Attributed strings for recognition of epileptic transients in EEG

This paper presents a structural pattern recognition system for signals interpreted and described by a human expert. For each pattern to be recognized, the signal is represented by a sequence of local information ('types of primitives') which is syntactically correct if some 'structural attributes' are satisfied. Types of primitives and structural attributes are defined. This representation method takes a priori knowledge of the expert's descriptions into account. Our system is concerned with EEG's analysis. Detection of transient events, spikes and spike-and-wave complexes helps the electroencephalographist by rejecting parts of the EEG trace which certainly do not contain these patterns. The transient events detection keeps a 16% average of the initial trace for pathological EEG's and 8.3% for non-pathological EEG's. Only 3.3% of pathological EEG's are not rejected after spikes and spike-and-wave complexes detection. The results of this parse are compared with human interpretations.