Edited by: Misha Tsodyks, Weizmann Institute of Science, Israel
Reviewed by: Meng Hu, Drexel University, USA; Abdelmalik Moujahid, University of the Basque Country, Spain
*Correspondence: Svyatoslav Vergun, Medical Physics, University of Wisconsin–Madison, 1122-Q2 Wisconsin Institutes for Medical Research, 1111 Highland Avenue, Madison, WI 53705, USA. e-mail:
This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in other forums, provided the original authors and source are credited and subject to any copyright notices concerning any third-party graphics etc.
The brain at rest consists of spatially distributed but functionally connected regions, called intrinsic connectivity networks (ICNs). Resting state functional magnetic resonance imaging (rs-fMRI) has emerged as a way to characterize brain networks without confounds associated with task fMRI such as task difficulty and performance. Here we applied a Support Vector Machine (SVM) linear classifier as well as a support vector machine regressor to rs-fMRI data in order to compare age-related differences in four of the major functional brain networks: the default, cingulo-opercular, fronto-parietal, and sensorimotor. A linear SVM classifier discriminated between young and old subjects with 84% accuracy (
Functional networks are defined by a temporal correlation of brain regions normally involved during a task and are observed when individuals are resting without performing a specific task (Biswal et al.,
Research efforts in functional magnetic resonance imaging (fMRI) are shifting focus from studying specific cognitive domains like vision, language, memory, and emotion to assessing individual differences in neural connectivity across multiple whole-brain networks (Thomason et al.,
Simultaneously, use of machine learning techniques for analyzing fMRI data has increased in popularity. In particular, Support Vector Machines (SVMs) have become widely used due to their ability to handle very high-dimensional data and their classification and prediction accuracy (Schölkopf and Smola,
With task-based fMRI data, LaConte et al. (
Resting state fMRI data has been shown viable in classification and prediction. Craddock et al. (
One advantage of resting state data as opposed to task-based data is that the acquiring of resting data is not constrained by task difficulty and performance. This provides a potentially larger group of subjects that are not able to perform tasks (e.g., Alzheimer’s Disease patients, patients with severe stroke) on which studies can be done. There has been a great amount of progress made in describing typical and atypical brain activity at the group level with the use of fMRI, but, determining whether single fMRI scans contain enough information to classify and make predictions about individuals remains a critical challenge (Dosenbach et al.,
We describe a classification and regression method implemented on aging adult rs-fcMRI data using SVMs, extracting relevant features, and building on the SVM/SVR study of children to middle-aged subjects (Dosenbach et al.,
Resting state data for 65 individuals (three scans each) were obtained from the ICBM dataset made freely accessible online by the 1000 Connectome Project
The analyses described in this work were performed on two data sets contained in the ICBM set. The same preprocessing algorithms were applied to both sets of data.
Data set 1 consisted of 52 right-handed individuals (age 19–85, mean 44.7, 23M/29F). This was the binary SVM set (both for age and gender classification) which contained a young group of 26 subjects (age 19–35, mean 24.7, 12M/14F) and an old group of 26 subjects (age 55–85, mean 64.7, 11M/15F).
Data set 2 consisted of 65 right-handed individuals (ages 19–85, mean 44.9, 32M/33F). This was the mcSVM set as well as the SVR age prediction set. It contained three age groups used for mcSVM: a young group of 28 subjects (age 19–37, mean 25.5, 14M/14F), a middle-aged group of 22 subjects (age 42–60, mean 52.4, 11M/11F), and an old group of 15 subjects (age 61–85, mean 69.9, 7M/8F).
Resting data were acquired with a 3.0 Tesla scanner using an echo planar imaging (EPI) pulse sequence. Three resting state scans were obtained for each participant, and consisted of 128 continuous resting state volumes (TR = 2000 ms; matrix = 64 × 64; 23 axial slices). Scan 1 and 3 had an acquisition voxel size = 4 mm × 4 mm × 5.5 mm, while scan 2 had an acquisition voxel size = 4 mm × 4 mm × 4 mm. All participants were asked to keep their eyes closed during the scan. For spatial normalization and localization, a T1-weighted anatomical image was acquired using a magnetization prepared gradient echo sequence (MP-RAGE, 160 sagittal slices, voxel size = 1 mm × 1 mm × 1 mm).
