Event-related potential noise reduction using the hidden Markov tree model

Event-related potentials (ERP) are brain signals in response to infrequent stimuli applied to a subject. These signals are usually small in amplitude and are embedded in background electroencephalographic (EEG) activity. To analyze them the most common method used is to perform a simple averaging of time aligned ERP segments. However, this method assumes that the ERP signal will not change from segment to segment. In reality, this assumption is not true since the ERP may reflect the activity of cognitive mechanisms in the brain that may change over time. The failure to satisfy this assumption will result in loss of information after the averaging. Here we describe a method to analyze a single segment ERP in noise using the wavelet transform and the hidden Markov tree model. The goal is to reduce the noise content in the signal. We present experimental results using this method on both synthetic and real single-trial visual ERP signals