. 2020 Dec 15;11:593336. doi: 10.3389/fpsyt.2020.593336

Table 3.

Mean absolute error (standard error) for the best linear unbiased predictor (BLUP) and support vector machine (SVM) models trained on gray matter maps for each fold.

	BLUP-mean	BLUP-quantiles	SVM	Ensemble learning
Fold 1	4.51 (0.16)^†^*	3.91 (0.14)^†	4.64 (0.17)^†^*	3.39 (0.13)
Fold 2	4.45 (0.16)^†^*	4.06 (0.15)^†	4.75 (0.16)^†^*	3.46 (0.13)
Fold 3	4.67 (0.17)^*	4.02 (0.16)	4.62 (0.17)^*	3.26 (0.13)
Fold 4	4.59 (0.16)^*	4.16 (0.16)^†	4.52 (0.16)^*	3.55 (0.14)
Fold 5	4.86 (0.18)^*	4.21 (0.17)	4.78 (0.17)^*	3.35 (0.14)
5-fold MAE	4.61	4.07	4.66	3.42

The standard error [SE = SD/sqrt(N)] reflects the uncertainty around the mean absolute error (MAE) estimate. A 95% confidence interval may be calculated as MAE ± 1.96 *SE, though it (falsely) assumes normality of the absolute error distribution. For the 5-fold combined MAE, we did not report the SE, as it is notoriously biased downward (54) due to the overlap of the different training/test samples.

^†

Algorithm trained on gray matter maps performs significantly better than the same algorithm trained on surface-based vertices (p <0.05/15).

Algorithm trained on gray matter maps performs significantly worse than Inception V1 (p < 0.05/15). Ensemble learning was performed using linear regression and included the seven algorithms considered in Tables 1, 2, in addition to the three introduced in this section.