gms | German Medical Science

Background: Functional magnetic resonance imaging (fMRI) is based on different magnetic blood properties before and after oxygen consumption caused by neuronal activity measured as spatial resolved “blood oxygen level dependent (BOLD)” signal. Thus, fMRI allows detecting areas of increased brain activation due to external stimulation or performing a mental task.

After learning the above causality between stimulus and brain area this knowledge can be used for the decoding to conclude (with short computation time) from the currently measured BOLD signals towards the on-going high-level brain function. This is called real-time fMRI decoding.

A well-established method (AH) uses the masks of previously fMRI-acquired significant voxels. For the real-time fMRI decoding the spatial means of the BOLD signal within the masks are calculated. We adapted and compared several multivariate methods to find the best solution for improving the decoding accuracy and minimizing the training time.

Methods: The large number of voxels (>150,000 per measurement) and the small sample size (100-500 measurements) prevent applying naÏve classical regression methods. Modern regularized regression methods need time consuming training. Furthermore, they need an inverted model which hampers the inclusion of a proper error model for the low frequency signal drift of the BOLD signal.

Here, both classical and modern methods were adapted to the problem. Multiple multivariate regression with restriction to diagonal covariance matrix (C2), principal component regression (PCR), elastic-net regularized regression (EN), and support vector machine regression (SVMR) have been combined each with a pre-whitening-method and quantile-detrending to deal with the signal drift at the training and during the real-time decoding. For pre-whitening and for adjusting the quantile-detrending a restricted maximum likelihood method were implemented to estimate the variance components spatial resolved. Further adaptions of the methods where implemented to use them also for paradigms with variable stimulus amplitudes.

These multivariate methods and the well-established method where applied to simulated real-time experiments using both simulated data and measured data from a motor activation experiment (finger tapping left and right) and a stress paradigm (video game with increasing difficulty). Motor and stress tasks with different stress levels were measured twice for 5 and 14 subjects to get training and test data.

Results: All approaches led to feasible methods allowing training and decoding of high-dimensional data with short training data series as well as dealing with the slow signal drift (both in the training and real-time decoding).

C2 works well without including spatial error correlation. EN showed good results with simulated data, but was less suited to measured data possibly as cross validation for the determination of the hyperparameters did not work well with the block-wise paradigm. PCR worked best with measured data and was 10 times more accurate than AH.

Training of EN and SVMR needed a factor of 150 and 50 of the time required for training in approach PCR.

Conclusion: Simulated and measured data led to the best results applying the PCR/pre-whitening approach to real-time fMRI decoding. The improved accuracy and relatively short training will be advantageous in future real-time fMRI-decoding experiments.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.