Data were preprocessed using AFNI (version AFNI_2009_12_31_1431
Nuisance signal [white matter, cerebrospinal fluid (CSF) and six motion parameters] was then removed from the preprocessed fMRI data. White matter and CSF masks were created using FSL by the segmentation of each individual’s structural image. These masks were then applied to each volume to remove the white matter and CSF signal. Following the removal of these nuisance signals, functional data were then transformed into Montreal Neurological Institute 152 (MNI152-brain template; voxel size = 3 mm × 3 mm × 3 mm) space using a two-step process. First a 6 degree-of-freedom affine transform was applied using FLIRT (Smith et al.,
One hundred functionally defined regions of interest (ROIs) encompassing the default mode, cingulo-opercular, fronto-parietal, and sensorimotor networks (see Figure
Average resting state blood oxygenation level dependent (BOLD) time series for each ROI were extracted. The BOLD time series for each ROI were then correlated with the BOLD time series of every other ROI (Pearson’s correlation) for every subject and every scan. This resulted in a square (100 × 100) symmetric matrix of correlation coefficients for each scan, but only 4950 ROI-pair correlation values from the lower triangular part of the matrix were retained (redundant elements and diagonal elements were excluded). These were then
The SVM is a widely used classification method due to its favorable characteristics of high accuracy, ability to deal with high-dimensional data and versatility in modeling diverse sources of data (Schölkopf et al.,
A SVM is an example of a linear two-class classifier, which is based on a linear discriminant function:
The vector
A brief description of the SVM optimization problem is given here and a more detailed one can be found in Vapnik’s (
In the soft margin SVM (Cortes and Vapnik,
The constant
This can be represented in a dual formulation in terms of variables αi (Cortes and Vapnik,
The dual formulation leads to an expansion of the weight vector in terms of input data examples:
The examples
The discriminant function then becomes:
The dual formulation of the optimization problem depends on the data only through dot products. This dot product can be replaced with a non-linear kernel function,
Some commonly used kernels are the polynomial kernel and the Gaussian kernel. In this work we used a linear kernel and a Gaussian kernel, which is also called a radial basis function (RBF):
We tuned the value of
With some datasets higher classification accuracies can be obtained with the use of non-linear discriminating boundaries (Ben-Hur and Weston,
Drucker et al. (
Epsilon-insensitive SVR defines a tube of width ε, which is user defined, around the regression line in high-dimensional space. Any points within this tube carry no loss. In essence, SVR performs linear regression in high-dimensional space using epsilon-insensitive loss. The
We used leave-one-out-cross-validation (LOOCV) to estimate the SVM classification and SVR prediction accuracy since it is a method that gives the most unbiased estimate of test error (Hastie et al.,
In a round, or fold, of LOOCV, an example from the example set is left out and is used as the entire testing set, while the remaining examples are used as the training set. So each example is left out only once and the number of folds is equal to the number of examples. In our work, LOOCV was performed across participants, not scans, so three scans per participant were removed in each fold and used only in the testing set to avoid “twinning” bias.
During each SVM LOOCV fold, two-sample
One important aspect of SVM and SVR is the determination of which features in the model are most significant with respect to example classification and prediction.
For linear kernel SVM and SVR features, the individual weights of the features as given by the SVM or SVR revealed their relative importance and contribution to the classification or prediction. In the linear kernel SVM and SVR method each node’s (ROI’s) significance, as opposed to each feature’s significance, was directly proportional to the sum of the weights of the connections to and from that node.
