main.py
-
from scripts/main_con.py : calls
con_matrix()-
Returns data that contains a all important variables used, plus the connectomes for each condition, and the difference post-pre node and edge-wise.
E.g. pre_connectomes = [conMatrix-sub01, ..., conMatrixSubXX]
Dict keys :
['subjects', 'conditions', 'pre_series', 'post_series', 'labels', 'atlas', 'seed_pre_series', 'seed_post_series', 'seed_to_pre_correlations', 'seed_pre_masker', 'seed_to_post_correlations', 'seed_post_masker', 'ttest_rel', 'connectivity_obj', 'pre_connectomes', 'post_connectomes', 'diff_weight_connectomes'] -
-
calls
connectome_analyses()- returns a dict
graphsthat contains dict for different metric/conditions
Dict keys :
(['pre_graphs', 'post_graphs', 'pre_nodes_metrics', 'post_nodes_metrics', 'change_nodes', 'randCon_pre', 'randCon_post', 'randCon_nodeChange_dfs', 'pval_ls_nodes'])pre_graphs : One key for each subject randomized arrays (e.g. ranCon_post) : One key/subject,
Dict[randCon_post][sub]is alistof permtuted connectome or graph of lenght perm=N - returns a dict
-
calls
prediction_analyses()
-
Input :
graphs: graph[change_nodes] = list of Node x node Metric arrays of len(subs) -
Specific columns e.g. nodeStrenght will be extracted for each subj, then inversed and used as a N sub X M nodes feature matrix to predict Y.
-
returns an embeded dict that has four levels :
regressionDict>Condition>Node metric>Yi variable> Performance mertrics
Dict['change'].keys(): Each node metric/ROI feature matricesdict_keys(['strengthnorm', 'eigenCent', 'edge_weight', 'Supramarginal gyrus', 'Anterior Cingulate Cortex', 'Cingulate gyrus mid-anterior', 'Cingulate cortex posterior'])That is, dict per node metric or ROI nodes that contains :
reg_dict['change']['strengthnorm'].keys()Contain a dict for each dependant variable/model (Yi~node metrics)dict_keys(['SHSS_score', 'total_chge_pain_hypAna', 'Chge_hypnotic_depth', 'Mental_relax_absChange', 'Abs_diff_automaticity'])reg_dict['change']['strengthnorm']['SHSS_score'].keys()Contains performance metrics, and input to visualize cross-val info :dict_keys(['plot_title', 'CV_mse', 'CV_rmse', 'CV_r2', 'CV_pearson_r', 'pca_n_components', 'r2p_value', 'rmse_p_value', 'y_preds', 'y_tests', 'coeffs', 'corr_coeffs'])
def load_dict(p,folder, name):
with open(os.path.join(p,folder, name), 'rb') as f:
results = pickle.load(f)
return results
con_dict = load_dict(p, resultsFolder, 'dict_connectomes.pkl')
graph_dict = load_dict(p, resultsFolder, 'graphs_dict.pkl')
reg_dict = load_dict(p, resultsFolder, 'cv_results.pkl')
Summary: "This project aims to better understand neural correlates of hypnosis, which is defined as an experiential experience of focused attention and heighten reponse to suggestions. Hypnotic experience can be assessed in part trough the feeling of automaticity associated with hypnotic experience, with hypnotic depth measures and with hypnotizalibility scores. The brain's functionnal connectivity might reflect which brain regions interact together to produce the subjective change in phenomenological experience associated with hypnosis. We used Arterial Spin Labeling fMRI (Siemens 3T) data acquired at rest (from Rainville et al., 2019), both before and after hypnotic induction. The functional connectivity matrix for each subject was computed, for each condition (pre and post hypnosis) and then substracted to extract the contrast connectivity (post-pre hypnosis) induced by hypnosis. We then used the contrast connectivity matrices and graph theory metrics extracted from the connectomes (degree, clustering and centrality) to predict hypnosis-related behavioral variables. A 10-fold (80:20 train-test split) ridge regression was trained to predict (separately) the dependent variables.'
tags: [ASL-fMRI, hypnosis, connectome, ML]
Hypnosis is defined as an experiential state of focused attention and heightened response to suggestions. Hypnotic experience can be assessed in part with the automaticity associated with hypnotic experience, with hypnotic depth and with hypnotizalibility scores. Efficiency of hypnosis to modulate pain is well grounded in litterature (see Thompson et al., 2019 for a systematic review).
