Figure 5.

Paired evaluation reveals that models trained to recognize disease stage are not confounded by age

(A) The distribution of age of death (AOD) for patients who were diagnosed with mild, moderate, or severe AD during postmortem pathology analysis.

(B) Schematic representation of rankable pairs, selected to be either confounder matched (red) or mismatched (black). Each patient is represented by a square, colored according to the corresponding pathology annotation. The value of AOD is censored at 90 years of age in the dataset.

(C) 2 × 2 contingency tables showing the correctly and incorrectly ranked test pairs for AOD-confounded and AOD-matched scenarios. The p value was computed using a one-sided Fisher’s exact test with the alternative hypothesis being that AOD-matched pairs were more likely to be misranked by the model.

See also Figure S8.