Feature connections and nodes were visualized using BrainNet Viewer (Version 1.1
Dosenbach et al. (
A holdout set of 20 randomly chosen subjects was used to tune the SVM age and gender classification parameters. We limited ourselves to number of features <1000 for two reasons: previous work (Dosenbach et al.,
A similar procedure for the SVR method was taken. A holdout set of 25 randomly chosen subjects was used to tune the SVR age prediction parameters. First, slope (of a linear regression line fitting the predicted age) as a function of top features was computed to reveal a peak performance area. Then, slope as a function of the number of features and ε was output with a grid search method. The number of features and value of ε that maximized the slope and
The binary SVM classifier, using a linear kernel, was able to significantly discriminate between young and old subjects with 84% accuracy (
The linear kernel SVM classifier outperformed the RBF kernel SVM classifier with this dataset and a comparison of the two classifiers is given in Table
Classifier | Accuracy (%) | Top features retained | |
---|---|---|---|
Linear SVM | 84 | 100 | 0.1 |
Rbf SVM | 81 | 62 | 1.0 |
Of the 100 total features retained per fold, 63 were present in every fold and these are called the consensus features. Table
Feature index | SVM feature number | ROI 1 | Connected with | ROI 2 | Weight |
---|---|---|---|---|---|
1 | 632 | L_precentral_gyrus_3 | L_vent_aPFC | 0.3119 | |
2 | 1037 | L_sup_frontal | R_sup_frontal | 0.4479 | |
3 | 1038 | M_ACC_2 | R_sup_frontal | 0.2472 | |
4 | 1047 | L_basal_ganglia_1 | R_sup_frontal | 0.1405 | |
5 | 1048 | M_mFC | R_sup_frontal | 0.203 | |
6 | 1231 | R_pre_SMA | M_ACC_1 | 0.0986 | |
7 | 1233 | M_SMA | M_ACC_1 | 0.1508 | |
8 | 1727 | R_vFC_2 | R_vFC_1 | 0.121 | |
9 | 1732 | L_mid_insula_1 | R_vFC_1 | 0.2313 | |
10 | 1795 | M_mFC | R_ant_insula | 0.0542 | |
11 | 1950 | M_mFC | L_ant_insula | 0.1294 | |
12 | 2110 | L_vFC_3 | L_basal_ganglia_1 | 0.1074 | |
13 | 2183 | R_basal_ganglia_1 | M_mFC | 0.0652 | |
14 | 2301 | L_post_cingulate_1 | R_frontal_1 | 0.0016 | |
15 | 2311 | R_precuneus_3 | R_frontal_1 | 0.1118 | |
16 | 2314 | R_post_cingulate | R_frontal_1 | 0.0027 | |
17 | 2315 | L_precuneus_2 | R_frontal_1 | 0.0074 | |
18 | 2441 | R_precuneus_1 | R_dFC_2 | 0.3302 | |
19 | 2509 | L_precuneus_1 | R_dFC_3 | 0.0548 | |
20 | 2511 | R_precuneus_1 | R_dFC_3 | 0.3977 | |
21 | 2542 | M_SMA | L_dFC | 0.1668 | |
22 | 2551 | R_precentral_gyrus_3 | L_dFC | 0.029 | |
23 | 2605 | L_basal_ganglia_2 | L_vFC_2 | 0.2421 | |
24 | 2606 | R_basal_ganglia_1 | L_vFC_2 | 0.1719 | |
25 | 2618 | L_precentral_gyrus_2 | L_vFC_2 | 0.1803 | |
26 | 2884 | L_mid_insula_2 | R_pre_SMA | 0.0787 | |
27 | 2887 | R_mid_insula_2 | R_pre_SMA | 0.0787 | |
28 | 2908 | L_precuneus_1 | R_pre_SMA | 0.112 | |
29 | 2935 | M_SMA | R_vFC_2 | 0.0752 | |
30 | 2989 | R_post_cingulate | R_vFC_2 | 0.0487 | |
31 | 3033 | L_precuneus_1 | M_SMA | 0.1055 | |
32 | 3094 | L_precuneus_1 | R_frontal_2 | 0.0269 | |
33 | 3256 | L_parietal_5 | L_mid_insula_1 | 0.1804 | |
34 | 3277 | R_precuneus_2 | L_mid_insula_1 | 0.0604 | |
35 | 3298 | L_parietal_1 | L_precentral_gyrus_1 | 0.1927 | |
36 | 3328 | L_precuneus_1 | L_precentral_gyrus_1 | 0.0331 | |
37 | 3330 | R_precuneus_1 | L_precentral_gyrus_1 | 0.1669 | |
38 | 3357 | R_precentral_gyrus_3 | L_parietal_1 | 0.1524 | |
39 | 3367 | L_parietal_4 | L_parietal_1 | 0.1008 | |
40 | 3368 | R_parietal_1 | L_parietal_1 | 0.0787 | |
41 | 3376 | R_parietal_3 | L_parietal_1 | 0.021 | |
42 | 3379 | L_parietal_7 | L_parietal_1 | 0.0593 | |
43 | 3546 | L_precuneus_1 | R_precentral_gyrus_3 | 0.0535 | |
44 | 3548 | R_precuneus_1 | R_precentral_gyrus_3 | 0.2019 | |
45 | 3598 | L_precuneus_1 | L_parietal_2 | 0.0234 | |
46 | 3835 | R_parietal_3 | R_mid_insula_2 | 0.2415 | |
47 | 3926 | R_parietal_3 | L_mid_insula_3 | 0.2598 | |
