Table 2.
Results of prediction models based on absolute gray matter volume for the PCA-based approach (first row) and for the atlas-informed feature construction method (second row)
| Network size | MSE | pperm | Range | MAE | RMSE | r | |
|---|---|---|---|---|---|---|---|
| Global model |
556,694 400 |
183 196 |
< 0.001* < 0.001* |
136–223 158–218 |
10.77 11.35 |
13.50 14.00 |
0.24 0.30 |
| Local models | |||||||
| Visual network |
52,753 61 |
202 203 |
.061 < 0.001* |
160–250 164–224 |
11.74 11.45 |
14.18 14.22 |
0.19 0.21 |
| Somatomotor network |
46,282 77 |
245 203 |
0.836 < 0.001* |
150–331 167–225 |
12.46 11.45 |
15.52 14.22 |
0.10 0.19 |
| Dorsal attention network |
36,374 46 |
226 203 |
0.446 < 0.001* |
165–327 163–234 |
12.43 11.43 |
15.00 14.22 |
0.13 0.28 |
| Ventral attention network |
32,345 47 |
206 199 |
0.018 < 0.001* |
142–286 162–220 |
11.72 11.36 |
14.25 14.10 |
0.22 0.22 |
| Limbic network |
27,296 26 |
236 205 |
0.490 < 0.001* |
182–277 174–226 |
12.49 11.46 |
15.33 14.30 |
0.08 0.29 |
| Fronto-parietal network |
45,921 52 |
196 206 |
0.007 0.002* |
153–291 158–233 |
11.28 11.52 |
13.96 14.31 |
0.20 0.27 |
| Default-mode network |
71,492 91 |
210 199 |
0.068 < 0.001* |
171–266 163–228 |
11.92 11.36 |
14.48 14.09 |
0.21 0.29 |
| Subcortical network |
20,361 – |
225 – |
0.267 – |
151–268 – |
12.14 – |
14.95 – |
0.16 – |
| Cerebellum |
57,851 – |
210 – |
0.054 – |
149–298 – |
11.86 – |
14.38 – |
0.15 – |
Network size is depicted in number of voxels for the PCA-based approach and in number of parcels for the atlas-based feature construction method. Note that in the PCA-based approach the number of features was independent from network size, i.e., features were always 277/278 principal components, whereas in the atlas-based approach the number of features corresponds to the number of parcels, i.e., the network size. Results indicating statistical significance are marked with an asterisk (Bonferroni-corrected for multiple comparisons). MSE mean squared error, pperm p value of statistical significance computed by non-parametric permutation test, range of MSE values resulting from different cross-validation folds, MAE mean absolute error in IQ-points, RMSE root mean squared error in IQ-points, r Pearson’s correlation coefficients between predicted and observed Full-Scale Intelligence Quotient (FSIQ) score. All model fit indices were calculated for each cross-validation fold separately and averaged across folds afterwards