Table 9.
Site prediction results.
| Meta-dataset and feature type | Balanced accuracy median (IQR) | ||
|---|---|---|---|
| Without harmonization | Harmonization with neuroHarmonize | Harmonization with harmonizer transformer | |
| CHILDHOOD | |||
| CT | 0.45 (0.03) | 0.09 (0.01) | 0.13 (0.02) |
| FD | 0.35 (0.02) | 0.09 (0.01) | 0.13 (0.02) |
| ADOLESCENCE | |||
| CT | 0.45 (0.03) | 0.13 (0.02) | 0.15 (0.03) |
| FD | 0.43 (0.04) | 0.16 (0.03) | 0.22 (0.04) |
| ADULTHOOD | |||
| CT | 0.40 (0.03) | 0.09 (0.01) | 0.09 (0.01) |
| FD | 0.29 (0.02) | 0.12 (0.01) | 0.13 (0.01) |
| LIFESPAN | |||
| CT | 0.28 (0.01) | 0.06 (0.01) | 0.07 (0.01) |
| FD | 0.22 (0.01) | 0.08 (0.01) | 0.10 (0.01) |
The median and the interquartile range of the balanced accuracy over 100 repetitions of the 5-fold CV have been reported. In bold, we have highlighted significant falsely overestimated performance due to data leakage (the median balanced accuracy in imaging site prediction using data harmonized with neuroHarmonize is lower, i.e., better performance, than that estimated using data harmonized with the harmonizer transformer within the CV – one-sided Wilcoxon signed-rank test p-values < 0.001 for all the analyses).
CT: cortical thickness; CV: cross-validation; FD: fractal dimension; IQR: interquartile range.