For a informative discussion about neural correlates involved in hypnosis and clinical implications, by David Speigel, see this link.
Data Used comes from Rainville et al. (2019) and hypnosis was used during the fMRI scanning acquisition with a Simens 3T scanner.
Outline
33 participants (female=17,male = 16, 27 y.o mean)
- 2 resting state scans (6 minutes each)
- Conditions : Pre-hypnosis, Post-hypnosis
- Hypnosis-related measures
- Standford hypnotizability scale (SHSS)
- Hypnotic depth (HDS)
- Hypnotic automaticity (HAS)
- Hypnotic experience (HEQ)
A 6 minutes resting state scan was first acquired at the beginning of the protocol (pre hypnosis). Other pain-related tasks detailed in Desmarteaux et al. (2021) were performed following the initial resting state acquisition. These tasks invlove a series of painful stimulations coupled with hypnotic suggestions to either increase or decrease the pain. Some variables related to this part of the protocol were used in the present project as dependent varibles (e.g. the efficiency of pain modulation induced by hypnosis). Other hypnosis-related measures are derived from this part of the protocol, e.g. measures of feeling of automaticity, mental relaxation and hypnotic depth. At the end of the protocol, participant took part in a 15 minutes hypnotic induction prior to a second (post hypnosis) resting state scan (6 minutes).
The functional MRI signal is Arterial splin labeling which measure crebral blood flow over time. Altough this signal is not as precise as the BOLD signal (lower S/N ratio), it has the advantage of being absolute and not relative to a baseline. This is particularly useful for the resting state scan, as the baseline is not well defined.
*Note on pain modulation : Other tasks in the scanning protocol - aside from the 2 resting state scans- included conditions of verbal suggestions either to increase (hyperalgesia cond.) or to decrease (analgesia cond.) pain experience prior to a series of painful stimulation. Subjective pain was assessed after each trial of [verbal suggestion + painful stimulations]. The difference between the pain ratings in the neutral condition and the pain ratings in the analgesia/hyperalgesia condition was computed for each subject. The absolute change in pain modulation was then used as a regressor in the analysis.
Data access is unfortunately restricted from public open access. However, group results can be shared on demand. Please contact me at [email protected]
- A github repository with all the code and documentation organized in a standard way; main.py, src, scripts, images, notebooks and a README.md (this document) file.
- A jupyter notebook with the prediction analyses and plots used in this report.
- Yeo7 atlas
- Dictionary of Functional Modes (DiFuMo) 64 ROIs atlas
- Pearson correlation
- Partial correlation
- Tangent space embedding
The following section on the Yeo atlas is mainly to ground this project in a knowledge sharing perspective and for transparacy, as this atlas was not utlimately the one chosen for the main analyses. The DiFuMo atlas, presented after, was used for the project.
-
Atlas from Yeo et al, 2011
from nilearn import datasets atlas = datasets.fetch_atlas_yeo_2011()["thick_7"]This atlas was first consider for its parcimony as it only comprises 7 networks. With that many ROIs, the number of possible pairs is 21 and it would not be desirable to have too much regressors in multiple regression models where we have few samples. is a deterministic atlas. As metionned in the nilearn docs :
"A deterministic atlas is a hard parcellation of the brain into non-overlaping regions, that might have been obtained by segmentation or clustering methods. These objects are represented as 3D images of the brain composed of integer values, called ‘labels’, which define the different regions."This implies the use of Nilearn.maskers.NiftiLabelsMasker class to extract timeseries from each regions of interest (ROI).
