Optimal EEG Segmentation for Microsleep Detection Based on Machine Learning

Abstract

Paroxysmal brain state changes, such as microsleep events in drivers, are presumably subcortically induced and accompanied by cortical processes. This raises questions of how stable and persistent are the cortical processes that can be observed with EEG in real time. For this purpose, recordings of four night-driving simulation studies including 79 subjects are used to analyze how large the time window and how the temporal offset of the EEG segment must be to achieve maximum classification rate. From each EEG segment, power spectral densities were estimated using modified periodogram method and averaged in narrow spectral bands. They were then processed using gradient boosting machines in order to map them to one of two brain state types: microsleep or sustained attention. Segment length and offset were found to have moderate and dramatic effects on recognition rate, respectively.

Keywords

• electroencephalography
• microsleep events
• driving simulations
• machine learning
• gradient boosting