; analysis and prevention. https://doi.org/10.1016/j.aap.2026.108407