*/ Symetrical bilateral mask
Since the atlas is symetrical, it is labeled bilaterally with the same label in the left and right hemispheres. When the connectivity matrix is directly computed on this atlas, it renders a symetrical connectome without inter-hemispheric connections, as illustrated in this connectome from a nilearn tutorial :
Compute inter-hemispheric connectivity
To account for this, I had to relabel each ROI with a different label for the right hemisphere. This was done locally with the scripts.func.make_mask_bilat(atlas) function, that take each ROI and duplicate it with a different label for the right hemisphere. Hence, after this tweak, the atlas had 14 ROIs and the number of possible pairs is now of 91!
Fine-grain atlas of functional modes for fMRI analysis
- This atlas is originally presented in Dadi et al, 2020
- Can be downloaded here
- Note that finer parcellation of the atlas are also available here
- Labels are provided Exemple of ROI/modes shown on the website
- It is a probabilstic atlas (4 dimenional Nifti file of shape Xdim, Ydim, Zdim, Number of ROIs). Each 3D slice of this 4D file contains a ROI. Hence, some voxels can be in more than one ROI.
Image from Dadi et al. (2020), e.i. the article introducing the DiFuMo atlas.
It shows the 'Comparison of modes overlap for all proposed DiFuMo atlases. The y-axis shows how many voxels are at the intersect of exactly n modes (x-axis), without thresholding the modes. On average, at least two modes are shared between voxels.'(Dadi et al. 2020)
This atlas was chosen for its balance between parsimony (64 ROIs) and for its anatomico-functional precision. Also, the process used for the development of this atlas is very rigorous in the computation of the functional ROIs (modes), but also in the labeling process of the ROIs.
The covariance matrix is computed with the Nilearn.connectome.ConnectivityMeasure class. This class has the kind parameter that can be set to covariance, correlation, partial correlation, tangent or precision. Code example :
from nilearn.connectome import ConnectivityMeasure
correlation_measure = ConnectivityMeasure(kind='correlation')
A word on Sparse inverse covariance (precision)
- Citation from 6.2.1 Nilearn :
"As shown in [Smith 2011], [Varoquaux 2010], it is more interesting to use the inverse covariance matrix, ie the precision matrix. It gives only direct connections between regions, as it contains partial covariances, which are covariances between two regions conditioned on all the others [...] To recover well the interaction structure, a sparse inverse covariance estimator is necessary."
"The sample correlation matrix is a poor estimate of the underlying population correlation matrix.[...] the sample correlation matrix captures a lot of sampling noise, intrinsic randomness that arises in the estimation of correlations from short time series. Conclusions drawn from the sample correlation matrix can easily reflect this estimation error" (Varoquaux et al., 2013)
Therefore, as shown in Varoquaux et al. (2010) and Smith et al. (2011), other estimator should be prefered to recover the connectivity structure
-
The inverse covariance matrix captures the direct link between ROIs, by removing the effect of other ROIs. This is done by computing the partial correlation between each pair of ROIs.
- "Using inverse-covariance matrices or partial correlations to understand brain connectivity makes the interpretation in terms of interactions between brain regions easier and more robust to the choice of confounds" (Varoquaux et al., 2013)
- In small samples, the removal of confounding indirect correlations is challenging from a statistical standpoint. To have a good inverse covariance estimate, a sparsity assumption (few connections to estimate) needs to be respected. (Varoquaux et al., 2013)
Computation of correlation assumes that the ROIs form a linear model, which is not always the case. The tangent embedding is a projection of the correlation matrix, that is more robust to non-linearities and spatial dependancies in the ROIs. Image from Rahim et al. (2019) to illustrate the tangent space
-
"The [Tangent method] is less frequently used but has solid mathematical foundations and a variety of groups have reported good decoding performances with this framework (Varoquaux et al., 2010a; Barachant et al., 2013; Ng et al., 2014; Dodero et al., 2015; Qiu et al., 2015; Rahim et al., 2017; Wong et al., 2018)" (Dadi et al., 2019)
-
"Using the tangent embedding as a connectivity measure gives a consistent improvement over correlations and partial correlation"(Dadi et al., 2016)
The tangent method was chosen for this study, as it is a mathematically well grounded method and as it is less sensitive to non-linearities and spatial dependancies in the ROIs.
