ON THE USE OF NEURAL-NETWORK TECHNIQUES TO ANALYZE SLEEP EEG DATA - 3RD-COMMUNICATION - ROBUSTIFICATION OF THE CLASSIFICATOR BY APPLYING ANALGORITHM OBTAINED FROM 9 DIFFERENT NETWORKS
R. Baumgartschmitt et al., ON THE USE OF NEURAL-NETWORK TECHNIQUES TO ANALYZE SLEEP EEG DATA - 3RD-COMMUNICATION - ROBUSTIFICATION OF THE CLASSIFICATOR BY APPLYING ANALGORITHM OBTAINED FROM 9 DIFFERENT NETWORKS, Neuropsychobiology, 37(1), 1998, pp. 49-58
This is the third communication on the use of neural network technique
s to classify sleep stages. In our first communication we presented th
e algorithms and the selection of the feature space and its reduction
by using evolutionary and genetic procedures. In our second communicat
ion we trained the evolutionary optimized networks on the basis of mul
tiple subject data in context with some smoothing algorithms in analog
y of Rechtschaffen and Kales (RK). In this third communication we coul
d demonstrate that the robustness concerning individual specific featu
res of automatically generated sleep profiles could be reasonably impr
oved by an additional modification of the procedure used by SASCIA (Sl
eep 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 me
dians of the values obtained in each output measure were selected for
the allocation to a sleep stage. The fitness criteria of 16 automatica
lly generated sleep profiles showed reasonable concordance with the ex
pert profile. Even though in single cases the concordance between conv
entional 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 clas
sifier is more robust against individuum-specific variability. Despite
the fact that the expert generally employs three channels (EEG, EMG a
nd EGG), at least to build up sleep profiles, the SASCIA system was ab
le to produce profiles on the basis of only one EEG channel with 80% c
oncordance and a correlation coefficient of 0.86. The feature selectio
ns were performed by genetic algorithms and the topologies of the netw
orks were optimized by evolutionary algorithms. This algorithm will no
w be used for larger sample forward classification.