Premium
Individual level prediction of cognitive impairment status from neuronal networks using machine learning in the Alzheimer’s Disease Connectome Project
Author(s) -
Kulkarni Arman P.,
Adluru Nagesh,
Nair Veena A.,
Li ShiJiang,
Meyerand Elizabeth M.,
Alexander Andrew L.,
Bendlin Barbara B.,
Prabhakaran Vivek
Publication year - 2021
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.057485
Subject(s) - artificial intelligence , connectome , random forest , cross validation , diffusion mri , dementia , human connectome project , binary classification , receiver operating characteristic , computer science , neuroscience , pattern recognition (psychology) , machine learning , psychology , disease , magnetic resonance imaging , medicine , pathology , support vector machine , functional connectivity , radiology
Background Alzheimer’s disease dementia (AD) and mild cognitive impairment (MCI) are characterized by degeneration of myelinated axonal fiber bundles which subserve network connectivity, and degeneration of which ultimately leads to the dementia syndrome. Most of the prior disease related findings derived from diffusion weighted magnetic resonance imaging (dMRI) are based on group comparisons, although individual‐level measures of abnormality would have greater clinical utility as well as serve clinical trial enrichment and provide novel outcome measures. In this study, fiber bundle capacity (FBC) was employed to derive a proxy for “communication bandwidth” between brain regions based on dMRI. The purpose of this study was to perform a preliminary investigation of individual level predictive value of FBC based neuronal networks in the Alzheimer’s disease connectome project using machine learning. Method Multi‐shell dMRI data (Table‐1) were processed using a state‐of‐the‐art ‘connectography’ pipeline using FSL, ANTS and MRtrix3. The IIT‐Desikan atlas with 84 nodes was used to derive 3486 neuronal connections (FBC‐edges). FBC‐edges, after feature preprocessing, were input into a stratified three‐fold nested leave‐one‐out cross‐validation framework with a random forest classifier for binary classification (cognitively unimpaired (CU) vs. MCI/AD). Recursive feature elimination was employed in the outer loop to aid with the interpretation of predictive features. Receiver‐operating‐characteristic curves were derived per outer fold, and the area under the curves (AUC) were reported. Result The average AUC was 0.73 as shown in Fig‐1, with additional cross‐validation accuracies and test‐set accuracies shown in Table‐2. The top (20) features per outer fold are reported in Fig‐2. Conclusion FBC based neuronal networks offer significant predictive value in differentiating the MCI/AD brain from that of the CU brain at an individual level, indicating alterations in information pathways across the brain. Depending on the fold, there were alterations in FBC connections 1) involving the right parietal lobe, and in particular, the precuneus to the left parietal lobe, 2) parietal‐parietal, parietal‐temporal, and parietal‐frontal interactions, and 3) left parietal interactions and intratemporal interactions. Connections of the precuneus with other nodes were predominant across all folds. Future work will determine the extent to which these metrics are predictive of amyloid burden among individuals with asymptomatic preclinical AD.