-
Computation of the connectivity matrix for each subject, for each condition (pre/post hypnosis)
-
Computation of the difference between the two conditions (post-pre hypnosis element-wise in the adjacency/connectivity matrices) for each subject's connectome.
A Fischer R to Z transfomation was applied to each connectivity matriz prior to the the (post - pre) substraction. This was done to avoid substraction of pearson r's, which is not recommended. The numpy.arctanh was used for such transformation. Code example :
# Element-wise substraction of the two matrices import numpy as np contrast_matrix = np.arctanh(post_matrix) - np.arctanh(pre_matrix) -
The edges' values of the resulted contrast connectome are thus the difference in Z scores obtained from the element-wise Z score substraction post-pre.
- Binarizing the matrices based on threshold (e.g. 20% strongest weights)
- Neg. weights are included and binarized (?)
- Absolute values of edges' weights (as in Centeno et al., 2022) --> keep potential biological relevent information
-
Degree : 'The weighted node degree is the sum of the edge weights for edges incident to that node'(NetworkX docs)
-
Centralities : 'Centrality measures are a way of detecting the most important nodes in a graph. Informally, a central node is one that has a large number of connections to other nodes in the network.' (NetworkX docs)
Image from 'hands‑on tutorial on network and topological neuroscience' by Centeno et al, 2022
- Consideration for centrality measures "An alternative is to take the absolute value of the correlation coefficient, in which case a strong negative correlation will be given an equal weighting to a strong positive correlation. The decision of which trans- formation to use depends on how we choose to interpret the polarity of the edge weights." Ch. 5, p.141, Fornito et al., 2016
-
Clustering : "The clustering coefficient assesses the tendency for any two neighbours of a vertex to be directly connected (or more strongly connected in the weighted case) to each other and can also be termed cliquishness (Hallquist and Hillary 2018; Watts and Strogatz 1998)" (Centeno et al, 2022) Clustering suited for negative weights, implemented in NetworkX library, based on Generalization of Clustering Coefficients to Signed Correlation Networks by G. Costantini and M. Perugini, PloS one, 9(2), e88669 (2014).
The main goal of this project, beside to explore the functional connectome related to hypnosis, is to predict hypnosis-related variables from the functional connectivity. The hypnosis-related variables used are the following :
Dependent variables :
-
SHSS score
-
Raw change in pain -ANA condition
-
Raw change in pain - Hyper condition
-
Absolute change in pain (Hyperalgesia, Analgesia, Ana+Hyper )
-
Change in hypnotic depth
-
Change in mental relaxation
-
Automaticity post induction
-
Change in automaticity
Variables names in the excel file : ['SHSS_score', 'raw_change_ANA','raw_change_HYPER', "Abs_chge_pain_hypAna", "Chge_hypnotic_depth", "Mental_relax_absChange", 'Automaticity_post_ind',"Abs_diff_automaticity"]
The graphs' metrics described in section 2.5 above were used to predict the change in hypnosis-related variables with the contrast (post-pre) connectivity.
-
Vectorize half of the connectivity matrices, excluding the main diagonal, for each subject and concatenate in a numpy array of shape (n_subjects, n_features). Example of code used for the vectorization :
# Triangular lower masker on connectivity matrix M = results["pre_connectomes"] tril_mask = np.tril(np.ones(M.shape[-2:]), k=-1).astype(bool) # Stack the vectorized matrices results["X"] = np.stack(M[i][..., tril_mask] for i in range(0, len(M)),axis=0) -
Principal component analysis (0.80% explained variance) was applied to the feature matrix to reduce the dimensionality of the data, because multiple regression is not suited when nb. of features > nb. of subjects.
from sklearn.decomposition import PCA pca = PCA(n_components=0.80) results["X_pca"] = pca.fit_transform(results["X"]) -
Multiple regession with Ridge regression model was used to predict the change in hypnosis-related variables with the contrast (post-pre) connectivity.
