Abstract
In this study we proposed a new noncontact sleep stage estimation method using environmental, forehead and nasal skin temperatures by Support Vector Machine (SVM). The forehead and nasal skin temperature and environmental temperature were measured by a thermography, and the sleep stage was estimated from these temperatures by use of nonlinear SVM with RBF kernel. The polysomnography and thermography were measured simultaneously. Sleep stages were classified into 6 classes:Wake, REM and Stages 1, 2, 3, 4 that are non-REM based on Rechtschaffen and Kales (R & K) scoring rules. When using SVM, measured environmental, forehead and nasal skin temperature data were input to SVM, and the reference data was given as the corresponding correct label. Subjects were five healthy males (age 24.0±7.9 years;mean±SD), and ten data sets were created from the data of all subjects using boot strap random sampling. SVM was trained using the ten data sets, and we evaluated the estimate accuracy as each class. Tuning parameters of the RBF kernel were decided heuristically by grid search, and the generalization capability was tested using cross validation test. Results showed that the classification rate of each sleep classes is as follows;Stage 1:73.2%, Stage 2:96.1%, Stage 3:45.5%, Stage 4:97.1%, REM:89.6%, Wake:63.6% and no error were found in the estimation of sleep state change from each sleep stage to wake stage. The total classification rate was 88.5%, and the 10-fold cross validation test was 81.7%. These results demonstrate that SVM could be classifying the sleep stages by using of the environmental, forehead and nasal skin temperatures.
