Premium
IC‐P‐030: MCI diagnosis via manifold based classification of functional brain networks
Author(s) -
Fan Yong,
Browndyke Jeffrey N.
Publication year - 2010
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 - 10.1016/j.jalz.2010.05.044
Subject(s) - discriminative model , pattern recognition (psychology) , artificial intelligence , independent component analysis , computer science , functional magnetic resonance imaging , subspace topology , parietal lobe , neuroscience , psychology
Background: Accumulated evidence from fMRI studies has demonstrated that Alzheimer’s disease (AD) is associated with abnormalities of brain functional networks (FNs) in medial temporal lobe, frontal, temporal, and parietal cortices. Identifying brain FNs affected by mild cognitive impairment (MCI) and optimally utilizing them in diagnosis could potentially improve early detection of AD. This study applied independent component analysis (ICA) to extract brain FNs from fMRI data of MCI and cognitively normal (CN) elderly subjects during their performance of a simple semantic memory task and resting visual fixation. Advanced pattern classification techniques identified the optimal combination of FNs for diagnostic determination. Methods: A manifold based classification algorithm was used to classify fMRI data of 12 MCIs and 12 CNs. A group ICA was used to extract FNs which were encoded by spatial independent components (ICs) and a back-reconstruction technique was used to compute subject specific FNs. The FNs of each individual were used as bases for spanning a linear subspace, referred to as a functional connectivity pattern (FCP), which facilitated a comprehensive characterization of temporal signals of fMRI data. The FCPs of different individuals were analyzed on the Grassmann manifold by adopting a principal angle based subspace distance. In conjunction with a k-nearest neighbor classifier, a forward component selection technique was used to select ICs for constructing the most discriminative FCP whose discriminative power was measured using leave-one-out cross validation. Results: We identified a FCP spanned by 7 FNs, including temporal, parietal, and frontal regions, which were most characteristic of MCI. This combined cognitive challenge and resting state FCP correctly distinguished 20 out 24 subjects, cross-validated using leave-one-out method. This result is better than those obtained by the classifiers built on FCPs spanned by either all available FNs (50%) or any individual FN (the best performance: 54.2%). Conclusions: The manifold based classification method has the potential to identify MCI associated FNs, which could serve as surrogate biomarkers for early detection of individuals at risk for AD. Furthermore, the FCP spanned by multiple discriminative FNs has superior diagnostic value compared to either any individual FN or the FCP spanned by all available FNs.