TABLE 2.
Biomarker model performance for predicting conversion from MCI to Alzheimer's
| Model | Predictors | Sensitivity | Specificity | Balanced accuracy | AUC | T‐test log likelihood, P |
|---|---|---|---|---|---|---|
| Individual biomarkers | ||||||
| Amyloid‐β42‐CSF | 0.625 | 0.895 | 0.760 | 0.749 | −51.10 | |
| Total tau | 0.708 | 0.625 | 0.671 | 0.730 | −58.88 | |
| P‐tau | 0.729 | 0.666 | 0.697 | 0.722 | −58.57 | |
| Amyloid‐β‐PET | 0.666 | 0.666 | 0.666 | 0.729 | −58.09 | |
| Hippocampal volume | 0.666 | 0.666 | 0.666 | 0.739 | −57.84 | |
| Brain‐PAD | 0.645 | 0.604 | 0.625 | 0.637 | −63.361 | |
| Main effects | ||||||
| 1 | Hippocampal volume, brain‐PAD, amyloid‐β42 | 0.712 | 0.842 | 0.777 | 0.866 | −39.47 |
| 2 | Hippocampal volume, brain‐PAD, amyloid‐β‐PET | 0.756 | 0.754 | 0.755 | 0.830 | −45.44 |
| 3 | Hippocampal volume, amyloid‐β‐PET, P‐tau | 0.770 | 0.762 | 0.766 | 0.830 | −45.44 |
| Full model | 0.770 | 0.729 | 0.760 | 0.838 | −46.58 | |
| Two‐way interactions | ||||||
| 1 | Brain‐PAD * amyloid‐β42 | 0.708 | 0.848 | 0.778 | 0.864 | −39.45, 0.797 |
| Amyloid‐β42 * hippocampal volume | 0.713 | 0.846 | 0.780 | 0.865 | −39.08, <0.001 | |
| 2 | Brain‐PAD * amyloid‐β‐PET | 0.759 | 0.749 | 0.754 | 0.826 | −45.39, 0.46 |
| Hippocampal volume * amyloid‐β‐PET | 0.723 | 0.788 | 0.756 | 0.830 | −44.60, <0.001 | |
| 3 | Amyloid‐β‐PET * P‐tau | 0.771 | 0.756 | 0.764 | 0.826 | −45.26, 0.015 |
| Hippocampal volume * amyloid‐β‐PET | 0.733 | 0.806 | 0.769 | 0.840 | −43.02, <0.001 | |
| P‐tau * hippocampal volume | 0.758 | 0.768 | 0.763 | 0.826 | −45.114, 0.003 | |
| Three‐way interactions | ||||||
| 1 | Hippocampal volume * brain‐PAD * amyloid‐β42 | 0.724 | 0.830 | 0.777 | 0.861 | −39.422, 0.745 |
| 2 | Hippocampal volume * brain‐PAD * amyloid‐β‐PET | 0.738 | 0.791 | 0.764 | 0.827 | −44.712, <0.001 |
| 3 | Hippocampal volume * amyloid‐β‐PET * P‐tau | 0.748 | 0.812 | 0.780 | 0.835 | −43.232, <0.001 |
Note: T‐tests (paired) based on comparison between accuracy and log likelihood scores on 100 bootstrapped sets are provided as a mean to assess the relative increase in generalisation attributable to the introduction of either two‐way interactions or three‐way interactions between biomarkers in comparison to main effects only variants for two‐way interactions, respectively main effects plus a summation of pairwise two‐way interactions terms between the three variables in question for three‐way interactions.
Abbreviation: AUC, area under receiver operator characteristic curve.