EEG-Based Lapse Detection With High Temporal Resolution

A warning system capable of reliably detecting lapses in responsiveness (lapses) has the potential to prevent many fatal accidents. We have developed a system capable of detecting lapses in real-time with second-scale temporal resolution. Data was from 15 subjects performing a visuomotor tracking task for two 1-hour sessions with concurrent electroencephalogram (EEG) and facial video recordings. The detector uses a neural network with normalized EEG log-power spectrum inputs from two bipolar EEG derivations, though we also considered a multichannel detector. Lapses, identified using a combination of video rating and tracking behavior, were used to train our detector. We compared detectors employing tapped delay-line linear perceptron, tapped delay-line multilayer perceptron (TDL-MLP), and long short-term memory (LSTM) recurrent neural networks operating continuously at 1 Hz. Using estimates of EEG log-power spectra from up to 4 s prior to a lapse improved detection compared with only using the most recent estimate. We report the first application of a LSTM to an EEG analysis problem. LSTM performance was equivalent to the best TDL-MLP network but did not require an input buffer. Overall performance was satisfactory with area under the curve from receiver operating characteristic analysis of 0.84 plusmn 0.02 (mean plusmn SE) and area under the precision-recall curve of 0.41 plusmn 0.08