Abstract
Ergonomic sleep studies benefit from long-term monitoring in the home environment to cope with daily variations and habituation effects. Polysomnography allows to asses sleep accurately, but is costly, time-consuming and possibly disturbing for the sleeper. Actigraphy is cheap and user friendly, but for many studies lacks accuracy and detailed information. This proof-of-concept study investigates Least-Squares Support Vector Machines as a tool for automatic sleep stage classification (Wake-N1-Rem to N2-N3 separation), using automatic trainingset-specific filtered features as derived from three easy to register signals, namely heart rate, breathing rate and movement. The algorithms are trained and validated using 20 nights out of a 600 night database from over 100 different healthy persons. Different training and test set strategies were analyzed leading to different results. The more person-specific the training nights to the test nights, the better the classification accuracy as validated against the hypnograms scored by experts from the full polysomnograms. In the limit of complete person-specific training, the accuracy of the algorithm on the test set reached 94%. This means that this algorithm could serve its use in long-term monitoring sleep studies in the home environment, especially when prior person-specific polysomnographic training is performed.