Identification of memory reactivation during sleep by EEG classification

Abstract

Memory reactivation during sleep is critical for consolidation, but also extremely difficult to measure as it is subtle, distributed and temporally unpredictable. This article reports a novel method for detecting such reactivation in standard sleep recordings. During learning, participants produced a complex sequence of finger presses, with each finger cued by a distinct audio-visual stimulus. Auditory cues were then re-played during subsequent sleep to trigger neural reactivation through a method known as targeted memory reactivation (TMR). Next, we used electroencephalography data from the learning session to train a machine learning classifier, and then applied this classifier to sleep data to determine how successfully each tone had elicited memory reactivation. Neural reactivation was classified above chance in all participants when TMR was applied in SWS, and in 5 of the 14 participants to whom TMR was applied in N2. Classification success reduced across numerous repetitions of the tone cue, suggesting either a gradually reducing responsiveness to such cues or a plasticity-related change in the neural signature as a result of cueing. We believe this method will be valuable for future investigations of memory consolidation.

Keywords: Consolidation; Machine learning; Memory reactivation; Pattern recognition; Sleep.

Fig. 1

A schematic illustration of the design of the experiment and the classifier. (A) In the Motor task, participants performed the SRTT task with finger presses. In the Imagery task they were instructed to remain motionless and imagine performing the task while experiencing the same audio-visual cues that were used in the Motor task. During subsequent SWS and N2, the sequence was repeatedly reactivated in blocks of 1.5 min on, 2 min off. (B) The visual cues used in the experiment. Visual cues were objects or faces (1 = face #1, 2 = lamp, 3 = face #2, 4 = water tap). Note that these 4 cues were always the same (e.g. each finger was paired with just one image, and that image was repeated every time the cue was repeated). (C) Mean learning curve showing performance (CS = RT/accuracy) for each block before and after sleep. Error bars indicate one standard deviation.

Fig. 2

Flow diagram of the classifier pipeline. We trained the classifier with EEG data from the wakeful imagery task (bluish colours), next we used EEG data from sleep (orange colours) to feed the trained algorithm and calculate the final accuracy results (purple colours). From the imagery data we extracted 3 types of features (temporal, spectral and wavelet-based features) that divided into training and testing sets were used to train the classifier after a selection process to reduce the number of features. The ranking and selection of features was done using join mutual information (JMI) algorithm and a wrapping methodology. Once the classifier (LDC) was trained we extracted the same type of features from the sleep dataset and used them to feed the trained classifier. An additional control step (permutation of labels) was added to be sure that the classification rates were not due merely to the chance probability.

Fig. 3

Behavioural results. (A) Correct classification rate (CCR) in the Motor and Imagery experiments shown as mean and standard error (SE). (B) Correct classification rate for SWS, N2 and Control and their corresponding random classifiers, shown as mean and SE.

Fig. 4

Frequency of selecting each family of features. After the feature extraction stage, a feature selection process determines which features were most suitable for classification. The X-axis (# Times Selected) represents the number of times each feature family appeared across participants. Y-axis (% participants) shows the proportion of participants in whom that particular number of features was selected.

Fig. 5

Electrode selection. (A) A plot of the frequency of selecting each of the 16 electrodes for the Imagery classifier. This was determined by accounting for each time a feature belonging to a particular electrode was selected by the classifier. The more often an electrode was selected (# Times Selected) across a large proportion of the participants (Proportion of Participants is indicated by the colour bar), the more important the electrode was deemed. This was objectively determined using hierarchical clustering (B).