Hyperalignment of motor cortical areas based on motor imagery during action observation

Multivariate Pattern Analysis (MVPA) has grown in importance due to its capacity to use both coarse and fine scale patterns of brain activity. However, a major limitation of multivariate analysis is the difficulty of aligning features across brains, which makes MVPA a subject specific analysis. Recent work by Haxby et al. (2011) introduced a method called Hyperalignment that explored neural activity in ventral temporal cortex during object recognition and demonstrated the ability to align individual patterns of brain activity into a common high dimensional space to facilitate Between Subject Classification (BSC). Here we examined BSC based on Hyperalignment of motor cortex during a task of motor imagery of three natural actions (lift, knock and throw). To achieve this we collected brain activity during the combined tasks of action observation and motor imagery to a parametric action space containing 25 stick-figure blends of the three natural actions. From these responses we derived Hyperalignment transformation parameters that were used to map subjects’ representational spaces of the motor imagery task in the motor cortex into a common model representational space. Results showed that BSC of the neural response patterns based on Hyperalignment exceeded both BSC based on anatomical alignment as well as a standard Within Subject Classification (WSC) approach. We also found that results were sensitive to the order in which participants entered the Hyperalignment algorithm. These results demonstrate the effectiveness of Hyperalignment to align neural responses across subject in motor cortex to enable BSC of motor imagery.


Sensitivity of the common model space to subject order during AOMI and MI
In our main analysis, the common model spaces created in section 2.6.6 were used to individually map the MI data of Session 2 into the common model spaces, and the distribution of BSC was estimated using classifiers trained and tested on the MI dataset of Session 2. To supplement this analysis, here we report the results of the same procedure applied to the AOMI data of Session 2. As shown in Figure S1, for the AOMI dataset classification accuracies are marginally higher than for the MI dataset. Figure S1. A box-whisker plot of BSC based on Hyperalignment of 2000 permutations of subject order showing the classification accuracies of different orders of subjects using AOMI and MI data of session 2.

Effect of subject order on classifier accuracy
To check whether the variation in Hyperalignment performance might be related to simply the first participant(s) used, we examined the participant orders used to obtain the low, middle and best classification results for the MI data. This is shown below in Figure S2, with three panels corresponding to sets of results at low, middle and high performance, where the y axis is classifier performance obtained for the subject order (left panel low classifier accuracy, right panel high classifier accuracy), the x axis is the subject order position and the colour scheme indicates subject number. No strong effect of subject order is evident, and accuracy appears independent of first subject entered into the algorithm. Figure S2. Illustration of the effect of subject order on classifier accuracy. Shown are three panels corresponding to low, middle and best results of classifier accuracy (shown on y axis). The ten colour coded items on each row show the subject order.

Between Subject Classification
Using the AOMI data of session 2, BSC based on anatomical alignment was 37.51% (SE=1.96%, chance=33%). In contrast, BSC based on Hyperalignment was 73.33% (SE=2.79%). The WSC accuracy was still below that of BSC based on Hyperalignment, as shown in Figure S3. In summary, our supplementary analyses confirm that Hyperaligment has a similar effect when applied to the MI and AOMI datasets. The trend for enhanced classification accuracies for the AOMI dataset is unsurprising, given that the Hyperalignment parameters were derived from an AOMI condition (Session 1). Figure S3. The classification accuracies (mean ±SEM) for averaged WSC, BSC based on Anatomical alignment and BSC based on Hyperalignment using AOMI and MI data of Session 2, dashed line indicates chance level (33%).

Univariate analysis of Session 2.
An exploratory univariate analysis was performed on the Session 2 data, in order to examine how brain activity varied between AO+MI and MI conditions as well as between each condition and baseline. For the first level analysis, all the pre-processed functional data of each subject were analysed using a General Linear Model (GLM) with six predictors (two modalities with three actions each), convolved with a hemodynamic response function. Group data were evaluated based on a second level random effect analysis general linear model (RFX-GLM). Three statistical maps were calculated which contrasting the AOMI vs baseline, MI vs baseline and AOMI vs MI respectively. These maps were corrected for multiple comparisons using FDR of 0.05.
The results are shown in Table S1. A single activation peak was found in the contrast MI > baseline. Reassuringly, this was coextensive with the peak for the supplementary motor area in the Searchlight map of BSC based on Hyperalignment shown in main Figure 5. The contrast AOMI > baseline mainly comprised activations in occipital cortex, and a small focus in the supplementary motor area was only sub-threshold in this contrast. These differences are backed-up by the direct contrasts between the two conditions.