An automatic EEG monitoring system in the pediatric ICU

A knowledge-based expert system was developed to automatically assess the level of EEG abnormality of pediatric patients monitored in the ICU. The system receives six hours of 8-channel EEGs and classifies the background EEG as one of seven abnormality levels.A total of 188, six-hour EEGs were visually interpreted by a neurologist and used as training examples. Spectral band activity was computed; artifacts were rejected using a median filter with a hard-limiter. Quantitative variables reflecting amplitude, symmetry, Front/Back differentiation and time variability were then extracted based on the study of Pasupathy (1994). Relationships between quantitative measures and the neurologist's assessment of amplitude, symmetry and Front/Back differentiation were established. A two-layer neural network having the measures of EEG variability as input was created for variability evaluation. A single-layer network was constructed to give the integrative interpretation of EEG abnormality based on the neurologist's assessment of the four features. Suitable knowledge base and inference engine were also constructed.Performance was tested using the rotation method of error estimation. 45% of testing instances were classified the same as the neurologist's interpretation. 46% were classified with an error of one abnormality level. Possible improvement and the clinical future of the system are discussed