run_benchmark

run_benchmark(
    result,
    *,
    condition_col='condition',
    control='HC',
    dataset_col=None,
    age_col='age',
    alpha=0.05,
)

The AA1 and AA2 tests, and the score that combines them.

AA2 – for a dataset with controls: is the condition group’s age acceleration higher than its own controls’? One-sided Mann-Whitney, BH corrected across datasets.

AA1 – for a dataset without controls: is the condition group’s acceleration above zero? One-sided Wilcoxon signed-rank.

MedAE and MedE – median absolute error and median signed error against chronological age, on healthy controls only. MedE is the bias, and it discounts the AA1 credit in the total:

.. code-block:: text

total = AA2 + AA1 * (1 - max(0, MedE) / MedAE)

Without that discount a clock that simply over-predicts everybody sweeps AA1, because every group looks accelerated when the baseline is wrong. This is the correction ComputAgeBench introduced and it is the reason the benchmark ranks differently from a plain “does it separate cases” tally.

Median absolute error against chronological age is reported but never ranked on. A perfect chronological oracle would score zero here and be useless – it would have no age acceleration to detect anything with.