48 | 4021 | L_precuneus_1 | L_parietal_4 | 0.2507 | |
49 | 4061 | L_temporal_2 | R_parietal_1 | 0.1886 | |
50 | 4063 | L_precuneus_1 | R_parietal_1 | 0.0089 | |
51 | 4065 | R_precuneus_1 | R_parietal_1 | 0.1549 | |
52 | 4095 | M_post_cingulate | L_parietal_5 | 0.241 | |
53 | 4104 | L_precuneus_1 | L_parietal_5 | 0.0656 | |
54 | 4249 | M_post_cingulate | R_post_insula | 0.1736 | |
55 | 4299 | L_post_cingulate_1 | R_basal_ganglia_2 | 0.3015 | |
56 | 4311 | L_post_cingulate_2 | R_basal_ganglia_2 | 0.2509 | |
57 | 4334 | L_post_cingulate_1 | M_post_cingulate | 0.3287 | |
58 | 4430 | R_precuneus_1 | L_post_insula | 0.1071 | |
59 | 4518 | L_precuneus_1 | L_post_parietal_1 | 0.1153 | |
60 | 4602 | L_IPL_1 | L_precuneus_1 | 0.1964 | |
61 | 4683 | L_IPL_2 | L_IPL_1 | 0.2273 | |
62 | 4802 | L_IPL_3 | L_parietal_8 | 0.379 | |
63 | 4812 | L_angular_gyrus_2 | L_parietal_8 | 0.0522 |
ROI index | ROI | Weight |
---|---|---|
7 | L_vent_aPFC | 0.1559 |
12 | R_sup_frontal | 0.5193 |
14 | M_ACC_1 | 0.1247 |
15 | L_sup_frontal | 0.2239 |
16 | M_ACC_2 | 0.1236 |
20 | R_vFC_1 | 0.1761 |
21 | R_ant_insula | 0.0271 |
23 | L_ant_insula | 0.0647 |
25 | L_basal_ganglia_1 | 0.0165 |
26 | M_mFC | 0.0229 |
27 | R_frontal_1 | 0.0591 |
29 | R_dFC_2 | 0.1651 |
30 | R_dFC_3 | 0.1714 |
31 | L_dFC | 0.0979 |
32 | L_vFC_2 | 0.1169 |
33 | L_basal_ganglia_2 | 0.1211 |
34 | R_basal_ganglia_1 | 0.1185 |
35 | L_vFC_3 | 0.0537 |
36 | R_pre_SMA | 0.072 |
37 | R_vFC_2 | 0.0473 |
38 | M_SMA | 0.0232 |
39 | R_frontal_2 | 0.0135 |
42 | L_mid_insula_1 | 0.0556 |
43 | L_precentral_gyrus_1 | 0.1963 |
44 | L_parietal_1 | 0.3025 |
46 | L_precentral_gyrus_2 | 0.0901 |
47 | R_precentral_gyrus_3 | 0.1649 |
48 | L_parietal_2 | 0.0117 |
50 | L_mid_insula_2 | 0.0394 |
53 | R_mid_insula_2 | 0.1601 |
55 | L_mid_insula_3 | 0.1299 |
57 | L_parietal_4 | 0.1758 |
58 | R_parietal_1 | 0.0181 |
59 | L_parietal_5 | 0.0631 |
60 | L_precentral_gyrus_3 | 0.1559 |
63 | R_post_insula | 0.0868 |
64 | R_basal_ganglia_2 | 0.2762 |
65 | M_post_cingulate | 0.043 |
66 | R_parietal_3 | 0.2401 |
68 | L_post_insula | 0.0536 |
69 | L_parietal_7 | 0.0296 |
71 | L_post_parietal_1 | 0.0577 |
72 | L_temporal_2 | 0.0943 |
74 | L_precuneus_1 | 0.4059 |
76 | R_precuneus_1 | 0.5722 |
77 | L_IPL_1 | 0.2118 |
79 | L_post_cingulate_1 | 0.0128 |
80 | R_precuneus_2 | 0.0302 |
83 | L_parietal_8 | 0.2156 |
86 | L_IPL_2 | 0.1137 |
88 | L_IPL_3 | 0.1895 |
89 | R_precuneus_3 | 0.0559 |
91 | L_post_cingulate_2 | 0.1255 |
92 | R_post_cingulate | 0.023 |
93 | L_precuneus_2 | 0.0037 |
98 | L_angular_gyrus_2 | 0.0261 |
We employed the same SVM method on gender classification as we did for age classification. A linear SVM classifier was not able to significantly discriminate between male and female subjects (55% accuracy,
Seeing that classification of age groups was successful, we decided to test whether age prediction of individuals is viable on a continuous scale with the use of only fcMRI data. That is, given an fMRI connectivity map, we wanted to determine the age in years of the individual on a continuous range rather than choose between two or three discrete classes. A SVR linear predictor (top features retained = 298, ε = 0.1) was applied to 65 subjects varying in age (19–85 years) and was able to predict subject age with a reasonable degree of accuracy, [
The SVR method had 185 features (out of the 298) present in every fold. These consensus features’ weights and the node weights were computed in the same way as for the SVM classifier (see Figures
Feature index | SVR feature number | ROI 1 | Connected with | ROI 2 | Weight | |
---|---|---|---|---|---|---|
1 | 200 | 6 | R_aPFC_2 | 3 | M_mPFC | 3.4199 |
2 | 208 | 14 | M_ACC_1 | 3 | M_mPFC | 4.0323 |
3 | 302 | 12 | R_sup_frontal | 4 | L_aPFC_2 | 0.1837 |
4 | 308 | 18 | L_vPFC | 4 | L_aPFC_2 | 6.7691 |
5 | 514 | 35 | L_vFC_3 | 6 | R_aPFC_2 | 1.9686 |
6 | 515 | 36 | R_pre_SMA | 6 | R_aPFC_2 | 3.9004 |
7 | 517 | 38 | M_SMA | 6 | R_aPFC_2 | 1.5754 |
8 | 523 | 44 | L_parietal_1 | 6 | R_aPFC_2 | 0.4170 |
