Semiautomated extraction of decision relevant features from a raw data based artificial neural network demonstrated by the problem of saccade detection in EOG recordings of smooth pursuit eye movements

Visual identification of saccades in electrooculographic (EOG) recordings of smooth pursuit eye movements (SPEM) is a very time consuming process for the individual experts. Algorithmic approaches to overcome this problem automatically produce high rates of false positive errors. Artificial neural networks (ANN) are excellent tools for pattern recognition problems when signal to noise ratio is low. An automated decision process based on modified raw data inputs showed successful proceeding of a backpropagation ANN with an overall performance of 87% correct classifications with previously unknown data. Investigating the specific influences of prototypical input patterns on a specially designed ANN led to a sparse and efficient data coding, based on a combination of expert knowledge and the internal representation structures of the ANN. Data coding obtained by this semiautomated procedure yielded a list of feature vectors, each representing the relevant information for saccade identification. The feature based ANN produced a reduction of the error rate of nearly 40% and reached an overall correct classification of 92% with unknown data. The proposed method of extracting internal ANN knowledge is not restricted to EOG recordings, and could be used in various fields of signal analysis