Because of the spread of electromagnetic signals through CSF and skull through volume conduction, EEG data recorded at different points on the scalp tend to be correlated. Bell and Sejnowski (1995) have recently presented an artificial neural network algorithm that identifies and separates statistically independent signals from a number of channels composed of linear mixtures of an equal number of sources. Here we present a first application of this Independent Component Analysis (ICA) algorithm to human EEG data. Conceptually, ICA filtering separates the problem of source identification in EEG data from the related problem of physical source localization. Three subjects performed a continuous auditory detection task in two half hour sessions. ICA filters trained on 14-channel EEG data collected during these sessions identified 14 statistically independent source channels which could then be further processed using event-related potential (ERP), event-related spectral perturbation (ERSP), and other signal processing techniques. One ICA source channel contained most eye movement activity, and another two collected line noise and muscle activity, while others were free of these artifacts. Changes in spectral power in several ICA channels covaried with changes in performance. If ICA sources can be shown to have distinct and consistent relationships to behavior or other physiological signals, ICA filtering may reveal meaningful aspects of event-related brain dynamics associated with sensory and cognitive processing but hidden within correlated EEG responses at individual scalp sites.

Thirteen subjects participated in an auditory simulation of a passive sonar target detection environment. Targets were 300 ms noise bursts presented at near threshold levels in a noise background at a mean rate of 10 per minute. Task-irrelevant probe tones were also presented at inter-stimulus intervals of 2-4 seconds. Each subject participated in two 28 minute test sessions, pressing a button whenever they detected a noise target. Prominent minute-scale fluctuations in performance (computed as changes in local error rate using a 32-s moving window) occurred in many of the sessions. Evoked responses to the irrelevant probe tones in thirteen runs with highest number of performance lapses were sorted by current local error rate and smoothed using a moving-average. The amplitude of the grand mean N2 response to the irrelevant probe tones increased monotonically with error rate. Averaged evoked responses to relatively frequent, task-irrelevant probe tones appear to allow an accurate estimate of level of alertness if adequate numbers of trials are available.

A total of ten male subjects (ages 18-34) participated in a simulation of a passive sonar auditory target detection task. Throughout each session, intermittent target and probe auditory stimuli were presented binaurally in the presence of a continuous white noise background at 62 dB nHL. Target noise bursts were presented at 6 dB SL and were 300 ms in duration with rise and fall times > 110 ms. Each subject participated in two simulated work sessions of 28 minutes, seated in a comfortable chair with eyes closed and pressing a button each time they detected a target noise burst. EEG and EOG signals were amplified 50K times with a 0.1-100 Hz bandwidth through Grass EEG amplifiers and converted continuously to 12-bit digital format at a sampling rate of 312.5 Hz per channel. EEG was collected from 13 central scalp locations of the International 10-20 system referred to the right mastoid. Periocular electrodes were used to record and reject from analysis electrical potentials generated by eye movements.

A continuous estimate of performance at regularly spaced time intervals, local error rate, was derived by computing the fraction of undetected targets within a time window with a constant width of 32 s which was advanced through the data in 1.64 s steps. Targets were considered detected when target reaction times were within 200-2000 ms of stimulus onset. Because of technical problems, eighteen of the twenty sessions were used in the analysis. Brain evoked responses to detected targets (Hits) were ordered by increasing local error rate. A moving average of the reordered epochs at site Pz (center panel) reveals that as error rate increases above 50%, mean P300 amplitude decreases, while peak latency remains nearly constant. Examination of individual sessions showed that between-subject variability in P300 latency was not responsible for the decrease in amplitude of P300 to Hits. P300 latency remained near constant in nearly every run, even though mean reaction time in these experiments increased linearly by 256 ms from zero-error to highest-error rate epochs. Except for Lapses occurring when local error rate was very low, P300 is absent from "Lapse" target responses (right panel) and a triphasic response pattern appears. This response may be an equivalent of the "sleep P2-N2" that occurred in response to those task-irrelevant probes preceding target Lapses in this experiment. (Makeig et al, submitted). In the moving-averaged response to all targets (left panel), as local error rate increases, P300 amplitude declines and the triphasic Lapse response pattern emerges.We conclude that in task environment in which target time and response information is available, P300 amplitude may be used to predict current local error rate and hence, vigilance level. In situations in which such information is not available, other electrophysiological measures, including task-irrelevant probe responses and the EEG spectrum itself may be used to predict vigilance.

Thirteen subjects detected noise burst targets presented in in a white noise background at a mean rate of 10 per minute. Within each session, local error rate, defined as the fraction of targets detected in a 33-second moving window, fluctuated widely. Mean coherence between slow mean variations in EEG power and in local error rate was computed for each EEG frequency and performance cycle length, and was shown by a Monte Carlo procedure to be significant for many EEG frequencies and performance cycle lengths, particularly in four well-defined EEG frequency bands, near 3, 10, 13, and 19 Hz, and at higher frequencies, in two cycle length ranges, one longer than 4 minutes and the other near 90 s per cycle. The coherence phase plane contained a prominent phase reversal near 6 Hz. Sorting individual spectra by local error rate confirmed the close relation between performance and EEG power and its relative within-subject stability. These results show that attempts to maintain alertness in an auditory detection task result in concurrent minute and multi-minute scale fluctuations in performance and the EEG power spectrum.

