Autodetection of characteristics of sleep EEG with integration of multichannel information by neural networks and fuzzy rules

The automation of electroencephalograms has been studied in order to make their interpretation more quantitative and objective and to reduce the workload of physicians. But in contrast to the techniques used by specialists, autodetection of characteristic waves in EEG in the conventional method that uses single-channel information is difficult for various reasons. In this report, we introduce a method in which multichannel information is integrated to achieve a significantly higher detection rate than in the conventional method. The data detected on each channel by neural networks were integrated by two methods: simple sum integration and integration by fuzzy rules learned using cross-error. Analysis of the results showed that the non-linearity of neural networks facilitates reliable autoselection of channels and the learning of fuzzy rules. Though the mean square error has been commonly used in the learning of fuzzy rules, we used cross-error in this experiment and ensured its effectiveness.