Driver Drowsiness Detection using Machine Learning and Deep Learning Techniques: A Systematic Review

Behavioral indicators have been increasingly utilized in machine learning (ML) and deep learning (DL) frameworks to enable automated driver drowsiness detection (DDD). This study aimed to investigate the available evidence on the modeling frameworks, datasets, input modalities, and performance metrics used in DDD systems.

were identified through systematic searches of PubMed, Scopus, Web of Science, and IEEE Xplore for English-language publications up to 31 August 2025. Eligible studies were original research that applied ML or DL techniques to image- or video-based behavioral features for automatic DDD. Studies focusing primarily on vehicle telemetry, physiological signals without behavioral imaging, reviews, editorials, and gray literature were excluded. We extracted the dataset, input data modality, driving context, inference mode, ML/DL methods, and performance metrics including accuracy, precision, recall, and F1-score. Risk of bias was evaluated using the PROBA-AI tool.

A total of 69 studies met the inclusion criteria. DL models outperformed ML, achieving higher median accuracy (94.48% vs. 91.80%) and significantly better F1-scores (93.15% vs. 84.00%). Median recall was comparable between DL and ML models (93.76% vs. 94.12%), and precision remained similarly high across both approaches (93.18% vs. 94.6%). Most DL methods employed Convolutional Neural Networks (CNN), Recurrent Neural Network (RNN), or hybrid CNN- Long Short-Term Memory (LSTM) architectures, whereas ML studies primarily relied on classical classifiers such as Support Vector Machine (SVM) and Random Forest (RF) supported by handcrafted behavioral features. PROBA-AI assessment revealed considerable methodological heterogeneity, with 27 studies rated as high risk, 25 as low risk, and 17 as moderate.

DL models demonstrate clear performance advantages over ML approaches, particularly in accuracy and F1-score. However, substantial methodological variability-reflected in inconsistent dataset design, annotation practices, and validation strategies-continues to limit comparability across studies.