Comparison of fuzzy reasoning and Autoassociative MLP in sleep spindle detection

Sleep spindles are important short-lasting waveforms in the sleep EEC They are the hallmarks of the so-called Stage 2 sleep. Automated methods for spindle detection presented in literature typically use some form of fixed spindle amplitude threshold. The problem with that approach is that it is poor against inter-subject variability in spindle amplitudes. In this work a spindle detection method without an amplitude threshold was considered. Two versions of the method were compared as fuzzy reasoning and an Autoassociative Multilayer Perceptron (A-MLP) network were both employed for the classification between sleep spindles and non-spindle EEG activities. A novel training procedure was developed to remove inconsistencies from the training data of the A-MLP. This improvement of training data was found to have a positive effect on the method performance on the test data. However, in this comparison the fuzzy reasoning produced a better spindle detection result, probably due to the small size of the A-MLP.