The presented connectome is the mean of each contrast (post - pre hypnosis) individual connectomes. The edges are the difference post-pre in Z scores. It hence reflects the mean hypnosis-induced change in connectivity.
Brain plot of the 1% strongest edge values (Z difference post-pre hypnosis)
The tresholded mean connectome (the one reported above) was binarized. The logic is to assign 1 to an edge/connection if it survives the 1% highest edge treshold, else edge = 0. From that binarized matrix, the frequency of the remaining connections is extracted for each node/ROI. The nodes with the most non-zero connections, e.i. having the most frequent high edge values (according to the 1% treshold) are reported below :
Calcarine sulcus anterior: 7 non-zero connections
Paracentral lobule superior: 7 non-zero connections
Parieto-occipital sulcus anterior: 5 non-zero connections
Cuneus: 5 non-zero connections
Occipital pole: 4 non-zero connections
Angular gyrus inferior: 4 non-zero connections
Central sulcus: 4 non-zero connections
Caudate: 4 non-zero connections
- 'In an undirected weighted network like our rsfMRI matrix, the vertex degree is analogous to the vertex strength (i.e., the sum of all edges of a vertex) and equivalent to its degree centrality' (Centeno et al, 2022)
Note : The nodes represented in the tresholded brain plot are the first six nodes displayed in the first plot, e.i. the nodes with the highest degree value.
- "Closeness (shortest path-based) centrality measures how closely or’ directly’ connected a vertex is to the rest of the network." (Centeno et al, 2022)
- "Betweenness (shortest path-based) centrality is the proportion of all vertex-pairs shortest paths in a network that pass through a particular vertex (Newman 2008; Freeman 1977)" (Centeno et al, 2022)
- "The clustering coefficient assesses the tendency for any two neighbours of a vertex to be directly connected (or more strongly connected in the weighted case)" (Centeno et al, 2022)
Dependent variables :
- SHSS score
- Raw change in pain -ANA condition
- Raw change in pain - Hyper condition
- Absolute change in pain (Hyperalgesia, Analgesia, Ana+Hyper )
- Change in hypnotic depth
- Change in mental relaxation
- Automaticity post induction
- Change in automaticity
-
Ridge regression was the chosen model for its simplicity and its ability to deal with multicollinearity by penalizing the coefficients.
-
Cross-validation (10 folds), with a shufflesplit spliting strategy (80:20) was applied.
-
PCA with 90% explained variance was applied prior to the regression to reduce the number of features.
pipeline = Pipeline([('std', StandardScaler()),('pca', PCA(n_components=0.90)), ('ridge', Ridge())]) k = 10 # Number of folds kf = ShuffleSplit(n_splits=k, test_size=0.20, random_state=random_seed)
The full connectivity matrix of each subject was used for prediction (before applying the PCA). Each edges (2016 in total) was considered as a feature.
Note : the metrics reported are the mean of the 10 folds and each regression line is the correlation (pearson) between predicted values and the real values for one fold.
The feature matrix used for prediction has the shape (N=31 x Nodes=64).
The Standford Hypnotic Susceptibility Scale (SHSS) score was used to classify the subjects in three groups : low (0-2), medium (3-5) and high (6-8), based on the connectivity matrices, on the node degree, on the node centrality and on the clustering coefficient.
- Ensemble classifier (VotingClassifier) including :
- LogisticRegression
- RandomForestClassifier
- SVC PCA was applied prior to the classification to reduce the number of features.
- A stratefiedKFold (5 folds) was used to split the data into train and test sets. This method accounts for the unbalanced distribution of the classes.
Accuracy: 0.4516 Precision: 0.4452 Recall: 0.4516 F1-score: 0.4415
Accuracy: 0.3871
Precision: 0.3903
Recall: 0.3871
F1-score: 0.3870
Accuracy: 0.3871
Precision: 0.3903
Recall: 0.3871
F1-score: 0.3870
Accuracy: 0.3548
Precision: 0.3592
Recall: 0.3548
F1-score: 0.3557
This project was in part executed as part of the BrainHack School (BHS) project. Thank you to all the BHS teaching assistants for the support and great advices on this project and also to the organizers who made such a interactive and dynamic course possible. This project was executed in collaboration and under the supervision of Pierre Rainville and Mathieu Landry, and some additional help from Jen-I Chen regarding data analysis and preprocessing.
