Abstract
The paper presents selected general mathematical tools for pattern recognition and machine learning applied in polysomnography (PSG) for detection of sleep stages and for classification of feature vectors into classes close to those specified by experienced neurologists. Data studied include the set of 85 PGS overnight records of patients with different disorders acquired in the sleep laboratory.
Specific features are then evaluated both in the time-frequency and time-scale domains. Resulting matrices of selected features and associated vectors of sleep stages (wake, NonREM1,2,3 and REM) are used for (i) optimization of the two layer neural network model with sigmoidal and probabilistic transfer functions based on the Bayes theorem and (ii) for evaluation of class boundaries.
The final part of the study includes analysis of classification accuracy and cross-validation of the modes proposed.