Adaptive hybrid system for automatic sleep staging

We present a new adaptive system for automated sleep staging. The proposed system relies on each subject's own data for self-training. Conventional automatic sleep staging algorithms are either rule based, which typically fail to accurately model the complex nature of sleep signals, or numerical methods that use multi-patient training schemes, which suffer from inaccuracies caused by inherent inter-patient variability. The proposed system employs two stages. The first stage is a rule based reasoning engine that can be tuned conservatively to decrease or eliminate false positives, generating just enough samples to train the second stage, which is comprised of a neural network classifier. Results show that this hybrid approach provides an adaptive training scheme that performs more accurately compared to one of the popular commercially available systems.