Neuroplasticity within and between Functional Brain Networks in Mental Training Based on Long-Term Meditation

Figures

• 

  Figure 1. Experimental paradigm and analysis schema. (A) Experimental procedure consisted of three blocks of the following sequence: 6 min FA and 6 min OM meditation blocks intermixed with 3 min non-meditative resting state block. (B) Analysis schema consisted in extracting the average time course of the preprocessed BOLD signal (see Section 2) from 90 ROIs and computing the pairwise Pearson correlation matrix between extracted time courses; then, the upper triangular part of the correlation matrix was used as a feature set for training a Support Vector Regression (SVR) model, to predict either the meditation expertise or the age of the participants. Before training, a correlation based feature selection was performed to reduce the feature set. We split the dataset into two parts: 75% of the subjects were used for training and 25% for training repeating this procedure 50 times, with permutation of the set of subjects for cross-validation. Finally, performance was evaluated using mean squared error and correlation, while feature evaluation was performed by extracting the selection frequency of the features and by inspecting the weights of the SVR model.

• 

  Figure 2. Feature selection frequency in FA and OM meditation styles. The figure shows the feature selection frequency for (a) focused attention and (b) open monitoring meditation, for age prediction (x-axis) and expertise prediction (y-axis). Each dot represents a connection (feature), and the position coordinates x and y represent the number of times this connection was selected in the regression, in all cross-validation folds. The maximum frequency value is equal to the number of cross-validation folds. The color of the dots indicates the number of features with that particular selection frequency. In these plots, dots representing features with similar frequencies in the two regression tasks lie close to the diagonal, while dots representing features with different frequencies lie far from the diagonal.

• 

  Figure 3. Feature selection frequency for expertise and age prediction. The figure shows feature selection frequency for (a) meditation expertise prediction and (b) age prediction for focused attention (x-axis) and open monitoring (y-axis) meditation styles. Each dot represents a connection (feature), and the position coordinates x and y represent the number of times this connection was selected in the regression, in all cross-validation folds. The maximum frequency value is equal to the number of cross-validation folds. The color of the dots indicates the number of features with that particular selection frequency. In these plots, dots representing features with similar frequencies in the two meditation styles lie close to the diagonal, while dots representing features with different frequencies lie far from the diagonal.

• 

  Figure 4. Within− and between−network contributions for different prediction tasks. (a) The average of connectivity weights within and between networks for expertise prediction in FA meditation style and (b) OM style and (c) for age prediction in both styles. These weights were obtained by extracting the weights learned by the model, averaging within and between networks, then normalizing them using the z-score function. Straight lines represent within-network contributions, while other lines represent between-network contributions.

• 

  Figure 5. Normalized absolute weights of the nodes for different prediction tasks. The figure shows normalized absolute weights for (a) expertise prediction in FA meditation style and (b) OM style and (c) for age prediction in FA style and (d) OM style. These weights were obtained by extracting the absolute weights learned by the model, averaging across folds, and then normalizing to have mean equal to 1 and standard deviation equal to 0. Finally, for each node, we averaged the weights between connections that included that particular node.

• 

  Figure 6. Meditation−specific weights of the expertise prediction model. The figure shows the normalized feature weights of the expertise prediction model averaged across cross-validation folds. These weights are specific for expertise prediction during (a) FA and (b) OM. A positive (red) weight indicates an increase of connectivity between nodes as the expertise increases, while a negative (blue) weight indicates a decrease of connectivity as the expertise increases. The circle size shows the absolute average weight of connections that include that specific node.

• 

  Figure 7. Connectivity weights shared across different meditation styles. The figure shows the normalized feature weights of the predictive connectivities that are shared in the meditation styles. Weights were averaged across cross-validation folds. These weights were learned by the model to predict (a) expertise and (b) age in both FA and OM meditation. The shown connectivities are those chosen by the feature selection algorithm in at least 95% of the folds in both FA and OM meditation. A positive (red) weight indicates an increase of connectivity between nodes as the expertise increases, while a negative (blue) weight indicates a decrease of connectivity as the expertise increases. The circle size shows the absolute average weight of the connections that include that specific node.