All patients originated from the AddNeuroMed project, part of InnoMed (Innovative Medicines in Europe), a European Union programme designed to make drug discovery more efficient. The project is designed to develop and validate novel surrogate markers in AD and includes a human neuroimaging strand (Simmons et al., 2009, 2011) which combines MRI data with other biomarkers and clinical data. Data was collected from six different sites across Europe: University of Kuopio, Finland, University of Perugia, Italy, Aristotle University of Thessaloniki, Greece, King's College London, United Kingdom, University of Lodz, Poland, and University of Toulouse, France. MR images from a total of 119 AD patients, 119 MCI patients, and 110 healthy controls (CTL) were available. Only a subsample of 62 AD patients, 73 MCI patients and 79 CTL individuals were followed longitudinally however. Table 1 provides demographics of the whole and longitudinal datasets, Table 2 summarizes the cognitive scores per diagnostic group, and Figure 1 depicts the disposition of the diagnostic groups. All AD and MCI subjects were recruited from local memory clinics of the six participating sites while the control subjects were recruited from non-related members of the patient's families, caregiver's relatives or social centers for the elderly. Written consent was obtained where the research participant had capacity, and in those cases where dementia compromised capacity then assent from the patient and written consent from a relative, according to local law and process, was obtained. This study was approved by ethical review boards in each participating country. The inclusion and exclusion criteria were as follows. Alzheimer disease: Inclusion criteria: (1) ADRDA/NINCDS and DSM- IV criteria for probable AD. (2) Mini Mental State Examination score range between 12 and 28. (3) Age 65 years or above. Exclusion criteria: (1) Significant neurological or psychiatric illness other than AD. This would exclude patients with vascular dementia or large infarcts, for example. (2) Significant unstable systematic illness or organ failure. All AD subjects had a Clinical Dementia Rating (CDR) scale score of 0.5 or above. Mild Cognitive Impairment and controls: Inclusion criteria: (1) Mini Mental State Examination score range between 24 and 30. (2) Geriatric Depression Scale score less than or equal to 5. (3) Age 65 years or above. (4) Medication stable. (5) Good general health. Exclusion criteria: (1) Meet the DSM- IV criteria for Dementia. (2) Significant neurological or psychiatric illness. (3) Significant unstable systematic illness or organ failure. The distinction between MCI and controls was based on: the subject scores 0 on Clinical Dementia Rating Scale = control. The subject scores 0.5 on Clinical Dementia Rating scale = MCI. Additionally, for the MCI subjects it was preferable that the subject and informant reported occurrence of memory problems. There is an overlap regarding CDR-sum-of-boxes and MMSE scores between AD and MCI patients.

At baseline, a health interview was administrated to all participants following a standardized protocol available at http://www.innomed-addneuromed.com and data on demographics, medical history, current health status, medication use and family history were collected. For AD and MCI cases, information on the duration and severity of cognitive decline was obtained from the subjects or informants through a detailed questionnaire: the presence of memory problems, time of their onset, subsequent course, and investigations performed.

For all participants, cognition, behavior, functional status, and global severity were assessed and the MMSE (Folstein et al., 1975) and the Clinical Dementia Rating scale (CDR) (Hughes et al., 1982) were administered. Evaluation of AD included the Hachinski ischemic scale (Hachinski et al., 1975), the AD assessment scale-cognitive subscale (ADAS-Cog) (Rosen et al., 1984), Neuropsychiatric Inventory (Cummings et al., 1994), and the Alzheimer disease Cooperative Study (ADCS)-activities of daily living scale (Galasko et al., 1997). Evaluation of MCI and CTL individuals comprised the Consortium to Establish a Registry for Alzheimer Disease (CERAD) Cognitive Battery (Morris et al., 1989) and the Geriatric Depression Scale (GDS) (Yesavage and Sheikh, 1986). The same procedure was used at 12 month follow-up. MRI findings were not known to clinicians, so as not to influence the clinical diagnosis.

The ADAS-cog and the CERAD battery use the same 10-word recall task. The only difference is that the scoring is inverted. The mean number of words not recalled in the CERAD word list immediate recall task was calculated. The variable obtained was named ADAS1, corresponding to the first subtest of ADAS-Cog. This was performed to obtain comparable measures between groups. CDR global was used for diagnosis as described above, but CDR sum of boxes were used in relation to the index.

Data acquisition for the AddNeuroMed study was designed to be compatible with the Alzheimer Disease Neuroimaging Initiative (ADNI) (Jack et al., 2008). Data was collected from 4 General Electric, 1 Siemens, and 1 Picker MR-scanner. The imaging protocol for both studies included a high resolution sagittal 3D T1-weighted MPRAGE volume (voxel size 1.1 × 1.1 × 1.2 mm³) and axial proton density/T2-weighted fast spin echo images. The MPRAGE volume was acquired using a custom pulse sequence specifically designed for the ADNI study to ensure compatibility across scanners. Full brain and skull coverage was required and detailed quality control carried out on all MR images according to the AddNeuroMed quality control procedures (Simmons et al., 2009, 2011).