9 | 632 | 60 | L_precentral_gyrus_3 | 7 | L_vent_aPFC | 2.8031 |
10 | 785 | 30 | R_dFC_3 | 9 | R_vlPFC | 0.5353 |
11 | 871 | 26 | M_mFC | 10 | R_ACC | 7.0751 |
12 | 881 | 36 | R_pre_SMA | 10 | R_ACC | 7.2001 |
13 | 910 | 65 | M_post_cingulate | 10 | R_ACC | 1.5674 |
14 | 955 | 21 | R_ant_insula | 11 | R_dlPFC_1 | 1.6761 |
15 | 957 | 23 | L_ant_insula | 11 | R_dlPFC_1 | 0.0260 |
16 | 961 | 27 | R_frontal_1 | 11 | R_dlPFC_1 | 4.6158 |
17 | 1037 | 15 | L_sup_frontal | 12 | R_sup_frontal | 11.408 |
18 | 1038 | 16 | M_ACC_2 | 12 | R_sup_frontal | 3.0458 |
19 | 1044 | 22 | R_dACC | 12 | R_sup_frontal | 3.5010 |
20 | 1047 | 25 | L_basal_ganglia_1 | 12 | R_sup_frontal | 6.4695 |
21 | 1048 | 26 | M_mFC | 12 | R_sup_frontal | 4.8864 |
22 | 1139 | 30 | R_dFC_3 | 13 | R_vPFC | 4.3102 |
23 | 1213 | 18 | L_vPFC | 14 | M_ACC_1 | 5.4143 |
24 | 1218 | 23 | L_ant_insula | 14 | M_ACC_1 | 3.4732 |
25 | 1231 | 36 | R_pre_SMA | 14 | M_ACC_1 | 4.6873 |
26 | 1233 | 38 | M_SMA | 14 | M_ACC_1 | 2.9394 |
27 | 1239 | 44 | L_parietal_1 | 14 | M_ACC_1 | 0.6638 |
28 | 1260 | 65 | M_post_cingulate | 14 | M_ACC_1 | 3.6222 |
29 | 1306 | 26 | M_mFC | 15 | L_sup_frontal | 1.8433 |
30 | 1387 | 23 | L_ant_insula | 16 | M_ACC_2 | 1.2731 |
31 | 1398 | 34 | R_basal_ganglia_1 | 16 | M_ACC_2 | 2.0425 |
32 | 1560 | 31 | L_dFC | 18 | L_vPFC | 0.7004 |
33 | 1727 | 37 | R_vFC_2 | 20 | R_vFC_1 | 2.9906 |
34 | 1730 | 40 | R_precentral_gyrus_1 | 20 | R_vFC_1 | 0.8001 |
35 | 1732 | 42 | L_mid_insula_1 | 20 | R_vFC_1 | 3.3381 |
36 | 1739 | 49 | R_mid_insula_1 | 20 | R_vFC_1 | 2.0195 |
37 | 1791 | 22 | R_dACC | 21 | R_ant_insula | 3.8335 |
38 | 1795 | 26 | M_mFC | 21 | R_ant_insula | 3.2795 |
39 | 1870 | 23 | L_ant_insula | 22 | R_dACC | 5.1255 |
40 | 1880 | 33 | L_basal_ganglia_2 | 22 | R_dACC | 0.8944 |
41 | 1881 | 34 | R_basal_ganglia_1 | 22 | R_dACC | 1.3206 |
42 | 1882 | 35 | L_vFC_3 | 22 | R_dACC | 2.0233 |
43 | 1883 | 36 | R_pre_SMA | 22 | R_dACC | 4.7918 |
44 | 1949 | 25 | L_basal_ganglia_1 | 23 | L_ant_insula | 2.8737 |
45 | 1950 | 26 | M_mFC | 23 | L_ant_insula | 1.0806 |
46 | 1954 | 30 | R_dFC_3 | 23 | L_ant_insula | 1.6872 |
47 | 1960 | 36 | R_pre_SMA | 23 | L_ant_insula | 6.0690 |
48 | 2006 | 82 | R_IPL_1 | 23 | L_ant_insula | 2.2798 |
49 | 2016 | 92 | R_post_cingulate | 23 | L_ant_insula | 4.7532 |
50 | 2110 | 35 | L_vFC_3 | 25 | L_basal_ganglia_1 | 3.9079 |
51 | 2113 | 38 | M_SMA | 25 | L_basal_ganglia_1 | 2.0904 |
52 | 2176 | 27 | R_frontal_1 | 26 | M_mFC | 0.2893 |
53 | 2182 | 33 | L_basal_ganglia_2 | 26 | M_mFC | 3.0670 |
54 | 2183 | 34 | R_basal_ganglia_1 | 26 | M_mFC | 0.8171 |
55 | 2184 | 35 | L_vFC_3 | 26 | M_mFC | 3.9006 |
56 | 2190 | 41 | L_thalamus_1 | 26 | M_mFC | 2.0477 |
57 | 2217 | 68 | L_post_insula | 26 | M_mFC | 12.328 |
58 | 2252 | 30 | R_dFC_3 | 27 | R_frontal_1 | 4.8758 |
59 | 2258 | 36 | R_pre_SMA | 27 | R_frontal_1 | 3.5564 |
60 | 2262 | 40 | R_precentral_gyrus_1 | 27 | R_frontal_1 | 2.4206 |
61 | 2267 | 45 | R_precentral_gyrus_2 | 27 | R_frontal_1 | 4.0491 |
62 | 2271 | 49 | R_mid_insula_1 | 27 | R_frontal_1 | 1.7481 |
63 | 2299 | 77 | L_IPL_1 | 27 | R_frontal_1 | 1.6080 |
64 | 2301 | 79 | L_post_cingulate_1 | 27 | R_frontal_1 | 9.5243 |
65 | 2302 | 80 | R_precuneus_2 | 27 | R_frontal_1 | 3.2888 |
66 | 2304 | 82 | R_IPL_1 | 27 | R_frontal_1 | 2.3834 |
67 | 2308 | 86 | L_IPL_2 | 27 | R_frontal_1 | 2.5869 |
68 | 2311 | 89 | R_precuneus_3 | 27 | R_frontal_1 | 1.6131 |
69 | 2313 | 91 | L_post_cingulate_2 | 27 | R_frontal_1 | 4.0015 |
70 | 2314 | 92 | R_post_cingulate | 27 | R_frontal_1 | 3.1354 |
71 | 2315 | 93 | L_precuneus_2 | 27 | R_frontal_1 | 1.4905 |
72 | 2317 | 95 | L_post_cingulate_3 | 27 | R_frontal_1 | 1.2658 |
73 | 2340 | 46 | L_precentral_gyrus_2 | 28 | L_vFC_1 | 2.9832 |
74 | 2343 | 49 | R_mid_insula_1 | 28 | L_vFC_1 | 0.9104 |
75 | 2344 | 50 | L_mid_insula_2 | 28 | L_vFC_1 | 1.5884 |
76 | 2374 | 80 | R_precuneus_2 | 28 | L_vFC_1 | 2.1640 |
77 | 2399 | 34 | R_basal_ganglia_1 | 29 | R_dFC_2 | 1.8317 |
78 | 2439 | 74 | L_precuneus_1 | 29 | R_dFC_2 | 5.2682 |
79 | 2441 | 76 | R_precuneus_1 | 29 | R_dFC_2 | 3.9293 |
