Event-based detection of lapses of responsiveness

EEG spectral power has been shown to correlate with level of arousal and alertness in humans. In this paper, we assess its usefulness in the detection of lapses of responsiveness ('lapses') on an event, rather than state, basis. Eight non-sleep-deprived normal subjects performed two 1-hour sessions of a continuous tracking task while EEG and facial video were recorded. Lapses were identified by the presence of tracking flat spots or clear instances of behavioural microsleeps as identified by a human rater observing video recordings of the subject. Spectral power in the standard EEG bands was calculated using the Burg method on 16 bipolar derivations to form an EEG feature matrix. Linear discriminant analysis was used to form a classifier for each subject. The 8 classifiers were combined using stacked generalization with constrained-least squares fitting to create an overall detection model. Estimation of lapse-event detection performance required the development of a novel procedure to account for the variable duration of lapses. Event detection for the concatenated data from all 8 subjects yielded an overall sensitivity of 73.5%, selectivity of 25.5%, and accuracy of 61.2%. While the performance of this detector is modest, and not yet sufficient for real-time detection, the detection of lapses at such a high temporal resolution (1 s) is encouraging relative to previous studies which have generally tended to estimate changes in alertness on a minute-scale basis.