FreeSurfer (version 4.5.0) was used for image processing and analysis. The pipeline produces regional cortical thickness and volumetric measures. Cortical reconstruction and volumetric segmentation includes removal of non-brain tissue using a hybrid watershed/surface deformation procedure (Segonne et al., 2004), automated Talairach transformation, segmentation of the subcortical white matter and deep gray matter volumetric structures (including hippocampus, amygdala, caudate, putamen, ventricles) (Fischl et al., 2002, 2004; Segonne et al., 2004), intensity normalization (Sled et al., 1998), tessellation of gray matter and white matter boundary, automated topology correction (Fischl et al., 2001; Ségonne et al., 2007) and deformation following intensity gradients to optimally place the gray/white and gray/CSF borders at the location where the greatest shift in intensity defines the transition to the other tissue class (Dale and Sereno, 1993; Dale et al., 1999; Fischl and Dale, 2000). Once the cortical models are complete, registration to a spherical atlas takes place, which utilizes individual cortical folding patterns to match cortical geometry across subjects (Fischl et al., 1999). This is followed by parcellation of the cerebral cortex into units based on gyral and sulcal structure (Fischl et al., 2004; Desikan et al., 2006). The pipeline generated 68 cortical thickness measures (34 from each hemisphere) and 50 regional volumes. Volumes of white matter hypointensities, optic chiasm, right and left vessel, and right and left choroid plexus were excluded from further analysis. White matter hypointensities were excluded since most subjects were characterized by zero values. Cortical thickness and volumetric measures from the right and left side were averaged. In total 57 MRI measures were used as input variables for multivariate analysis, see Table 3. Using tkmedit, the image processing steps were visually inspected to ensure they had been carried out correctly and corrected for skull-stripping errors and gray/white matter boundary (cases where significant image artifacts or gross brain abnormalities result in segmentation errors were excluded). These segmentation approaches have been used for multivariate classification of Alzheimer's disease and healthy controls (Westman et al., 2010; Mangialasche et al., 2013), neurophysiological image analysis (Liu et al., 2011; Ferreira et al., 2014), imaging-genetic analysis (Liu et al., 2010b; Velayudhan et al., 2013), and biomarker discovery (Thambisetty, 2008; Thambisetty et al., 2010).

All volumetric measures from each subject were normalized by the subject's intracranial volume (ICV). The ICV was estimated based on an affine transform in Freesurfer, while cortical thickness measures were not normalized (Westman et al., 2013). Preprocessing of each variable was performed using mean centering and unit-variance scaling.

Orthogonal partial least squares to latent structures (OPLS) (Trygg and Wold, 2002; Bylesjö et al., 2006; Wiklund et al., 2008) is a supervised multivariate data analysis method included in the software package SIMCA (Umetrics AB, Umea, Sweden). In OPLS, the information related to class separation is found in the first component of the model, the predictive component. The orthogonal components in the model, if any, relate to variation in the data not connected to class separation. Focusing the information related to class separation on the first component makes data interpretation easier (Wiklund et al., 2008).

Each OPLS model receives a Q²(Y) value that describes its statistical significance for separating groups. Q²(Y) values > 0.05 are regarded as statistically significant (Eriksson et al., 2006), where Q2(Y)=1−PRESS/SSY

PRESS (predictive residual sum of squares) = ∑ (y_actual − y_predicted)² and SSY is the total variation of the Y matrix after scaling and mean centering (Eriksson et al., 2006). Q²(Y) is the fraction of the total variation of the Ys (expected class values) that can be predicted by a component according to cross-validation (CV).

An OPLS model was created at baseline for AD vs. CTL (119 AD and 110 CTL). Cortical thickness measures and subcortical volumes were analyzed separately, which means a separate OPLS model is created for each set of variables. The two models are then combined using hierarchical modeling (Westman et al., 2012). The OPLS model creates an index for each individual subject, which tends to be close to 1 for AD patients and 0 for CTL individuals. Index values above 0.5 indicate a more AD-like characteristic while values below 0.5 a more CTL-like characteristic. The AD vs. CTL OPLS model derived from the training data was applied on the longitudinal data of AD, CTL, and MCI individuals. This produced an index reflecting the degree to which the individual's pattern resembled the pattern of AD patients or the pattern of CTL individuals. Although age and education were significantly different between groups, the model did not consider them as confounding variables since multivariate models seemed to be robust to the presence of differences in age and/or years of education because the overall pattern of atrophy is a strong predictor of AD in and of itself (Klöppel et al., 2008; Vemuri et al., 2008; Misra et al., 2009; Aguilar et al., 2013).

Basic descriptive statics were used to characterize diagnostic groups and stratified according to a specific criterion (e.g., gender or time point). Linear regression was used to illustrate dependences and quantified by computing linear correlation coefficients. Statistical significance was determined by applying repeated One-Way ANOVA and non-parametric matched-paired test. Area under the ROC (receiver operating characteristic) curve (AUC) was calculated using the web-based application JROCFIT (Eng J. ROC analysis: web-based calculator for ROC curves. Baltimore: Johns Hopkins University; www.jrocfit.org).

