Normative modeling#

Spectral descriptors become clinically useful when you can say how unusual a given brain is. Normative modeling answers that by learning the expected distribution of a descriptor across a healthy reference cohort, then scoring each new subject against it.

Why normative, not matched controls#

Matched control groups force a binary, case-vs-control framing and carry a control-selection confound. A normative model instead gives a dimensional z-score per subject (and per vertex or region), models its uncertainty explicitly, and removes the need to hand-pick a matched group:

\[ z(x) = \frac{\, y(x) - \hat{\mu}(x)\,}{\hat{\sigma}(x)} , \]

where \(\hat\mu\) and \(\hat\sigma\) are the normative mean and dispersion estimated from the reference cohort at location \(x\). A subject’s deviation map is then a continuous field of how far each location departs from normative expectation.

In SpectralBrain#

The statistics.normative module estimates the reference model and produces z-deviation maps for new subjects; pair it with the harmonization tools (ComBat / ComBat-GAM) when the reference cohort spans multiple sites, so that deviations reflect biology rather than scanner.

from spectralbrain.statistics import normative
# fit on a reference cohort, then z-score new subjects against it