Table 3.
Behavior prediction based on the significant clusters of mass-univariate VLSM, SVR-LSM, and SVR-MLSM.
| Model | Independent variables | Baseline MoCA | Year 1 MoCA | ||
|---|---|---|---|---|---|
| Accuracy | p | Accuracy | p | ||
| 1 | Age, gender, education year | 0.5897 | 2.08E-08 | 0.6075 | 5.94E-09 |
| 2 | Model 1 + Total infarct volume | 0.6219 | 2.02E-09 | 0.6999 | 1.99E-12 |
| 3 | Model 2 + Total WMH volume | 0.6020 | 8.83E-09 | 0.7108 | 6.33E-13 |
| 4 | Model 2 + VLSM SVOI-AIL (noVol) | 0.7173 | 3.10E-13 | 0.7208 | 2.10E-13 |
| 5 | Model 2 + VLSM SVOI-AIL (totalVol) | 0.6214 | 2.11E-09 | 0.7000 | 1.98E-12 |
| 6 | Model 2 + SVR-LSM SVOI-AIL (noVol) | 0.7328 | 5.31E-14 | 0.7067 | 9.82E-13 |
| 7 | Model 2 + SVR-LSM SVOI-AIL (voxelwise) | 0.7264 | 1.12E-13 | 0.7428 | 1.57E-14 |
| 8 | Model 2 + SVR-LSM SVOI-AIL (totalVol) | 0.6779 | 1.73E-11 | 0.7121 | 5.49E-13 |
| 9 | Model 2 + SVR-LSM SVOI-AIL (nonlinear) | 0.7009 | 1.80E-12 | 0.7355 | 3.81E-14 |
| 10 | Model 3 + MLSM SVOI-AIL (noVol) | 0.7026 | 1.50E-12 | 0.7290 | 8.25E-14 |
| 11 | Model 3 + MLSM SVOI-AIL (voxelwise) | 0.7074 | 9.07E-13 | 0.7593 | 1.89E-15 |
| 12 | Model 3 + MLSM SVOI-AIL (totalVol) | 0.7086 | 8.01E-13 | 0.7622 | 1.27E-15 |
| 13 | Model 3 + MLSM SVOI-WMH (noVol) | 0.7935 | 1.26E-17 | 0.8140 | 3.91E-19 |
| 14 | Model 3 + MLSM SVOI-WMH (voxelwise) | 0.8467 | 5.71E-22 | 0.8525 | 1.52E-22 |
| 15 | Model 3 + MLSM SVOI-WMH (totalVol) | 0.7906 | 2.00E-17 | 0.8730 | 8.89E-25 |
| 16 | Model 3 + MLSM SVOI-AIL + MLSM SVOI-WMH (noVol) | 0.8237 | 6.58E-20 | 0.8443 | 9.70E-22 |
| 17 | Model 3 + MLSM SVOI-AIL + MLSM SVOI-WMH (voxelwise) | 0.8600 | 2.56E-23 | 0.8826 | 5.74E-26 |
| 18 | Model 3 + MLSM SVOI-AIL + MLSM SVOI-WMH (totalVol) | 0.8112 | 6.42E-19 | 0.8750 | 5.10E-25 |
The prediction performance was evaluated using support vector regression through leave-one-out cross-validation. The prediction accuracy was calculated as the Pearson correlation coefficient of the real MoCA score and the predicted MoCA score, and the corresponding p-value was also provided. SVOI, significant clusters-based volume of interest; noVol, without volume control; voxelwise, voxelwise normalization by weighting each voxel with inverse proportion to the square root of the corresponding lesion size; totalVol, volume control by regressing out the total lesion size from baseline or year 1 MoCA.