Automatic Feature Selection for Sleep/Wake Classification with Small Data Sets

This paper describes an automatic feature selection algorithm integrated into a classification framework developed to discriminate between sleep and wake states during the night. The feature selection algorithm proposed in this paper uses the Mahalanobis distance and the Spearman’s ranked-order correlation as selection criteria to restrict search in a large feature space. The algorithm was tested using a leave-one-subject-out cross-validation procedure on 15 single-night PSG recordings of healthy sleepers and then compared to the results of a standard Sequential Forward Search (SFS) algorithm. It achieved comparable performance in terms of Cohen’s kappa (κ = 0.62) and the Area under the Precision-Recall curve (AUCPR = 0.59), but gave a significant computational time improvement by a factor of nearly 10. The feature selection procedure, applied on each iteration of the cross-validation, was found to be stable, consistently selecting a similar list of features. It selected an average of 10.33 features per iteration, nearly half of the 21 features selected by SFS. In addition, learning curves show that the training and testing performances converge faster than for SFS and that the final training-testing performance difference is smaller, suggesting that the new algorithm is more adequate for data sets with a small number of subjects.