A first laboratory version of an Alertness Monitoring/Management (AMM) system has been designed and implemented. The system continually estimates the level of alertness of a human subject using EEG spectral information recorded from the subject's scalp, and delivers auditory feedback to assist the subject in managing his or her own level of alertness in work environments requiring constant vigilance. The system allows experimenters to monitor its input and output via real-time color graphics displays.

As a first demonstration and evaluation of the system, six subjects participated in five half-hour sessions (three training and two feedback sessions), which involved dual detection tasks simulating the passive sonar environment. Auditory targets, 300-ms noisebursts presented at 6 dB above a noise background, were presented at a mean rate of 10 targets per minute. A continuous visual waterfall display presented illuminated vertical line targets at a mean rate of one per minute. Subjects pressed one response button to report noisebursts and another to report visual targets.

Neural net estimation algorithms were trained for each subject to estimate the current probability of detecting auditory targets using electroencephalogram (EEG) and performance data collected during one or more of the initial training sessions. During feedback sessions, real-time signal processing and individualized neural network analysis of EEG recorded from a central scalp electrode were used to estimate continuously, in near real-time, the current probability-of-detection of auditory targets. Whenever this probability-of-detection measure declined below a preset threshold (e.g., when it predicted more than a 40% chance of failure to detect the auditory targets), the system sounded an alarm in the subject's headphones. When training sessions comprising a relatively wide range of detection rates were used to train the estimation algorithms, the alertness estimates followed changes in observed detection probability relatively accurately.

Four of the six subjects reported that the alertness feedback helped them to maintain detection performance. A fifth subject did not produce enough detection lapses to fairly evaluate the system. Review of data from the sixth subject suggested that future versions of the system may be able to provide useful feedback to this subject as well. Review of results of the demonstration experiment have suggested several improvements to signal processing and training procedures used in the system. Effects of these enhancements on system performance are being evaluated.

A new measure of event-related brain dynamics, the event-related spectral perturbation (ERSP), is introduced to study event-related dynamics of the EEG spectrum induced by, but not phase-locked to, the onset of the auditory stimuli. The ERSP reveals aspects of event-related brain dynamics not contained in the ERP average of the same response epochs. Twenty-eight subjects participated in daily auditory evoked response experiments during a four day study of the effects of 24-hour free-field exposure to intermittent trains of 89 dB low frequency tones. During evoked response testing, the same tones were presented through headphones in random order at 5 s intervals. No significant changes in behavioral thresholds occurred during or after free-field exposure. ERSPs induced by target pips presented in some inter-tone intervals were larger than, but shared common features with ERSPs induced by the tones, most prominently a ridge of augmented EEG amplitude from 11 to 19 Hz, peaking 1-1.5 s after stimulus onset. Following 3-11 hours of free-field exposure, this feature was significantly smaller in tone-induced ERSPs; target-induced ERSPs were not similarly affected. These results therefore document systematic effects of exposure to intermittent tones on EEG brain dynamics even in the absence of changes in auditory thresholds.

Changes in EEG power collected at two sites from 10 subjects accompany slow and irregular fluctuations in alertness level. By merging power spectrum estimation, principal component analysis, and artificial neural networks, it is very feasible to accurately estimate shifts in an operator's level of alertness.

Real-time monitoring of alertness is highly desirable in a variety of operational environments where operators must sustain readiness during periods of low active decision-making. This paper demonstrates the feasibility of using physiological information, mainly the electroencephalogram (EEG), to accurately estimate, in real time, significant shifts in an operator's global level of alertness in a visual/auditory target detection task.

During drowsiness, human performance in responding to above-threshold auditory targets tends to vary irregularly over periods of 4 minutes and longer. These performance fluctuations are accompanied by distinct changes in the frequency spectrum of the electroencephalogram (EEG) on three time scales:

(1) During minute-scale and longer periods of intermittent responding, mean activity levels in the (< 4 Hz) delta and (4-6 Hz) theta bands, and at the sleep spindle frequency (14 Hz) are higher than during alert performance.

(2) In most subjects, 4-6 Hz theta EEG activity begins to increase, and gamma band activity above 35 Hz begins to decrease, about 10 s before presentations of undetected targets, while before detected targets, 4-6 Hz amplitude decreases and gamma band amplitude increases. Both these amplitude differences last 15-20 s and occur in parallel with event-related cycles in target detection probability. In the same periods, alpha and sleep-spindle frequency amplitudes also show prominent 15-20 s cycles, but these are not phase locked to performance cycles.

(3) A second or longer after undetected targets, amplitude at intermediate (10-25 Hz) frequencies decreases briefly, while detected targets are followed by a transient amplitude increase in the same latency and frequency range.