Principal component analysis in mild and moderate Alzheimer's disease — A novel approach to clinical diagnosis

Abstract

Principal component analysis (PCA) provides a method to explore functional brain connectivity. The aim of this study was to identify regional cerebral blood flow (rCBF) distribution differences between Alzheimer's disease (AD) patients and controls (CTR) by means of volume of interest (VOI) analysis and PCA. Thirty-seven CTR, 30 mild AD (mildAD) and 27 moderate AD (modAD) subjects were investigated using single photon emission computed tomography with ^99mTc-hexamethylpropylene amine oxime. Analysis of covariance (ANCOVA), PCA, and discriminant analysis (DA) were performed on 54 VOIs. VOI analysis identified in both mildAD and modAD subjects a decreased rCBF in six regions. PCA in mildAD subjects identified four principal components (PCs) in which the correlated VOIs showed a decreased level of rCBF, including regions that are typically affected early in the disease. In five PCs, including parietal-temporal-limbic cortex, and hippocampus, a significantly lower rCBF in correlated VOIs was found in modAD subjects. DA significantly discriminated the groups. The percentage of subjects correctly classified was 95, 70, and 81 for CTR, mildAD and modAD groups, respectively. PCA highlighted, in mildAD and modAD, relationships not evident when brain regions are considered as independent of each other, and it was effective in discriminating groups. These findings may allow neurophysiological inferences to be drawn regarding brain functional connectivity in AD that might not be possible with univariate analysis.

Introduction

Comparisons between regional cerebral blood flow (rCBF) patterns in Alzheimer's disease (AD) patients and normal controls (CTR) by single photon emission computed tomography (SPECT) or positron emission tomography (PET) have in the past mainly been carried out by either visual evaluation or outlining the regions of interest (ROIs) in a manual or semiautomatic mode (Syed et al., 1992, Salmon et al., 1994, Bonte et al., 1997, Jagust et al., 1997, Jagust et al., 2001). In recent years, several digitized spatial standardization software programs have been introduced, and some of them have been extensively used both in research and clinical investigations (Friston et al., 1995, Worsley et al., 1996, Hirsch et al., 1997, Houston et al., 1998, Bartenstein et al., 1997, Imran et al., 1999). Among them, the Computerized Brain Atlas (CBA) (Greitz et al., 1991, Thurfjell et al., 1995, Andersson and Thurfjell, 1997), when registered to imaging data, defines volumes of interest (VOIs), corresponding to anatomical and functional areas of the brain (Brodmann areas and deep grey structures). It thus allows investigating rCBF relationships between anatomically distributed but physiologically correlated brain regions using principal component analysis (PCA) (Pagani et al., 2002). Moreover, the univariate approach treats all variables as if they represent independent measures across brain regions, thus ignoring the extensive and structured connections between them. Possible networking is better investigated by multivariate analysis describing pattern of covariance at both the voxel and VOI levels.

Multivariate spatial covariance methods have been proposed in functional neuroimaging by several authors (Moeller et al., 1987, Moeller et al., 1999, Eidelberg et al., 1990, Eidelberg et al., 1994, Friston et al., 1993, Friston et al., 1999, Strother et al., 1995, McIntosh et al., 1996, Jones et al., 1998, Zuendorf et al., 2003, Greicius et al., 2004, Lozza et al., 2004, Habeck et al., 2005, Hu et al., 2005, Kerrouche et al., 2006). All such studies were meant to overcome the concept of functional segregation as derived by univariate analysis of independent variables and to interpret the results in the frame of knowledge of brain circuitry and thus from a more holistic point of view.

The application of PCA aims to reduce the dimensionality of the data matrix through the grouping of VOIs into principal components that will undergo further statistical analysis. In AD, the rCBF distribution is likely to be the result of a combination of pathological lesions and the effects they can have on distant parts of the brain (Stoub et al., 2006). Such an approach has been applied to ¹⁸F-FDG PET data to characterize patterns of covariance in AD (Salmon et al., 2007) and to distinguishing between AD and vascular dementia (VaD) (Kerrouche et al., 2006). In studies of resting state rCBF with H₂¹⁵O PET (Scarmeas et al., 2004, Devanand et al., 2006) or SPECT (Huang et al., 2007), AD patients and subjects with mild cognitive impairment showed a distinct covariance pattern versus controls. PCA has been successfully performed implementing both a voxel-based analysis (Kerrouche et al., 2006, Scarmeas et al., 2004, Devanand et al., 2006) and using discrete regions of interest (Huang et al., 2007). These studies have consistently shown that PC composition may allow inferences on the peculiar functional re-organization of brain networking in both AD dementia and pre-dementia states, as well as in VaD (Kerrouche et al., 2006) and frontotemporal dementia (Salmon et al., 2006). Therefore, we hypothesised that investigating AD by means of multivariate analysis (i.e. PCA) might add knowledge about the rCBF pattern and neural connectivity underlying this complex neurodegenerative disorder.

The aim of the present study is to use multivariate statistical methodology to assess rCBF distribution differences between two groups of AD patients with different levels of disease severity and a group of normal controls.