Figure 2. Classification results.

(a) The content of a previous learning experience can be determined from sleep EEG during the second and fourth 90-min segment of the night. At these times, classification accuracy for all sleep stages is significant or approaches significance. The hatched area shows the 95% confidence interval. Classification accuracies for S4 sleep as well as REM sleep in the second sleep segment remain significant after Bonferroni–Holm correction considering all tests (S4: P=0.048, REM: P=0.014). (b) Significance was assessed using permutation tests to ensure that classification rates are higher than can be expected from data sets with random labelling of the data, that is, not containing any information. To estimate the displayed null-distribution from which exact significance levels of classification results can be determined, the MVPC analysis was repeated 1,001 times on the actual data with randomly shuffled condition labels. Dark grey areas show those randomizations during which classification accuracy on randomly labelled data exceeded accuracy obtained on correctly labelled data. (c) If classification accuracies are similar between the training and validation sets, generalizable information could be extracted and the classifier was not overfitted on the training data set. This was the case for all analyses that were significant, that is, for data from the second (circles) and fourth (stars) 90-min segments of the night. Here patterns detected in one set of subjects during classifier training can be generalized to data from a new set of subjects. Data from the first (triangles) and third (squares) 90-min segments show low training accuracy and low accuracy on validation data, indicating that the classifier could not extract information about previous learning content from these periods of the night. *P<0.05, **P<0.01, ***P<0.001.