Fig 1.
The connectivity matrices for each subject were manipulated to obtain a matrix of features for both task and resting data separately and both these matrices were used subsequently to fit different machine learning algorithm and make predictions.
Table 1.
Classification accuracy scores.
Basic statistics of the algorithms’ performance across the 5 × 10 folds for the most optimal hyperparameter configuration in parenthesis.
Fig 2.
Confusion matrix for the best model in task fMRI.
Confusion matrix using exclusively task fMRI data for the selection of the best classification model, which in this case corresponds to a NN with 4 hidden layer with 512, 256, 128 and 64 units.
Fig 3.
Pearson similarity between brain regions.
Representation of correlations amongst the 268 pattern connectivities from the subjects-averaged connectivity matrix. On the left, the full correlation matrix, ordered according to their group label, whose members’ interactions are outlined by a bold rectangle. On the right, the intra-group correlation distribution corresponding to the upper off-diagonal entries in each bold rectangle.
Fig 4.
Confusion matrix using resting data as test.
Confusion matrix for the best classification model, which corresponded to a neural network with four hidden layer of 512, 256, 128 and 64 units, but now using task-based fMRI data as training set and resting-state fMRI data as testing set.
Fig 5.
ROC curves for each class separately.
The areas under these curves can be found in the legend located on the right side. Grey crosses display the model with the specific threshold yielding the results shown before. Those curves above the red dashed line represent exhibit a discriminating power.
Fig 6.
PR curves for each class separately.
The areas under these curves can be found in the legend located on the right side. Grey crosses display the model with the specific threshold yielding the results shown before. Those curves above the red dashed line exhibit discriminating power.
Fig 7.
Classification accuracy of each node.
The best model is used to average each node’s accuracy performance across all subjects contained in the resting fMRI test set in each fold of the cross-validation scheme.