Abstract
Introduction:
Four semi-stable states explain consistently around 80% of spontaneous electroencephalography (EEG) topographies. These states are referred to as microstates and are suggested to be the building blocks of brain functions (Khanna, et al., 2015). The oscillation of microstates show long range dependencies and are postulated to be scale-free (Van de Ville, et al., 2010). Given that their time course of appearance is correlated with some of the well-known BOLD resting state networks (Britz, et al., 2010), the trajectory of microstate represents an indicator of whole brain dynamics. Deep Neural Networks (DNNs) have recently achieved excellent performance on various learning tasks and present the possibility to learn temporal sequences that have long range dependencies. Here, we show that DNNs can learn long range dependencies in microstate sequences with a high accuracy and that the microstates trajectory is subject invariant.
Methods:
Data acquisition: Data is from 12-min eyes-closed resting-state of 34 healthy male volunteers over one session of simultaneous 3 Tesla fMRI and 64-channel EEG with 5k Hz sampling rate. MRI induced EEG artifacts were removed by the carbon-wire-loop procedure (van der Meer, et al., 2016). A second step of artefact rejection was also done using custom MATLAB scripts based on routines provided by EEGLAB (Delorme and Makeig, 2004) to remove conventional EEG artefacts.
Microstate identification: EEG data was filtered between 1-40 Hz and resampled to 250 Hz. We identified four group level microstate classes using EEGLAB plugin by Thomas Koenig and calculated the time course of these microstates by back projecting group level microstates to the EEG signal of single subjects.
DNNs: We used Long-Short-Term-Memory (LSTM) encode-decoder network with a single hidden layer with each LSTM comprising of 40 units with mean squared loss function and Adam optimizer with a learning rate of 0.001. This method is used for microstate sequence reconstruction. All models are trained using NVIDIA GeForce GTX 1080 and Python based Tensorflow package.
Results:
Experiment 1: Microstates for each subject was divided into sequences of fixed length of 50, 100, and 500 which were used for within-subject sequence reconstruction. For each subject/sequence length, 70% of data was randomly used as training data and remaining was used for test. The mean reconstruction accuracy over all 34 subjects was 98.75±0.39%, 93.43±1.61% and 35.09±6.6% respectively.
Experiment 2: To make sure that the reconstructed sequences in Experiment 1 are not random, we repeated Experiment 1 with 10 randomly generated microstates sequences (sequence length of 100). This time, the accuracy drops to 39.61±2.8% (from originally 93.43%) which is significantly lower than the reconstruction accuracy for real microstates.
Experiment 3: We aimed to do between-subject reconstruction to see if microstate trajectories are subject invariant. We trained a between subject reconstruction model using 5-fold cross validation (sequence length 100). The mean accuracy for this experiment is 99.21% which shows the microstates temporal trajectories are subject invariant and can be generalized across subjects. The higher accuracy in the between-subject classification compared to Experiment 1 is due to more number of training data which helps to achieve better accuracy.
Conclusions:
We show that EEG microstates can be reconstructed and are subject-invariant for long sequence lengths. This suggests that the information encoded in microstates is far beyond the conventional univariate measures (e.g. mean duration). The microstates trajectories can be reconstructed optimally for sequence length of 100 (approximately 10s of time). However, we recognize the need to repeat our experiments with more number of hidden layers to see if the drop in the accuracy for very long sequence length (500 microstates, approximately 50s) is related to the properties of EEG dynamics or the limitations of the current DNN architecture.
