Fig. 2: Model performance in predicting amyloid and tau positivity.

a, b Receiver operating characteristic (ROC) and precision-recall (PR) curves for Aβ and meta-τ predictions are shown. The area under the ROC curve (AUC) and the average precision (AP) values for Aβ and meta-τ are displayed in the legends, respectively. c Heatmap presenting the AUROC and AP values for Aβ and meta-τ predictions using various combinations of clinical features, starting with person-level history alone and incrementally adding features such as MRI, neuropsychological battery, and plasma data. d Heatmap displaying the AUROC and AP values for Aβ and meta-τ predictions when specific feature sets are removed from the full model. Each row represents the model performance after excluding one feature set, showing how the absence of that data type impacts prediction accuracy. e, f ROC and PR curves showing micro-average, macro-average, and weighted-average calculations based on the regional τ labels. A portion of the NACC dataset used for internal testing, along with data from the ADNI and HABS cohorts for external validation, contributed to generating these results. In panels c and d, FAQ stands for functional activities questionnaire, and CDR stands for clinical dementia ratings.