A likelihood based computer approach to conventional scoring of sleep

We describe a computer algorithm for scoring sleep stages from polysomnographic records. Our algorithm is based on the likelihood of correct detection of features and the relative impact these features have on decisions necessary to implement R&K rules. We compared results of the computerized scoring to that done manually by 2 experienced scorers. Epoch-by-epoch agreement between computer and humans was 83.4% and between humans was 86.7%. Patterns of misscores were similar. The computer was also able to characterize ambiguity by comparing the probabilities of the two most likely stage scores for each epoch.