Stacking pre-trained classifiers across subjects improves fMRI task decoding accuracy

1432 

Abstract Submission 

Himanshu Aggarwal¹, Bertrand Thirion¹

¹INRIA Saclay, France, Palaiseau, Ile-de-France

Himanshu Aggarwal  
INRIA Saclay, France
Palaiseau, Ile-de-France

Bertrand Thirion  
INRIA Saclay, France
Palaiseau, Ile-de-France

Decoding refers to the process of inferring the stimuli from an individual's brain signals. In the context of task fMRI, high-dimensional voxel-wise BOLD activation patterns are used as features to decode the stimuli presented to the subject during the task. Imaging sessions are constrained by the time and resources available to researchers, resulting in a limited number of samples per subject. This leads to a suboptimal feature-to-instance ratio, causing low decoding accuracy. Many such small fMRI datasets are publicly available but hard to use due to their small sample sizes and domain-specific nature. Therefore, there is a need for methods that can leverage these data and can be used to efficiently transfer learning across subjects, tasks, and cohorts (Gu et al. 2022). In this study, we test a stacking approach for transfer learning across subjects within four small cohorts with a few trials of a given task.

In a conventional decoding setting, a classifier is trained to learn the mapping between stimuli labels and features in the voxel space. Here, we present a stacking approach where we train a classifier to learn the mapping between true stimuli labels and the stimuli predictions from classifiers pre-trained on other subjects' voxel-space features (Figure 1). This converts a high-dimensional problem (where the number of features is the number of voxels) to a low-dimensional one (where the number of features is one minus the number of subjects in the cohort).

We compare the stacking approach against the conventional one in four different cohorts and tasks: BOLD5000 (Chang et al. 2019), Forrest (Hanke et al. 2015), Neuromod (Bellec and Boyle 2019), and RSVP (Pinho et al. 2020). We use trial-by-trial GLM parameter maps from each of these tasks. Each of these datasets has about 4-6 classes of stimuli to be decoded and hence has similar decoding complexity. We also compare the two decoding settings across two different feature spaces - full voxel space and 1024 modes of the DiFuMo atlas (Dadi et al. 2020) and use two different classifiers for decoding - linear SVC and random forest (Pedregosa et al. 2011).

Within each cohort, we vary the size of the training set in increments of 10% of the samples available for each subject and always test the trained model on 10% of the samples.

As can be seen in Figure 2 (a), the stacking approach achieves the highest average accuracy within three (Neuromod, Forrest, and RSVP) of the four cohorts.

For the Neuromod dataset, the best-performing scenario at an average accuracy of 80% is that of the stacking approach using the full voxel space and linear SVC classifier (Figure 2 (a)). For this dataset, we also see a maximum average gain (accuracy of stacking - accuracy of conventional setting) ranging between 15-22% (Figure 2 (b) and (c)).

In contrast, for BOLD5000 (having only 3 subjects), the stacking approach performs worse than the conventional one across all classification scenarios.

The stacking approach presented in this study improves decoding performance, specifically in small training samples. This demonstrates a direct application for cohorts with small samples (like Neuromod). Since the final classifier in this method encounters a low dimensional feature space (one minus the number of subjects), the performance can be further improved by using a random forest classifier instead of linear SVC. This is especially true for relatively larger cohorts (like Forrest and RSVP). While the stacking approach works as well on reduced data (as with the DiFuMo approach), the benefits are less clear in the latter case, given that overfitting risks are lower. Finally, if the number of subjects in the cohort is too low, the method performs worse than conventional, as observed in BOLD5000. Overall, this suggests that off-the-shelf classifiers can chart the space of cognitive domains, which calls for a generalization of this approach across datasets.

Classification and Predictive Modeling ¹

Methods Development

Other Methods

Brain Machine Interface ²

Computational Neuroscience

Data analysis

FUNCTIONAL MRI

Machine Learning

Modeling

Multivariate

Statistical Methods

Other - Decoding, Transfer learning