Lapses in Responsiveness: Characteristics and Detection from the EEG

Performance lapses in occupations where public safety is paramount can have disastrous consequences, resulting in accidents with multiple fatalities. Drowsy individuals performing an active task, like driving, often cycle rapidly between periods of wake and sleep, as exhibited by cyclical variation in both EEG power spectra and task performance measures. The aim of this project was to identify reliable physiological cues indicative of lapses, related to behavioural microsleep episodes, from the EEG, which could in turn be used to develop a real-time lapse detection (or better still, prediction) system. Additionally, the project also sought to achieve an increased understanding of the characteristics of lapses in responsiveness in normal subjects. A study was conducted to determine EEG and/or EOG cues (if any) that expert raters use to detect lapses that occur during a psychomotor vigilance task (PVT), with the subsequent goal of using these cues to design an automated system. A previously-collected dataset comprising physiological and performance data of 10 air traffic controllers (ATCs) was used. Analysis showed that the experts were unable to detect the vast majority of lapses based on EEG and EOG cues. This suggested that, unlike automated sleep staging, an automated lapse detection system needed to identify features not generally visible in the EEG. Limitations in the ATC dataset led to a study where more comprehensive physiological and performance data were collected from normal subjects. Fifteen non-sleep-deprived male volunteers aged 18-36 years were recruited. All performed a 1-D continuous pursuit visuomotor tracking task for 1 hour during each of two sessions that occurred between 1 and 7 weeks apart. A video camera was used to record head and facial expressions of the subject. EEG was recorded from electrodes at 16 scalp locations according to the 10-20 system at 256 Hz. Vertical and horizontal EOG was also recorded. All experimental sessions were held between 12:30 and 17:00 hours. Subjects were asked to refrain from consuming stimulants or depressants, for 4 h prior to each session. Rate and duration were estimated for lapses identified by a tracking flat spot and/or video sleep. Fourteen of the 15 subjects had one or more lapses, with an overall rate of 39.3 ± 12.9 lapses per hour (mean ± SE) and a lapse duration of 3.4 ± 0.5 s. The study also showed that lapsing and tracking error increased during the first 30 or so min of a 1-h session, then decreased during the remaining time, despite the absence of external temporal cues. EEG spectral power was found to be higher during lapses in the delta, theta, and alpha bands, and lower in the beta, gamma, and higher bands, but correlations between changes in EEG power and lapses were low. Thus, complete lapses in responsiveness are a frequent phenomenon in normal subjects - even when not sleep-deprived - undertaking an extended, monotonous, continuous visuomotor task. This is the first study to investigate and report on the characteristics of complete lapses of responsiveness during a continuous tracking task in non-sleep-deprived subjects. The extent to which non-sleep-deprived subjects experience complete lapses in responsiveness during normal working hours was unexpected. Such findings will be of major concern to individuals and companies in various transport sectors. Models based on EEG power spectral features, such as power in the traditional bands and ratios between bands, were developed to detect the change of brain state during behavioural microsleeps. Several other techniques including spectral coherence and asymmetry, fractal dimension, approximate entropy, and Lempel-Ziv (LZ) complexity were also used to form detection models. Following the removal of eye blink artifacts from the EEG, the signal was transformed into z-scores relative to the baseline of the signal. An epoch length of 2 s and an overlap of 1 s (50%) between successive epochs were used for all signal processing algorithms. Principal component analysis was used to reduce redundancy in the features extracted from the 16 EEG derivations. Linear discriminant analysis was used to form individual classification models capable of detecting lapses using data from each subject. The overall detection model was formed by combining the outputs of the individual models using stacked generalization with constrained least-squares fitting used to determine the optimal meta-learner weights of the stacked system. The performance of the lapse detector was measured both in terms of its ability to detect lapse state (in 1-s epochs) and lapse events. Best performance in lapse state detection was achieved using the detector based on spectral power (SP) features (mean correlation of φ = 0.39 ± 0.06). Lapse event detection performance using SP features was moderate at best (sensitivity = 73.5%, selectivity = 25.5%). LZ complexity feature-based detector showed the highest performance (φ = 0.28 ± 0.06) out of the 3 non-linear feature-based detectors. The SP+LZ feature-based model had no improvement in performance over the detector based on SP alone, suggesting that LZ features contributed no additional information. Alpha power contributed the most to the overall SP-based detection model. Analysis showed that the lapse detection model was detecting phasic, rather than tonic, changes in the level of drowsiness. The performance of these EEG-based lapse detection systems is modest. Further research is needed to develop more sensitive methods to extract cues from the EEG leading to devices capable of detecting and/or predicting lapses