Classification of EEG recordings without perfectly time-locked events

This paper considers the problem of classification of electroencephalography (EEG) recordings without the precise time locking between stimulus presentation times and the recorded EEG waveforms. Traditionally, time locking, or perfect timing, information between stimulus and EEG recordings have been crucial in locating the region of possible neural response. In reality, the stimulus' time information is usually unavailable and the latency of test subjects may not be constant (due to fatigue, concentration, interference, etc.). Therefore, new classification approaches that do not depend on stimulus' time information are needed. To tackle this problem, we firstly characterized the brain response pattern of the target event using the EEG data, in which the timing information is available. Then, based on the pattern, a sliding window was applied to the EEG recordings to detect possible target image response started from each individual location. Finally, the probability of a target image event appeared during an entire EEG recording epoch is estimated by summarizing all the possible locations. The results show that, for classification of EEG epochs of 5s, the approach we proposed can obtain a median area under ROC 0.96, a result that comparable to that with perfect stimulus time information.