Neural network analysis of oxygenation signals in infants during sleep

The use of artificial neural networks (ANNs) to interpret sleep monitoring signals is described. Recordings from ten infants with apparent life threatening episodes were assigned into training feedforward R-PROP networks. In order to separate good signal from artefact, 60 second time frames of SaO2 and TcPO2 signals were processed and the mean and standard deviation values were used as inputs to the networks. Intra-human errors were minimized using this method whilst inter-human errors remained significant. To decrease the latter, the number of hidden units was increased to eight. Sensitivity figures of the SaO2 network were 0.93 and 0.9 for the training and test sets respectively whilst the specificity figures were 0.7 and 0.65 respectively. For the TcPO2 signals the above figures were 0.92, 0.85, 0.77 and 0.61 respectively.