Predicting MCI conversion with CSF, MRI and PET together

The question is simple. If you already have a basic conversion model, how much better does it get when you add each kind of biomarker in turn? The answer is measured with Harrell's C, the share of participant pairs the model orders correctly by time to conversion. A C of 0.5 is chance and 1.0 is perfect. This page builds on the single-marker ranking in From MCI to Alzheimer's.

Harrell's C for the conversion model as CSF, MRI and amyloid PET are added one at a time. The final bar is the full multimodal model.

• 0.669Base model
• 0.785+ CSF markers
• 0.827+ MRI
• 0.837+ amyloid PET

• 0.837Harrell's C, held-out (95% CI 0.749–0.914)
• 0.890AUC, held-out (95% CI 0.792–0.967)

Held-out means the model was scored on participants it was not fitted to. These two figures are the ones to quote.

1. Base model — 0.669
2. + CSF markers — 0.785
3. + MRI — 0.827
4. + amyloid PET — 0.837

The base model starts at 0.669. CSF markers take the biggest step, to 0.785. MRI adds a smaller step, to 0.827, and amyloid PET a smaller one still, to 0.837. The order matters: each modality is judged on what it adds after the ones before it, so PET looks small partly because CSF already carries much of the amyloid signal.

For comparison, the best single baseline marker on the conversion page is CSF Aβ42 on its own. Combining modalities is what moves a model from fair to good.

← Mild cognitive impairment: the research library

• Related ADNI analyses: conversion rates and single markers · what baseline cognition adds · how the measures correlate