An Ensemble Machine Learning Approach to Predict Cognitive Impairment and Amyloid Deposition through Metabolic and Vascular Deficits in ADNI Cohorts

Abstract

Metabolic and vascular changes in the brain have been shown to pre-date the development of cognitive impairment and amyloid-β deposition in Alzheimer's disease by several decades. However, these changes are not well characterized and currently lack clinical applications. Thus, the aim of this study was to elucidate that metabolic and vascular deficits that are indicative of Alzheimer's disease progression through an ensemble machine learning approach. Multimodal neuroimaging data (MRI and PET) from 193 patients ranging from normal cognition (n = 63) to mild cognitive impairment (n = 101) to Alzheimer's disease (n = 29) were obtained from ADNI, and analyzed for glucose uptake, cerebral blood flow, and florbetapir concentration (amyloid-β indicator). A meta-analysis was then performed to determine brain regions that typically exhibit changes in longitudinal Alzheimer's studies. These regions of interest, along with neuropsychological examinations, genetics, and demographic information were used to engineer features for a prediction framework that implemented nearest-neighbor algorithm, k-means clustering, and a multilayer perceptron neural network. The framework achieved 71.0% accuracy in identifying mild cognitive impairment patients with a receiver operating characteristic of 0.722, and achieved 92.3% accuracy in predicting conversion from mild cognitive impairment to Alzheimer's disease within three years, with a receiver operating characteristic of 0.959. This approach not only significantly outperformed existing methods, but can be used for cohort selection in future clinical drug trials. Linear correlation plots of FDG and CBF (A.), FDG and Florbetapir (B.), and CBF and Florbetapir (C.). FDG: Fluorodeoxyglucose (18F); CBF: Cerebral blood flow; AV 45: Florbetapir, SUV: standardized uptake value; Aβ: Amyloid-beta. Color-coded linear regression plots of CBF against FDG for male CN (A.), MCI (B.), and AD (C.) and female CN (D.), MCI (E.), and AD (F.) patients. Red circles indicate left inferior parietal, orange circles are right inferior parietal, black circles are posterior cingulate, green circles are left inferior temporal, and blue circles correspond to right inferior temporal. P-values are significant in male MCI, female MCI, and female AD. The predictor space (A.) shows the relationship between gender, left angular glucose uptake, and right angular glucose uptake. A four-layer MLP achieved 71.0% accuracy in identifying MCI patients, with receiver operating characteristics of 0.722 (B.). Normalized importance of each feature was then calculated and plotted (C.), with the highest five features originating in neuroimaging. An ensemble approach that implemented k-means clustering and a multilayer perceptron neural network achieved 92.3% accuracy with a receiver operating characteristic of 0.959 (A.). Normalized importance of each feature was then calculated and plotted (B.). This abstract is from the Experimental Biology 2018 Meeting. There is no full text article associated with this abstract published in The FASEB Journal.