Spatial parcellations, spectral filtering, and connectivity measures in fMRI: Optimizing for discrimination

Abstract The analysis of Functional Connectivity (FC) is a key technique of fMRI, having been used to distinguish brain states and conditions. While many approaches to calculating FC are available, there have been few assessments of their differences, making it difficult to choose approaches, and compare results. Here, we assess the impact of methodological choices on discriminability, using a fully controlled data set of continuous active states involving basic visual and motor tasks, providing robust localized FC changes. We tested a range of anatomical and functional parcellations, including the AAL atlas, parcellations derived from the Human Connectome Project and Independent Component Analysis (ICA) of many dimensionalities. We measure amplitude, covariance, correlation, and regularized partial correlation under different temporal filtering choices. We evaluate features derived from these methods for discriminating states using MVPA. We find that multidimensional parcellations derived from functional data performed similarly, outperforming an anatomical atlas, with correlation and partial correlation (p < .05, FDR). Partial correlation, with appropriate regularization, outperformed correlation. Amplitude and covariance generally discriminated less well, although gave good results with high‐dimensionality ICA. We found that discriminative FC properties are frequency specific; higher frequencies performed surprisingly well under certain configurations of atlas choices and dependency measures, with ICA‐based parcellations revealing greater discriminability at high frequencies compared to other parcellations. Methodological choices in FC analyses can have a profound impact on results and can be selected to optimize accuracy, interpretability, and sharing of results. This work contributes to a basis for consistent selection of approaches to estimating and analyzing FC.


ID
Description Name RSN

Testing different versions derived from HCP atlas.
We compared the classification performance obtained different versions derived from the HCP atlas, a symmetrized version, one using all the features and one in combination with subcortical segmentations.
We report classification results obtained with amplitude, covariance, correlation and partial correlation for the three versions of the HCP atlas.
We observe that for covariance, correlation and partial correlation, the symmetrized HCP version performed better than the full HCP atlas. For Amplitude, both results were close to chance level, but the full HCP performed better (Supplementary Table II). Adding subcortical structures to the atlas did not improve the results for covariance, correlation or partial correlation. However, the performance for amplitude was significantly higher. We only included the symmetrized version in the manuscript, due to its overall good performance and because it represented less computational load compared to the full version. Optimizing regularization levels for each parcellation scheme.

Symmetrized HCP Full HCP HCP + Subcort
We tested a range of regularization levels and observed good performance at many of them for the different atlases or parcellation strategies (Supplementary Figure 4).
We selected the optimal level of regularization for each parcellation using cross-validation.
That is, we selected, across a range of regularization levels, the one being more discriminant in the corresponding training set.
As reported in the main text, the atlases required higher parcellations than the ICA parcellation. Across ICA parcellations, higher dimensionalities needed in general higher regularization. In almost all parcellations the classification results reached the maximum levels at regularization lower than 5. For the HCP atlas, we needed to test a longer range of values, and optimal rhos were found between 7 and 8. In this case, we repeated the full pipeline including cross-validated classification and the results were not significantly higher (classification=70.67%, p=0.125). Similarly, for the Study-ROIs and the AAL atlases classification remains high at higher regularization. However, in these cases, the best value picked by the algorithm was <5.

Supplementary Figure 4. Use of different regularization levels and their implication for classification across parcellations. For each parcellation (A-J), we performed sequential classification using a range of regularization values (from 0 to 5, except for HCP and Study
ROIs where we tested values from 0 to 8).

Spectral characteristics of Study-ROIs
We studied the power spectra of the regions in the Study-ROIs. For each region, we obtained the mean spectra profile at the different tasks. Besides looking at overall power changes located in low frequencies, we observed small differences in frequencies >0.2 Hz (Supplementary Figures 5 and 6). In addition, we evaluated classification with amplitude and covariance with bandpass filtered data. The results showed good performance of both measures at high-frequencies.

Testing correlation and partial correlation with band-pass filtered data
We investigated the classification performance of correlation and regularized partial correlation with filtered data. The two lowest bands showed a similar pattern than the one observed with non-filtered data, with good classification results in almost all dimensionalities and with partial correlation often outperforming full correlation. The boost in performance given by the regularized partial correlation was significant for ICA150

Additional tests:
We include in this section some additional results that are mentioned but not included in the main manuscript We first tested classification with other two widely used methods, namely the K-Nearest Neighbor (K-NN) and Random Forest (RF), both implemented in MATLAB. For that, we used the connectivity matrices obtained from the different parcellation schemes and without any additional filtering.
The classification results are similar to those obtained with SVM, being the K-NN slightly lower and the RF less stable than SVM (Supplementary Table IV).