80 | 2472 | 37 | R_vFC_2 | 30 | R_dFC_3 | 1.2744 |
81 | 2509 | 74 | L_precuneus_1 | 30 | R_dFC_3 | 3.1864 |
82 | 2511 | 76 | R_precuneus_1 | 30 | R_dFC_3 | 10.158 |
83 | 2540 | 36 | R_pre_SMA | 31 | L_dFC | 0.0347 |
84 | 2542 | 38 | M_SMA | 31 | L_dFC | 4.5939 |
85 | 2551 | 47 | R_precentral_gyrus_3 | 31 | L_dFC | 4.2930 |
86 | 2561 | 57 | L_parietal_4 | 31 | L_dFC | 2.0379 |
87 | 2562 | 58 | R_parietal_1 | 31 | L_dFC | 0.1465 |
88 | 2570 | 66 | R_parietal_3 | 31 | L_dFC | 0.6321 |
89 | 2573 | 69 | L_parietal_7 | 31 | L_dFC | 3.8353 |
90 | 2606 | 34 | R_basal_ganglia_1 | 32 | L_vFC_2 | 1.6084 |
91 | 2617 | 45 | R_precentral_gyrus_2 | 32 | L_vFC_2 | 5.5595 |
92 | 2618 | 46 | L_precentral_gyrus_2 | 32 | L_vFC_2 | 6.3606 |
93 | 2620 | 48 | L_parietal_2 | 32 | L_vFC_2 | 5.3524 |
94 | 2806 | 36 | R_pre_SMA | 35 | L_vFC_3 | 5.9420 |
95 | 2829 | 59 | L_parietal_5 | 35 | L_vFC_3 | 2.5231 |
96 | 2876 | 42 | L_mid_insula_1 | 36 | R_pre_SMA | 2.0429 |
97 | 2884 | 50 | L_mid_insula_2 | 36 | R_pre_SMA | 0.8906 |
98 | 2887 | 53 | R_mid_insula_2 | 36 | R_pre_SMA | 2.8127 |
99 | 2889 | 55 | L_mid_insula_3 | 36 | R_pre_SMA | 0.6196 |
100 | 2908 | 74 | L_precuneus_1 | 36 | R_pre_SMA | 3.0879 |
101 | 2935 | 38 | M_SMA | 37 | R_vFC_2 | 1.3538 |
102 | 2977 | 80 | R_precuneus_2 | 37 | R_vFC_2 | 2.6368 |
103 | 2989 | 92 | R_post_cingulate | 37 | R_vFC_2 | 1.0198 |
104 | 2992 | 95 | L_post_cingulate_3 | 37 | R_vFC_2 | 1.2270 |
105 | 3001 | 42 | L_mid_insula_1 | 38 | M_SMA | 0.0421 |
106 | 3009 | 50 | L_mid_insula_2 | 38 | M_SMA | 2.5175 |
107 | 3012 | 53 | R_mid_insula_2 | 38 | M_SMA | 1.0882 |
108 | 3013 | 54 | R_temporal_1 | 38 | M_SMA | 1.5278 |
109 | 3022 | 63 | R_post_insula | 38 | M_SMA | 0.5381 |
110 | 3033 | 74 | L_precuneus_1 | 38 | M_SMA | 1.6784 |
111 | 3094 | 74 | L_precuneus_1 | 39 | R_frontal_2 | 1.1764 |
112 | 3160 | 80 | R_precuneus_2 | 40 | R_precentral_gyrus_1 | 1.1140 |
113 | 3172 | 92 | R_post_cingulate | 40 | R_precentral_gyrus_1 | 0.7809 |
114 | 3181 | 42 | L_mid_insula_1 | 41 | L_thalamus_1 | 7.0436 |
115 | 3255 | 58 | R_parietal_1 | 42 | L_mid_insula_1 | 3.3164 |
116 | 3256 | 59 | L_parietal_5 | 42 | L_mid_insula_1 | 3.9536 |
117 | 3268 | 71 | L_post_parietal_1 | 42 | L_mid_insula_1 | 1.1748 |
118 | 3274 | 77 | L_IPL_1 | 42 | L_mid_insula_1 | 0.6810 |
119 | 3276 | 79 | L_post_cingulate_1 | 42 | L_mid_insula_1 | 5.5190 |
120 | 3277 | 80 | R_precuneus_2 | 42 | L_mid_insula_1 | 2.2915 |
121 | 3289 | 92 | R_post_cingulate | 42 | L_mid_insula_1 | 2.2737 |
122 | 3298 | 44 | L_parietal_1 | 43 | L_precentral_gyrus_1 | 4.8264 |
123 | 3320 | 66 | R_parietal_3 | 43 | L_precentral_gyrus_1 | 2.3284 |
124 | 3328 | 74 | L_precuneus_1 | 43 | L_precentral_gyrus_1 | 4.3556 |
125 | 3330 | 76 | R_precuneus_1 | 43 | L_precentral_gyrus_1 | 3.8301 |
126 | 3357 | 47 | R_precentral_gyrus_3 | 44 | L_parietal_1 | 1.4419 |
127 | 3363 | 53 | R_mid_insula_2 | 44 | L_parietal_1 | 3.9555 |
128 | 3366 | 56 | L_parietal_3 | 44 | L_parietal_1 | 3.5900 |
129 | 3367 | 57 | L_parietal_4 | 44 | L_parietal_1 | 8.5639 |
130 | 3368 | 58 | R_parietal_1 | 44 | L_parietal_1 | 0.9669 |
131 | 3372 | 62 | R_parietal_2 | 44 | L_parietal_1 | 6.2692 |
132 | 3376 | 66 | R_parietal_3 | 44 | L_parietal_1 | 2.2942 |
133 | 3377 | 67 | L_parietal_6 | 44 | L_parietal_1 | 4.2380 |
134 | 3379 | 69 | L_parietal_7 | 44 | L_parietal_1 | 3.3560 |
135 | 3386 | 76 | R_precuneus_1 | 44 | L_parietal_1 | 2.3440 |
136 | 3521 | 49 | R_mid_insula_1 | 47 | R_precentral_gyrus_3 | 1.7159 |
137 | 3537 | 65 | M_post_cingulate | 47 | R_precentral_gyrus_3 | 1.2226 |
138 | 3542 | 70 | R_temporal_2 | 47 | R_precentral_gyrus_3 | 2.9690 |
139 | 3546 | 74 | L_precuneus_1 | 47 | R_precentral_gyrus_3 | 2.8436 |
140 | 3548 | 76 | R_precuneus_1 | 47 | R_precentral_gyrus_3 | 1.9436 |
141 | 3553 | 81 | R_temporal_3 | 47 | R_precentral_gyrus_3 | 0.1955 |
142 | 3598 | 74 | L_precuneus_1 | 48 | L_parietal_2 | 2.8843 |
143 | 3600 | 76 | R_precuneus_1 | 48 | L_parietal_2 | 5.3379 |
144 | 3633 | 58 | R_parietal_1 | 49 | R_mid_insula_1 | 3.0962 |
