Optimized driver fatigue detection method using multimodal neural networks
DOI: 10.1038/s41598-025-86709-1
URL: https://www.nature.com/articles/s41598-025-86709-1
archive: archived pipeline: cataloged verified
Summary
Multimodal deep-learning study (Scientific Reports 2025) detecting driver fatigue from physiological and facial data in the DROZY sleep-deprivation dataset. Authors compare a feature-combination model that concatenates EEG, ECG, and facial-image features against a feature-coupled model in which each modality dynamically weights the others contributions. A majority-voting ensemble aggregates predictions for the deployment phase.
Key finding
A multimodal feature-coupled neural network that lets EEG, ECG, and facial features mutually weight one another reaches 98.41% accuracy (precision 98.38%, recall 98.39%, F1 98.38%) on DROZY, outperforming a parallel feature-combination model (94.87% accuracy) and demonstrating the benefit of cross-modal coupling for fatigue classification.
Methodology
Multimodal deep neural networks trained on the DROZY sleep-deprivation dataset (EEG, ECG, facial images); two architectures compared (feature combination vs. feature coupling with mutual-weight mechanism); majority-voting ensemble in the decision phase.
Sample size: DROZY dataset (sleep-deprivation lab data; sample size per dataset specification)
Quality score: 5 / 5