Participants

Participants were recruited from mailing lists and were split into regular meditators (19 subjects) or non-meditators (20 subjects) dependent on their responses. Regular meditators were considered to be those who practised meditation three times a week, and had been practising for at least three years. Non-meditators were those who reported practising less than once a week, or not at all.

The groups were matched for age (meditators: 45.47±9.47; non-meditators: 43.80±9.35), gender (meditators: 8M/11F; non-meditators: 9M/11F) and education level (meditators: 78% undergraduate degree, 22% post-graduate; non-meditators: 65% undergraduate degree, 25% post-graduate, 10% secondary school). There was no statistically significant difference between groups on any of these factors. On average, the meditator group had been regularly meditating for 8.5±4.1 years. The styles of meditation used in this group were: “zazen” (N = 4), mantra meditation (N = 2) mindfulness of breathing (N = 6), kriya yoga meditation (N = 4) and meditation associated with hatha yoga (N = 3).

Participants were screened for possible mental health problems, on-going psychological or psychiatric treatment, and use of psychotropic drugs under the supervision of a psychologist and a neuropsychiatrist. In addition, all participants were evaluated on the day of MRI scanning for depression (Beck Depression Inventory [10]), anxiety (Beck Anxiety Inventory [11]), mindfulness (Mindfulness Attention Awareness Scale [12]), and self-compassion (Self-Compassion Scale [13]). There was a significant difference in anxiety levels between the groups, with greater anxiety reported by non-meditators. However, neither group exhibited clinically relevant anxiety (Table S1).

Interviews were conducted a posteriori to assess diet and exercise habits. As the interviews took place after the MRI scan, only 13 regular meditators and 15 non-meditators could be contacted. There was no statistically significant difference in diet (vegan, lacto vegetarians, ovo-lacto vegetarians, meat eaters) between the groups (p-value = 0.583, Fisher’s Exact Test). There was also no significant difference in exercise habits (proportion undertaking physical activity at least once a week) between the groups (p-value = 1.000, Fisher’s Exact Test). Furthermore, the groups were similar in the categories of physical activity undertaken by participants (aerobic, anaerobic, stretching, or more than one category of activity; p-value = 0.373, Fisher’s Exact Test). See Table S2 for more detailed information on these possible confounders.

MRI Acquisition

A high resolution MR image was acquired for each participant using a Siemens 3.0T Magnetom Tim Trio System, We used a MPRAGE T1-weighted sequence (matrix 1×1×1 mm voxel, TR = 2500 ms, TE = 3.45 ms, FOV = 265 mm, inversion time = 1100, flip angle 7 degrees).

Structural Image Processing

The T1 weighted structural images of all participants were processed using automated cortical and subcortical segmentation (aseg.volume.stats and bilateral aparc.volume.stats files, see Information S1) within the recon-all pipeline of the Freesurfer package (http://surfer.nmr.mgh.harvard.edu). This procedure includes cortical surface modelling, spherical coordinate transformation, nonlinear curvature registration, and automated segmentation of cortical and subcortical structures. The estimated volume of each segmented region obtained using this routine was then used as the input to the classifiers. Further details about recon-all pipeline can be found at [14], [15], [16]. Note: Freesurfer software labels the basal putamen as ‘vessel’, since it is an area with prominent vascular space.

Classification

The pattern recognition method used in this study was the linear two-classes (regular meditators/non-meditators) Support Vector Machine (SVM). To implement this method, we used e1071 package (which provides an interface to the libSVM library [17], www.csie.ntu.edu.tw/~cjlin/libsvm/) within the open source R environment (www.r-project.org). Detailed information about SVM implementation can be found in previous publications [3].

In brief, the basic idea when using pattern recognition analyses is to try to predict the class of an observation (e.g. controls vs patients) based on selected predictor variables (e.g. image features). The two-classes SVM method used in this study is a supervised machine learning approach. This means that the method “learns” to discriminate between two classes based on correctly categorised training data (accurate class and predictor variables from example data). When these classification rules are sufficiently “learnt” then SVM is able to generate class predictions for novel observations (test data). SVM has been used in conjunction with neuroimaging to discriminate patients from controls [18], [19], [20] and also to differentiate distinct brain states based on functional MRI [21], [22]. In the majority of these studies, the predictor variables were specific measures at each voxel of the brain (e.g. gray-matter coefficients, normalized fMRI signal, etc.) and the classes were the disease (present or absent) or experimental condition (e.g.: Task A or Task B in fMRI studies). One of the appealing properties of pattern recognition methods compared to conventional t-tests is that the former is able to generate predictions (and thus assess the amount of predictive information contained within a set of variables), and not only evaluate whether a variable is statistically different between groups.