145 | 3634 | 59 | L_parietal_5 | 49 | R_mid_insula_1 | 1.2744 |
146 | 3683 | 58 | R_parietal_1 | 50 | L_mid_insula_2 | 0.4134 |
147 | 3684 | 59 | L_parietal_5 | 50 | L_mid_insula_2 | 6.6085 |
148 | 3690 | 65 | M_post_cingulate | 50 | L_mid_insula_2 | 0.1353 |
149 | 3705 | 80 | R_precuneus_2 | 50 | L_mid_insula_2 | 4.0167 |
150 | 3835 | 66 | R_parietal_3 | 53 | R_mid_insula_2 | 7.7145 |
151 | 3926 | 66 | R_parietal_3 | 55 | L_mid_insula_3 | 0.6183 |
152 | 3973 | 69 | L_parietal_7 | 56 | L_parietal_3 | 0.6484 |
153 | 4021 | 74 | L_precuneus_1 | 57 | L_parietal_4 | 5.2334 |
154 | 4061 | 72 | L_temporal_2 | 58 | R_parietal_1 | 2.2855 |
155 | 4062 | 73 | L_temporal_3 | 58 | R_parietal_1 | 1.2126 |
156 | 4063 | 74 | L_precuneus_1 | 58 | R_parietal_1 | 3.3874 |
157 | 4065 | 76 | R_precuneus_1 | 58 | R_parietal_1 | 0.0311 |
158 | 4095 | 65 | M_post_cingulate | 59 | L_parietal_5 | 1.0257 |
159 | 4104 | 74 | L_precuneus_1 | 59 | L_parietal_5 | 6.1957 |
160 | 4249 | 65 | M_post_cingulate | 63 | R_post_insula | 4.0141 |
161 | 4253 | 69 | L_parietal_7 | 63 | R_post_insula | 1.0645 |
162 | 4255 | 71 | L_post_parietal_1 | 63 | R_post_insula | 0.5269 |
163 | 4264 | 80 | R_precuneus_2 | 63 | R_post_insula | 3.5363 |
164 | 4286 | 66 | R_parietal_3 | 64 | R_basal_ganglia_2 | 2.2027 |
165 | 4299 | 79 | L_post_cingulate_1 | 64 | R_basal_ganglia_2 | 1.9985 |
166 | 4311 | 91 | L_post_cingulate_2 | 64 | R_basal_ganglia_2 | 2.8954 |
167 | 4334 | 79 | L_post_cingulate_1 | 65 | M_post_cingulate | 11.855 |
168 | 4335 | 80 | R_precuneus_2 | 65 | M_post_cingulate | 0.2271 |
169 | 4400 | 78 | R_parietal_4 | 67 | L_parietal_6 | 1.0522 |
170 | 4430 | 76 | R_precuneus_1 | 68 | L_post_insula | 1.6719 |
171 | 4441 | 87 | L_angular_gyrus_1 | 68 | L_post_insula | 1.6202 |
172 | 4516 | 72 | L_temporal_2 | 71 | L_post_parietal_1 | 2.2346 |
173 | 4518 | 74 | L_precuneus_1 | 71 | L_post_parietal_1 | 1.3602 |
174 | 4521 | 77 | L_IPL_1 | 71 | L_post_parietal_1 | 0.1952 |
175 | 4530 | 86 | L_IPL_2 | 71 | L_post_parietal_1 | 2.9293 |
176 | 4552 | 80 | R_precuneus_2 | 72 | L_temporal_2 | 1.2155 |
177 | 4602 | 77 | L_IPL_1 | 74 | L_precuneus_1 | 3.9390 |
178 | 4609 | 84 | L_post_parietal_2 | 74 | L_precuneus_1 | 3.9741 |
179 | 4651 | 77 | L_IPL_1 | 76 | R_precuneus_1 | 0.8857 |
180 | 4683 | 86 | L_IPL_2 | 77 | L_IPL_1 | 0.2140 |
181 | 4686 | 89 | R_precuneus_3 | 77 | L_IPL_1 | 3.7542 |
182 | 4759 | 99 | R_precuneus_4 | 80 | R_precuneus_2 | 1.0681 |
183 | 4802 | 88 | L_IPL_3 | 83 | L_parietal_8 | 8.7927 |
184 | 4812 | 98 | L_angular_gyrus_2 | 83 | L_parietal_8 | 2.3099 |
185 | 4814 | 100 | L_IPS_2 | 83 | L_parietal_8 | 1.5265 |
ROI index | ROI | Weight |
---|---|---|
3 | M_mPFC | 0.3062 |
4 | L_aPFC_2 | 3.4764 |
6 | R_aPFC_2 | 1.7403 |
7 | L_vent_aPFC | 1.4016 |
9 | R_vlPFC | 0.2676 |
10 | R_ACC | 0.8462 |
11 | R_dlPFC_1 | 3.1330 |
12 | R_sup_frontal | 14.747 |
13 | R_vPFC | 2.1551 |
14 | M_ACC_1 | 1.7177 |
15 | L_sup_frontal | 6.6258 |
16 | M_ACC_2 | 3.1807 |
18 | L_vPFC | 5.7415 |
20 | R_vFC_1 | 2.5547 |
21 | R_ant_insula | 4.3946 |
22 | R_dACC | 0.0952 |
23 | L_ant_insula | 10.848 |
25 | L_basal_ganglia_1 | 3.7628 |
26 | M_mFC | 2.6519 |
27 | R_frontal_1 | 4.1057 |
28 | L_vFC_1 | 1.3242 |
29 | R_dFC_2 | 3.6829 |
30 | R_dFC_3 | 0.6411 |
31 | L_dFC | 3.3619 |
32 | L_vFC_2 | 1.4715 |
33 | L_basal_ganglia_2 | 1.0863 |
34 | R_basal_ganglia_1 | 0.9506 |
35 | L_vFC_3 | 1.7332 |
36 | R_pre_SMA | 0.6284 |
37 | R_vFC_2 | 1.6506 |
38 | M_SMA | 2.2000 |
39 | R_frontal_2 | 0.5882 |
40 | R_precentral_gyrus_1 | 0.1372 |
41 | L_thalamus_1 | 2.4979 |
42 | L_mid_insula_1 | 2.1402 |
43 | L_precentral_gyrus_1 | 0.9863 |
44 | L_parietal_1 | 5.0775 |
45 | R_precentral_gyrus_2 | 4.8043 |
46 | L_precentral_gyrus_2 | 4.6719 |
47 | R_precentral_gyrus_3 | 1.8094 |
48 | L_parietal_2 | 3.9030 |
49 | R_mid_insula_1 | 2.0883 |
50 | L_mid_insula_2 | 6.3615 |
53 | R_mid_insula_2 | 0.0710 |
54 | R_temporal_1 | 0.7639 |
55 | L_mid_insula_3 | 0.6189 |
56 | L_parietal_3 | 2.1192 |
57 | L_parietal_4 | 5.8797 |
58 | R_parietal_1 | 2.3834 |