Gael Varoquaux, Flore Baronnet, Andreas Kleinschmidt, Pierre Fillard, and Bertrand Thirion. Detection of brain functional-connectivity difference in post-stroke patients using group-level covariance modeling. In Tianzi Jiang, Nassir Navab, Josien P. W. Pluim, and Max A. Viergever, editors, Medical image computing and computer-assisted intervention - MICCAI 2010, Lecture notes in computer science, 200–208. Berlin, Heidelberg, 2010. Springer. doi:10/cn2h9c
G. Varoquaux, A. Gramfort, J.B. Poline, B. Thirion, Brain covariance selection: better individual functional connectivity models using population prior, in: NIPS, 2010.
Kamalaker Dadi, Alexandre Abraham, Mehdi Rahim, Bertrand Thirion, Gaël Varoquaux. Comparing functional connectivity based predictive models across datasets. PRNI 2016: 6th International Workshop on Pattern Recognition in Neuroimaging, Jun 2016, Trento, Italy. hal-01319131
Kamalaker Dadi, Mehdi Rahim, Alexandre Abraham, Darya Chyzhyk, Michael Milham, Bertrand Thirion, Gaël Varoquaux, Benchmarking functional connectome-based predictive models for resting-state fMRI, NeuroImage,Volume 192,2019,Pages 115-134,ISSN 1053-8119,https://doi.org/10.1016/j.neuroimage.2019.02.062.
Mehdi Rahim, Bertrand Thirion, Gaël Varoquaux, Population shrinkage of covariance (PoSCE) for better individual brain functional-connectivity estimation, Medical Image Analysis,Volume 54,2019,Pages 138-148,ISSN 61-8415, https://doi.org/10.1016/j.media.2019.03.001.
S. Smith, K. Miller, G. Salimi-Khorshidi, M. Webster, C. Beckmann, T. Nichols, J. Ramsey, M. Woolrich, Network modelling methods for fMRI, Neuroimage 54 (2011) 875.
Thompson T, Terhune DB, Oram C, Sharangparni J, Rouf R, Solmi M. The effectiveness of hypnosis for pain relief: A systematic review and metaanalysis of 85 controlled experimental trials. Neurosci Biobehav Rev. (2019). 99:298–310 .doi: 10.1016/j.neubiorev.2019.02.013
Thresholded connectome
-
Pair of nodes with highest change in edge value (Z difference post-pre):
[Occipital pole - Paracentral lobule superior]; Z=0.1647 [Paracentral lobule - Paracentral lobule superior]; Z=0.1628 [Occipital pole - Paracentral gyrus RH]; Z=0.1596 [Fusiform gyrus - Occipital pole]; Z=0.1588 [Intraparietal sulcus RH - Superior parietal lobule anterior]; Z=0.1573 [Paracentral lobule superior - Precuneus superior]; Z=0.1556 [Cingulate gyrus mid-anterior - Calcarine sulcus anterior]; Z=0.1486 [Superior temporal sulcus with angular gyrus - Paracentral lobule]; Z=0.1475 [Middle temporal gyrus - Heschl’s gyrus]; Z=0.1457 [Inferior frontal sulcus - Descending occipital gyrus]; Z=0.1444 [Superior parts of Postcentral and Precentral gyri - Paracentral lobule superior]; Z=0.1433-------Negative edges-------
[Fusiform gyrus posterior - Inferior occipital gyrus]; Z=-0.2077 [Fusiform gyrus - Cerebellum Crus II]; Z=-0.2096 [Occipital pole - Descending occipital gyrus]; Z=-0.2570 [Occipital pole - Inferior occipital gyrus]; Z=-0.2711 -
Nodes with the most connections (>3) from the thresholded connectome:
Occipital pole: 7 non-zero connections Calcarine cortex posterior: 4 non-zero connections Superior temporal sulcus with angular gyrus: 4 non-zero connections Paracentral lobule superior: 4 non-zero connections
Density
-
5% nodes with highest density values(pos. and neg. values) :
