Fig. 7

Braak NFT stage prediction results using imaging features. a Prediction results on the Emory-Train cohort when compared against the Braak NFT stage assigned during initial neuropathology autopsy, 52 cases. b Prediction results on the UC Davis cohort, 23 cases. Green boxes with hatches are used to highlight the diagonal. The weighted Cohen’s kappa is shown in the title. c Heatmap of weighted Cohen’s kappa for the Emory-Train cohort between pairs of expert raters and the ML model. The average and standard deviation of all Cohen’s kappas is shown in the title. d Top 10 most important features for predicting Braak stages. The random forest classifier reports the feature importance, with the feature value (x-axis) being a normalized value where the sum of all feature importances equals 1. E: expert, ML: random forest ML classifier, k: weighted Cohen’s kappa, r: radius used when calculating the average clustering coefficient, FOV: field of view (see methods, ML Braak NFT Staging section), coef: coefficient. Predicting Braak NFT Stages with Imaging Features and ML