Classification of Sleep Apnea Events by Means of Radial Basis Function Networks.
Thomas Zemen, Markus Clabian, Helmut Pfützner · 1998
Sleep apneas, cessations of breathing during sleep, which lead to shortened life expectation (adults) and indicate SIDS-risks (premature born babies) must be diagnosed reliably. Therefore a monitoring system is established and attempts are made for automatic scoring to reduce clinical efforts. This paper describes the detection and classification of sleep apnea events by means of radial basis function networks (RBFN). A monitoring device, based on electric field plethysmography, which detects respiratory and cardiac activity is presented. In addition, heart rate and blood oxygen saturation are recorded. Pre-processing is performed by several feature extraction algorithms including time dependent Fourier transform, cepstrum transform, linear predictive coding, linear predictive coding cepstrum (LPCC) and wavelet transform. RBFNs are trained with learning vector quantization (LVQ) and self organizing map (SOM). To achieve high classification rates, optimization of number and spread of neurons are performed. Best results with classification rates of 64% \\Sigma 3:4% (adults) and 62:6% \\Sigma 6:8% (babies) is obtained using LPCC and data reduction through principle components analysis (PCA). The net consisted of 20 hidden neurons and is trained with LVQ. All results are derived using 10-fold cross validation.