Paracentral lobule superior: 0.7882 Superior occipital gyrus: -0.6706 Parieto-occipital sulcus middle: -0.6943
Centrality (betweeness)
-
5% nodes with highest centrality values :
Occipital pole: 0.0753 Paracentral lobule superior: 0.0445 Superior temporal sulcus with angular gyrus: 0.0394
Clustering
-
5% nodes with highest clustering values :
Superior parietal lobule anterior: 0.1670 Paracentral lobule: 0.1635 Paracentral lobule superior: 0.1519-
(Not displayed on graph) 5% nodes with lowest clustering values :
Cingulate gyrus mid-anterior: -0.013501491129290272 Inferior frontal gyrus: -0.013918529960704935 Anterior Cingulate Cortex: -0.01404254150619912
-
Thresholded connectome
-
Pair of nodes with highest change in edge value (Z difference post-pre):
[Middle frontal gyrus - Superior fornix and isthmus]; Z=0.1566 [Heschl’s gyrus - Paracentral gyrus RH]; Z=0.1521 [Middle frontal gyrus - Parieto-occipital sulcus anterior]; Z=0.1494 [Fusiform gyrus - Cingulate cortex posterior]; Z=0.1475 [ventricles - Middle frontal gyrus]; Z=0.1466 [Superior frontal sulcus - Dorsomedial prefrontal cortex antero-superior]; Z=0.1461 [Middle frontal gyrus - Putamen]; Z=0.1448 [Cerebellum I-V - Parieto-occipital sulcus anterior]; Z=0.1428 [Middle frontal gyrus anterior - Angular gyrus inferior]; Z=0.1419 [Middle frontal gyrus anterior - Putamen]; Z=0.1401 -
Nodes with the most connections (>3) from the thresholded connectome:
Paracentral lobule superior: 8 non-zero connections Caudate: 5 non-zero connections Superior frontal sulcus: 4 non-zero connections Heschl’s gyrus: 4 non-zero connections Middle frontal gyrus: 4 non-zero connections
Degree 'The weighted node degree is the sum of the edge weights for edges incident to that node'
- 5% nodes with highest degree values :
Middle frontal gyrus: 3.4630
Angular gyrus inferior: 3.2940
Calcarine sulcus anterior: 3.1590
Middle frontal gyrus anterior: 3.1503
Centrality (betweeness)
-
5% nodes with highest centrality values :
Calcarine sulcus anterior: 0.0661 Middle frontal gyrus: 0.0584 Parieto-occipital sulcus anterior: 0.0568
Clustering
-
5% nodes with highest clustering values :
Superior parietal lobule anterior: 0.1670 Paracentral lobule: 0.1635 Paracentral lobule superior: 0.1519
(For comparison purpose with the contrast results)
Parieto-occipital sulcus superior: 0.30290769035702064 Superior parietal lobule posterior: 0.2847421670281322 Precuneus anterior: 0.27548093040763855 Cingulate gyrus mid-posterior: 0.2749985525257869 Parieto-occipital sulcus anterior: 0.2708608648412702
Find below sections A and B of the results of the prediction of hypnosis-related variables with the connectivity matrix and the node degree as independent variables.
The full connectivity matrix of each subject was used for prediction (before applying the PCA). Each edges (2016 in total) was considered as a feature.
Note : the metrics reported are the mean of the 10 folds and each regression line is the correlation (pearson) between predicted values and the real values for one fold.
The feature matrix used for prediction has the shape (N=31 x Nodes=64).
Seed choices : Main regions reported in Desmarteaux et al. (2021)
- PO (54, -28, 26)
- HPG : (-20, -26, -14)
- aMCC : (-2, 20, 32)
- ACC : (-8. 44 28)
- PCC : (-6, -26, 46)