59 | L_parietal_5 | 9.7648 |
60 | L_precentral_gyrus_3 | 1.4016 |
62 | R_parietal_2 | 3.1346 |
63 | R_post_insula | 0.8258 |
64 | R_basal_ganglia_2 | 1.3456 |
65 | M_post_cingulate | 6.4323 |
66 | R_parietal_3 | 5.6008 |
67 | L_parietal_6 | 1.5929 |
68 | L_post_insula | 6.1384 |
69 | L_parietal_7 | 3.8037 |
70 | R_temporal_2 | 1.4845 |
71 | L_post_parietal_1 | 1.9759 |
72 | L_temporal_2 | 2.8678 |
73 | L_temporal_3 | 0.6063 |
74 | L_precuneus_1 | 5.8246 |
76 | R_precuneus_1 | 15.035 |
77 | L_IPL_1 | 4.7624 |
78 | R_parietal_4 | 0.5261 |
79 | L_post_cingulate_1 | 2.9254 |
80 | R_precuneus_2 | 2.2972 |
81 | R_temporal_3 | 0.0978 |
82 | R_IPL_1 | 0.0518 |
83 | L_parietal_8 | 4.7881 |
84 | L_post_parietal_2 | 1.9870 |
86 | L_IPL_2 | 2.6511 |
87 | L_angular_gyrus_1 | 0.8101 |
88 | L_IPL_3 | 4.3964 |
89 | R_precuneus_3 | 2.6837 |
91 | L_post_cingulate_2 | 0.5530 |
92 | R_post_cingulate | 0.4474 |
93 | L_precuneus_2 | 0.7453 |
95 | L_post_cingulate_3 | 1.2464 |
98 | L_angular_gyrus_2 | 1.1549 |
99 | R_precuneus_4 | 0.5340 |
100 | L_IPS_2 | 0.7632 |
To check for agreement with previous studies (see Dosenbach et al.,
ROI index | ROI |
---|---|
23 | L_ant_insula |
26 | M_mFC |
44 | L_parietal_1 |
66 | R_parietal_3 |
69 | L_parietal_7 |
74 | L_precuneus_1 |
77 | L_IPL_1 |
However, we use the linear SVR predictor for feature and node significance output since weights extracted from the linear SVR have a direct proportionality between absolute weight and significance in variable prediction. The same cannot be said about the RBF SVR weights, which are not as readily interpreted.
In the present study, we examined the ability of a SVM to classify individuals as either young or old, and to predict age solely on their rs-fMRI data. Our aim was to improve the discriminatory ability and accuracy of the multivariate vector machine method by parameter tuning and feature selection and also output interpretable discriminating features.
Support vector machine classification (using temporal correlations between ROIs as input features) of individuals as either children or adults was found 91% accurate in a study by Dosenbach et al. (
Although age classification was very significant (
We found that the SVM method predicted subject age on a continuous scale with relatively good performance. A perfect predictor has a linear regression fit of
From the linear regression plot (Figure
The SVM method allows for detection of the most influential features and nodes which drive the classifier or predictor. We utilized this approach to find the “connectivity hubs,” or nodes with the most significant features that influenced age classification. Tables
Of note is the difference in distributions of the node weights for the linear SVM and linear SVR methods (Figures
The improvement in accuracy due to the reduction of the dimension of the feature space, in general, reveals that the classification performance is related to the number of features used and the “quality” of the features used. Our work, using the
The growing number of imaging-based binary classification studies of clinical populations (autism, schizophrenia, depression, and attention-deficit hyperactivity disorder) suggests that this is a promising approach for distinguishing disease states from healthy brains on the basis of measurable differences in spontaneous activity (Shen et al.,
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
This work was supported by the National Research Service Award (NRSA) T32 EB011434 to Svyatoslav Vergun, University of Wisconsin Institute for Clinical and Translational Research National Institutes of Health (UW ICTR NIH)/UL1RR025011 Pilot Grant from the Clinical and Translational Science Award (CTSA) program of the National Center for Research Resources (NCRR) and KL2 Scholar Award to Vivek Prabhakaran, and RC1MH090912 National Institutes of Health-National Institute of Mental Health (NIH-NIMH) ARRA Challenge Grant to Elizabeth Meyerand. We are thankful to the 1000 Functional Connectome Project for their data set.
1
2
3
4
5