P2‐023: Visualization of predictive statistical models of biomarker data from ADNI

Abstract

Measurement of Alzheimer's disease (AD) biomarkers has been challenging because existing CSF assays suffer from inter-lot variability, interferences, and matrix intolerance. This report describes the construction and visualization of predictive statistical models of the ADNI database, including several co-variates beyond the three measured CSF markers of Abeta42, pTau, tTau. The ADNI data table of demographic data, CSF, imaging and cognitive results was downloaded and transferred into JMP Pro v9 statistical discovery software from SAS. Using the new predictive statistical models in JMP Pro, including tree methods (eg, Random Forests, boosting), discriminant analysis, Neural Nets and stepwise linear regression, several predictive models were built and assessed using K-Fold internal cross-validation. The Profiler tool in JMP Pro was successfully used to create dynamic visualizations of various models in a Flash module format for embedding into a website or PowerPoint slide deck. Models were built with three tree based methods (Random Forest, Decision Tree, and boosted Tree), Discriminant analysis, 2 layer Neural Networks, and stepwise linear regression methods. Signature accuracies from near chance (50%) to reasonable clinical sensitivity and specificity (75%/75%) were observed after 10-Fold internal cross-validation. We believe the Profiler tool has enabled visualization of the age dependent inter-relationship between biomarker candidates for the first time. 2x2 “truth tables” and model results will be presented along with a dynamic visualization of model parameters. JMP Pro's ability to visualize predictive statistical models and its Profiler tool enables a more intuitive approach to understanding the relationships between various co-variates and biomarker factors. Flash compatible Profiler modules can be embedded into PPT slides or presented on a website. The results enable a more deep understanding of the inter-relationship between variables.