This is the third communication on the use of neural network techniques to classify sleep stages. In our first communication we presented the algorithms and the selection of the feature space and its reduction by using evolutionary and genetic procedures. In our second communication we trained the evolutionary optimized networks on the basis of multiple subject data in context with some smoothing algorithms in analogy of Rechtschaffen and Kales (RK). In this third communication we could demonstrate that the robustness concerning individual specific features of automatically generated sleep profiles could be reasonably improved by an additional modification of the procedure used by SASCIA (Sleep Analysis System to Challenge Innovative Artificial Networks). The outputs of nine different networks that were created by the data of 9 different subjects were used simultaneously for classification. The medians of the values obtained in each output measure were selected for the allocation to a sleep stage. The fitness criteria of 16 automatically generated sleep profiles showed reasonable concordance with the expert profile. Even though in single cases the concordance between conventional RK classifications and automatically generated profiles were a few percentages lower, the average correct classification of the 12 classified subjects improved substantially, thus proving that the classifier is more robust against individuum-specific variability. Despite the fact that the expert generally employs three channels (EEG, EMG and EOG), at least to build up sleep profiles, the SASCIA system was able to produce profiles on the basis of only one EEG channel with 80% concordance and a correlation coefficient of 0.86. The feature selections were performed by genetic algorithms and the topologies of the networks were optimized by evolutionary algorithms. This algorithm will now be used for larger sample forward classification.

To automate sleep stage scoring, the system sleep analysis system to challenge innovative artificial networks (SASCIA) has been developed and implemented. The aims of our investigation were twofold: In addition to automatic sleep stage scoring the hypothesis was tested that the information of only 1 EEG channel (C4-A2) should be sufficient to automatically generate sleep profiles which are comparable with profiles made by sleep experts on the basis of at least 3-channel EEG (C4-A2), EOG and EMG, as EOG and EMG are seen as epiphenomena during sleep and the full information about the sleep stage should--according to our hypothesis--be available in the EEG. The main components of the SASCIA sleep analysis system are designed to meet the requirements of flexible adaptation to the interindividual differences of the sleep EEG. The core of the SASCIA sleep analysis system consists of neural networks. Supervised learning was implemented and the experts' scorings were included into the learning set and test set. The feature selections out of a large number (118) are performed by genetic algorithms and the topologies of the networks are optimized by evolutionary algorithms. Different mathematical procedures were used to evaluate and optimize the efficiency of the system. The profiles generated by SASCIA are in reasonable agreement with the sleep stages scored by experts according to RKR. The development of the system is communicated in three parts: the first communication deals with the application of the neural network techniques using evolutionary and genetic algorithms and with the selection of feature space. The second communication shows the training of these evolutionary optimized network techniques with multiple subjects and the application of context rules, while the third communication shows an improvement in the robustness by the simultaneous application of 9 different networks obtained from 9 subject types which were used in combination with context rules.