hide
Free keywords:
-
Abstract:
Introduction:
The variable resting state fMRI (r-fMRI) signal encompasses various kinds of measurement noise (Caballero-Gaudes & Reynolds, 2017) and physiological changes (Nikolaou F, 2016), as well as the fluctuation of spontaneous neural activity. Although most of the r-fMRI acquisitions were performed by presuming the wakeful brain state of the subjects, previous studies (Tagliazucchi et al., 2012, 2014) have shown that subjects are not able to maintain the wakeful state during the entire scanning period.
Therefore, to acquire the high-quality data, it is suggested to measure the arousal level using simultaneous EEG-fMRI with an eye tracker. However, such experimental setup demands specialized hardware, time and skilled workforce and therefore it is not feasible to most scan facilities and for the all-resting state experiments they perform. The monitoring of the arousal state also plays an important role in a better understanding of the attention and memory tasks.
In this study, we aim to investigate the wakefulness brain state (vigilance level) with the r-fMRI data under machine learning framework.
Methods:
The 28 healthy subjects were scanned using simultaneous EEG-fMRI in resting condition.Using the VIGALL 2.1 algorithm (Hegerl et al., 2008) on resting-state EEG data, five different vigilance scores ranging from relaxed wakefulness (A1, A2, A3) to drowsiness (B1, B2/3) were assigned to each repetition time (TR=2.4s). In the next step, all the fMRI volumes corresponding vigilance score of A1 or B23 were averaged for all the subjects. One subject was discarded due to the absence of the B23 vigilance state. Averaged A1 and B2/3 volumes were compared using paired t-test. The mean signal from each brain region that showed increased activation in high vigilance score was extracted, together with the individual global brain signal. Furthermore, a Support-Vector-Machine classifier was used to predict the wakefulness level based on the region's mean signal. The SVM classifier was trained across subjects using 60% of the data and the prediction accuracy was tested on the remaining 40% as test data. Classification accuracy was compared with that of using the global brain signal instead.
Results:
A GLM contrast between higher and lower vigilance states resulted in higher activations in the thalamus, anterior cingulate cortex (ACC), midbrain, and locus coeruleus in high (A1) vs. low (B2/3) vigilance stage (Fig 1A). The SVM classifier was trained using the 60% data and the prediction accuracy was tested on the remaining 40% test data. The prediction accuracy converged up to 70.2369% for the mean voxel signals and interestingly similar for the global brain signal i.e. 70.2562%.
Conclusions:
The SVM driven classification revealed a good prediction accuracy between the higher vigilance and low vigilance state. Such Classification can improve data quality by estimating the vigilance state directly from the r-fMRI data and resulting in better interpretations of neural signal. However, further work is required to methodically advances the prediction accuracy by employing multi-label classification, higher order deep learning method like transfer learning. Secondly, the relationship between the mean brain area signal and the global brain signal should be further understood.
