Drowsiness Detection in Drivers: A Systematic Review of Deep Learning-Based Models

Deep learning (DL) models show considerable promise in detecting driver drowsiness, a major contributor to road traffic crashes. This systematic review evaluates the performance, contexts of application, and implementation challenges of DL-based drowsiness detection systems. Conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, the review includes peer-reviewed empirical studies published between 2015 and 2025 that develop and validate DL models using data collected in real or simulated driving environments. Studies were identified through systematic searches in PubMed, Scopus, Web of Science, ScienceDirect, and IEEE Xplore, last updated in March 2025. Due to methodological heterogeneity, findings are synthesized narratively. Eighty-one studies meet the inclusion criteria. Most employ Convolutional Neural Networks, Recurrent Neural Networks, or hybrid architectures and use behavioral, physiological, or multimodal inputs. Reported median values for accuracy and F1-score exceed 0.95 under both simulated and real-world conditions. However, studies frequently lack demographic diversity, standardized performance reporting, and robust validation protocols. Key limitations include limited dataset transparency, inconsistent evaluation metrics, and insufficient attention to ethical and privacy considerations. While DL models exhibit strong predictive performance, their real-world deployment remains limited by practical and methodological constraints. Future research should place emphasis on the development of inclusive datasets, the conduct of multi-context evaluations, the advancement of real-world deployment strategies, and the rigorous adherence to ethical standards.