Fig. 6: Model alignment with postmortem findings.
From: AI-driven fusion of multimodal data for Alzheimer’s disease biomarker assessment
The swarm and box plots display predicted probabilities of amyloid-beta positivity, P(Aβ), and meta-temporal tau positivity, P(τ), with respect to various AD neuropathological grades in the ADNI (n = 41) and NACC (n = 147) neuropathological validation cohorts. a Kruskal-Wallis tests revealed significant differences in model-predicted probabilities across Thal phases for amyloid plaques for both P(Aβ) (H = 48.32, p = 3.05 × 10−9) and P(τ) (H = 42.02, p = 5.82 × 10−8). b With respect to Braak stage for neurofibrillary degeneration, Kruskal-Wallis tests also showed significant differences in P(Aβ) (H = 54.81, p = 5.05 × 10−10) and P(τ) (H = 54.05, p = 7.19 × 10−10). c Model probabilities were again significantly different across CERAD scores for density of neocortical neuritic plaque for P(Aβ) (H = 52.18, p = 2.74 × 10−11), and P(τ) (H = 50.37, p = 6.68 × 10−11). d For cerebral amyloid angiopathy, Kruskal-Wallis tests yielded significant differences in model-derived probabilities across pathology burden for both P(Aβ) (H = 26.46, p = 7.62 × 10−6) and P(τ) (H = 25.36, p = 1.30 × 10−5). e Finally, CERAD scores for diffuse plaques were also significantly associated with model probabilities: H = 37.84, p = 3.05 × 10−8 for P(Aβ) and H = 29.61, p = 1.66 × 10−6 for P(τ). Pairwise statistical annotations denote the results of two-sided post hoc Dunn tests with Holm-Bonferroni corrections following the Kruskal-Wallis test, with significance levels denoted as * for p < 0.05; ** for p < 0.01; *** for p < 0.001; and **** for p < 0.0001. In addition, trend lines and text boxes in the bottom right of each subplot indicate the Spearman correlation coefficient ρ and associated two-sided p-value for the overall strength of the correlation between model probabilities and neuropathological grades. Each boxplot includes a box presenting the median value and interquartile range (IQR), with whiskers extending from the box to the maxima and minima no further than a distance of 1.5 times the IQR. Detailed statistics regarding median values and IQRs can be found in Supplementary Fig. S21. Additional statistics and p-values for Spearman correlation and Kruskal-Wallis tests can be found in Supplementary Table S22.