The current study evaluated whether the information contained in structural (T1 weighted) images was capable of predicting or discriminating between regular meditators and non-meditators. The volumes of each segmented region (121 areas, expressed in cubic millimeters) were used as the variables (features) for group prediction. The names of these predictor variables can be found in Supplementary Information. A feature selection step was included during classification analysis to reduce the influence of irrelevant variables, and also highlight the brain regions containing the most discriminant information. In this way, feature selection can be used as a brain mapping tool.

One of the main dangers when performing classification analysis is double dipping and overfitting, which may lead to unreliable estimates of classifier’s accuracy. These problems can be even worse when a feature selection step is included. In order to avoid these problems, the classifier’s accuracy was estimated based on a first-level leave-one-subject-out procedure and the feature selection was carried out in a second-level nested-leave-one-subject out procedure. This second process was required to guarantee that the information from the specific test subject removed at first-level leave-one-out analysis was only used to estimate prediction accuracy and was not contained within the SVM training data. The feature selection, classification, and accuracy estimation were performed via the following steps:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Classification of regular meditators and non-meditators using support vector machines (SVM).

Regions identified by the SVM as containing discriminative information used to consistently predict the groups (right precentral gyrus, left entorhinal cortex, right pars opercularis cortex, right basal putamen, and bilateral thalamus). These five regions were selected by SVM in an all leave-one-subject-out iterations, with 94.87% accuracy. The bottom of the figure depicts the projection values of each subject and the decision boundary.

https://doi.org/10.1371/journal.pone.0039832.g001

1. Step 1) . Leave one subject out of the sample (first-level leave-one-out);
2. Step 2) . Remove the effects of gender and age from each feature (predictor variable) of the training data by using a multiple linear regression analysis. The corrected training data are the residuals of this regression;
3. Step 3) . Normalize each feature of the corrected training data to have mean zero and variance one. This data is referred as the normalized training data;
4. Step 4) . Leave-one-out implementation:
5. Step 4.1) . Leave another subject out of the normalized training data (second-level leave-one-out);
6. Step 4.1.a) . Train the linear SVM using the respective normalized training data and its label vector (which specifies the groups).
7. Step 4.1.b) . Rank the SVM decision function coefficients (hyperplane coefficients) by their absolute values. This step will provide a rank vector describing the relevance of each feature to the groups’ discrimination;
8. Step 4.1.c) . Feature selection: Remove the most irrelevant feature from the normalized training data;
9. Step 4.1.d) . Train the SVM using the normalized training data obtained in step 4.1.c;
10. Step 4.1.e) . Predict the class of the subject left out in step 4.1;
11. Step 4.1.f) . Return to step 4.1.c and repeat until all features have been removed;
12. Step 4.2) . Return to step 4.1 until all iterations of second-level leave-one-subject-out have been carried out;
13. Step 5) . Compute the classification accuracies of the second-level leave-one-out for different numbers of features. Obtain the number of features, Q, which maximizes the second-level leave-one-out accuracy;
14. Step 6) . Train the linear SVM using the normalized training data obtained in step 3;
15. Step 7) . Obtain the rank vector in the same way as step 4.1.b;
16. Step 8) . Use the rank vector to build a normalized training data consisting solely of the Q (estimated in step 5) most discriminant features;
17. Step 9) . Train the linear SVM using the normalized training data from step 8;
18. Step 10) . Apply the covariate correction and normalization (based on the parameters from step 2 and 3) to the features of the subject left out in step 1 (first-level leave-one-out);
19. Step 11) . Classify the subject left out (first-level) using the test data from step 10;
20. Step 12) . Return to step 1 until all iterations of first-level leave-one-subject-out have been carried out;
21. Step 13) . Compute the first-level leave-one-out accuracy;

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Boxplot illustrating the volumetric information of the regions containing the greatest discriminative information, and ROC curves.

https://doi.org/10.1371/journal.pone.0039832.g002

Finally, the p-value for the significance of the first-level leave-one-out accuracy was obtained by using the Binomial distribution. One important point to be mentioned is that the number of features (see step 5) used by SVM is different for each iteration of the first-level leave-one-out. The most discriminant features (brain regions) referred to in the Results and Discussion sections are those which were selected by SVM across all iterations. However, the classification results are based on all the discriminant features found in all leave-one-out iterations.