Decoding traces of memory during offline continuous electrical brain activity (EEG)

Abstract

Continuous electroencephalogram (EEG) provides an excellent possibility to track memory traces from brain rhythmic activity and to study the underlying neural signatures of memory processes. To do so, a promising approach is to employ multivariate pattern classification (MVPC). These methods lend themselves very well to decode the information that resides within the whole distributed spatiotemporal patterns of activity. Using these methods, it is possible to detect traces of memory during sleep or wakefulness, which will reveal valuable insights about the memory function in these brain states. However, there are several methodological problems to decode memory traces from brain activity in paradigm-free (offline) periods. Continuous EEG is prone to elevated levels of noise and distortions and has much higher dimension than single-trial EEG, because of the longer recording time and lack of prior information about relevant time points that are informative for classification. In this case, detecting traces of memory involves searching the whole spatiotemporal feature space to find where memory representations reside. Such high-dimensional data, especially when signal-to-noise ratio and sample size are low, pose problems for classification and interpretation of MVPC result. To address these problems, in this thesis we aim: 1) to develop a proper classification algorithm that enables decoding of continuous EEG to detect memory traces in paradigm-free periods 2) to find EEG correlates of material-specific memory representations during offline periods of sleep and wakefulness, and 3) to provide a systematic method to interpret and validate the specificity of the MVPC results. In chapter 2, we used our MVPC method to detect the ‘when’ and ‘where’ of sleep-dependent reprocessing of memory traces in humans. Although replay of neuronal activity during sleep has been shown in animal experiments, its dynamics and underlying mechanism is still poorly understood in humans. We applied MVPC to human sleep EEG to see if the brain reprocesses previously learned information during sleep and looked for dynamics, neural signatures and relevance of different sleep stages to such process. Here, we developed a two-step classification algorithm that incorporates channel-based feature weighting as well as a tailored preprocessing scheme that is optimized to decode continuous EEG data for between-subject classification. With this method, we demonstrate that the specific content of previous learning episodes is reprocessed during post-learning sleep. We find that memory reprocessing peaks during two distinct periods in the night and both Rapid Eye Movement (REM) and non-REM sleep are involved in this process. To detect traces of short-term memory representations, we employed MVPC in chapter 3 to test whether electrical brain activity during short-term memory maintenance satisfies the necessary conditions for mnemonic representations; i.e. coding for memory content as well as retrieval success. We found that it is possible to decode the content maintained in memory during delay period and if it is subsequently recalled mainly from temporal, parietal, and frontal areas. Importantly, the only overlap between electrodes coding for retrieval success and memory content was found in parietal electrodes, indicating that a dedicated short-term memory representation resides in parietal cortex. Finally, chapter 4 aims at providing a systematic approach to validate the specificity of MVPC result. We investigate the consequences of the high sensitivity of MVPC for stimulus-related differences, which may confound estimation of class differences during decoding of cognitive concepts. We propose a method, which we call concept-response curve, to determine how much decoding performance is specific to the higher-order category processing and to lower-order stimulus processing. We show that this method can be used to quantify the relative contribution of concept- and stimulus-related components and to investigate the spatiotemporal dynamics of conceptual and perceptual processing.