may yield important insights into genetic influences on the biology of Alzheimer’s disease (AD).We investigated the accuracy of a novel unsuper- vised multimodal biomarker classifier for differentiating cognitively normal elderly (NC) from subjects with amnestic mild cognitive impairment

C. Clark,Daniel Skrovonsky,Y. Quiroz,M. Pontecorvo,A. Ruiz,F. Lopera,M. Mintun,P. Thiyyagura,E. Reiman,P. Tariot,A. Roontiva,Wendy Lee,Xiaofen Liu,Keith A. Johnson,P. Doraiswamy

Abstract:Alzheimer’s disease (AD).We investigated the accuracy of a novel unsupervised multimodal biomarker classifier for differentiating cognitively normal elderly (NC) from subjects with amnestic mild cognitive impairment (aMCI). By combining imaging and genetic biomarker data, we hypothesized that we would achieve greater accuracy in differentiating the diagnostic groups.Methods:Using automated segmentation techniques, we derived hippocampal and lateral ventricle volumes from the T1-weighted MRI data of 46 NC and 35 aMCI subjects. We collected gene expression (GE) data from all subjects and single nucleotide polymorphism (SNP) data on common variants in ApoE, TOMM40, PICALM, CLU, CR1, MAPT and PCDH11X from 44NC and 28 aMCI subjects. Using a novel automated support vector machine algorithm , we developed unimodal andmultimodal imaging and genetic diagnostic classifiers. All classifiers included age and sex. Results: In the N 1⁄4 72 imaging/SNP dataset, a classifier that used hippocampal volume only achieved 76.4% diagnostic accuracy (area under the curve, AUC 1⁄4 0.67) compared to the classifier based on ventricular volume only accuracy 69.4% (AUC 1⁄4 0.56) and the combined hippocampalventricular classifier accuracy 74% (AUC1⁄4 0.7). The addition of SNP variables led to a hippocampal-SNP classifier accuracy of 76% (AUC 1⁄4 0.74) and a hippocampal-ventricular-SNP classifier accuracy of 72% (AUC 1⁄4 0.69). Of the 7 SNPs entered, PICALM was selected by the hippocampalSNP classifier, ApoE by the hippocampal-ventricular-SNP classifier while TOMM40 was selected by both classifiers. The remaining SNPs were not included in the optimal classification algorithm. In the N 1⁄4 81 imaging/ gene expression dataset the hippocampal-only classifier achieved 69% diagnostic accuracy (AUC1⁄4 0.63) and ventricular-only classifier achieved 69% accuracy (AUC 1⁄4 0.57) compared to the hippocampal-GE classifier accuracy 78% (AUC 1⁄4 0.79), and the combined hippocampal-ventricular-GE classifier accuracy 84% (AUC 1⁄4 0.82). 12 expressed genes and 8 expressed genes were selected as being useful for improving classification, by the final hippocampal-GE and hippocampal-ventricular-GE combined classifiers, respectively. Conclusions: As hypothesized, NC vs. aMCI classifier performance improved when combining imaging and genetic biomarkers. Automated classifiers show great promise for diagnostic analyses and potentially for predicting future conversion to AD.

Biology,Medicine