Prediction of immediately occurring microsleep events from brain electric signals

Abstract

This contribution addresses the question if imminent changes of the cortical state are predictable. The analysis is based on 1484 examples of microsleep (MS) and 1940 counterexamples of sustained attention (SA), both observed during overnight driving in the simulator. EEG segments (8 s in length) immediately before each respective event were included. Features were extracted by (i) modified periodogram and (ii) Choi-Williams distribution. Machine learning algorithms, namely the optimized learning vector quantization (OLVQ) and the support-vector machine with Gaussian kernel function (SVM), were trained in order to map signal features to the event type (MS or SA). Cross validation analysis yielded test set classification accuracies of 87.5 ± 0.1 % and 82.7 ± 0.1 % for feature set (i) and (ii), respectively. In general, SVM outperformed OLVQ. In conclusion, EEG contains enough information to predict immediately upcoming microsleep events.

Keywords

• driving simulation
• EEG
• microsleep
• neural networks
• sleepiness
• support-vector machines