Objective. The objective of this paper is to present a driver sleepiness detection model based on electrophysiological data and a neural network consisting of convolutional neural networks and a long short-term memory architecture.

Approach. The model was developed and evaluated on data from 12 different experiments with 269 drivers and 1187 driving sessions during daytime (low sleepiness condition) and night-time (high sleepiness condition), collected during naturalistic driving conditions on real roads in Sweden or in an advanced moving-base driving simulator. Electrooculographic and electroencephalographic time series data, split up in 16 634 2.5 min data segments was used as input to the deep neural network. This probably constitutes the largest labeled driver sleepiness dataset in the world. The model outputs a binary decision as alert (defined as ≤6 on the Karolinska Sleepiness Scale, KSS) or sleepy (KSS ≥ 8) or a regression output corresponding to KSS [1-5, 6, 7, 8, 9].

Main results. The subject-independent mean absolute error (MAE) was 0.78. Binary classification accuracy for the regression model was 82.6% as compared to 82.0% for a model that was trained specifically for the binary classification task. Data from the eyes were more informative than data from the brain. A combined input improved performance for some models, but the gain was very limited.

Significance. Improved classification results were achieved with the regression model compared to the classification model. This suggests that the implicit order of the KSS ratings, i.e. the progression from alert to sleepy, provides important information for robust modelling of driver sleepiness, and that class labels should not simply be aggregated into an alert and a sleepy class. Furthermore, the model consistently showed better results than a model trained on manually extracted features based on expert knowledge, indicating that the model can detect sleepiness that is not covered by traditional algorithms.