The resulting classification model enabled high discrimination between AD and CTL individuals (7-fold CV: 84% sensitivity, 91% specificity, 0.93 AUC). As part of the model, the top 10 most important variables for discrimination were: hippocampus, entorhinal cortex, amygdala, temporal pole, superior temporal gyrus, inferior lateral ventricle, middle temporal gyrus, fusiform gyrus, inferior temporal gyrus, and parahippocampal gyrus (see Table 3 for a list of all variables included in the analysis).

The index was computed for all individuals in the longitudinal dataset. Figure 2A and Table 4 summarize the distribution of the index per diagnostic group, per time-point; and Figure 2B summarizes the distribution of the index annual rate of change per diagnostic group. Overall, we observed increased atrophy over time, AD patients showed the fastest increase (One-Way ANOVA P = 0.044). The index was highest in the AD group, followed by MCI converters (MCI-c), MCI non-converters (MCI-nc), and the CTL group. CTL subjects had values below 0.5. However, in older individuals with normal cognition AD-like values (>0.5) were more frequent (Figure 3).

Taking the average index from baseline and follow-up data and stratifying the diagnostic groups according to pattern of atrophy resulted in good separation of AD-like AD patients and CTL-like CTL subjects, as well as AD-like MCI-c and CTL-like MCI-nc patients, and revealed subgroups of subjects with values in between (Figure 4A). Also, for each diagnostic group, having a different pattern of atrophy resulted in a different index of annual rate of change, although non-significant, except for the CTL group (Figure 4B). Those CTL individuals having an AD-like pattern of atrophy had a significantly higher rate of atrophy (Index annual rate of change: AD-like > CTL-like; Mann-Whitney p = 0.028).

Sensitivity, specificity and AUC are reported for classification (AD vs. CTL) and prediction of MCI conversion to AD (Table 5).

Individual classification accuracy was significantly lower (p < 0.001) at baseline (7-fold CV: sensitivity 81%, specificity 90%, 0.90 AUC) compared to 1-year follow-up (sensitivity 92%, specificity 75%, 0.93 AUC). Classification accuracy increased due to an increase in sensitivity, but there was a decrease in specificity as well.

MCI individuals were labeled as AD-like or CTL-like based on their individual index. Furthermore, they were stratified as converters (MCI-c) or non-converters (MCI-nc) based on the 1-year follow-up diagnosis. We report the number of MCI-c subjects identified as AD-like. Out of 13 MCI-c subjects 11 were predicted as AD-like from baseline data, while 12 were classified as AD at follow-up.

Given that changes in brain structure (e.g., atrophy) should relate to cognitive performance, an exploratory statistical analysis was carried out on the cognitive data available. After stratification by diagnostic group, only CDR-sum showed significantly differences between baseline and 1-year scores for AD (P < 0.001) and MCI (P < 0.001) patients, reflecting an increased impairment; there were no significant changes in ADAS1 or MMSE.

The association between the index and CDR-sum was assessed based on the linear correlation coefficient for each diagnostic group at each time point. At baseline, there was no correlation for MCI patients (Spearman's r = 0.220; P = 0.093), while a mild non-significant correlation was found for AD patients (Spearman's r = 0.29; P = 0.063) (Figure 5A). At 1-year follow-up, there was a fair correlation for MCI patients (Spearman's r = 0.40; P = 0.001), and a good correlation for AD patients (Spearman's r = 0.59; P < 0.001) (Figure 5B). This suggests that at 1-year follow-up AD and MCI patients became more atrophied and more impaired, and the index and CDR-sum scores became more correlated.

Since computing the average index from baseline and follow-up data and stratifying by diagnostic groups yielded significant differences for the index analysis, this procedure was also applied on the cognitive measures. Computing the average index and average cognitive scores, there were significant correlations between the index and MMSE (Figure 6A; Spearman's r = −0.614, p < 0.001), ADAS1 (Figure 6B; Spearman's r = 0.562, p < 0.001) and CDR-sum (Figure 6C; Spearman's r = 0.679, p < 0.001). When the analysis was restricted to the AD group, significant correlations were found between the average index and average MMSE (Spearman's r = −0.438, p < 0.001), average ADAS1 (Spearman's r = 0.549, p < 0.001) and average CDR-sum (Spearman's r = 0.412, p = 0.001). In the CTL group there was significant correlation only between the average index and average CDR-sum (Spearman's r = 0.236, p = 0.036). For the MCI groups there were no significant correlations.

Having a different pattern of atrophy may be reflected in having a different cognitive performance. In particular, there was a faster rate of cognitive decline in AD-like subjects compared to CTL-like in MMSE (for AD Mann-Whitney p = 0.035; for CTL Mann-Whitney p = 0.05, one-sided T-test p = 0.03), ADAS1 (for MCI-c Mann-Whitney p = 0.013) and CDR-sum (for AD Mann-Whitney p < 0.001). Additionally, stratification of CTL individuals by pattern of atrophy resulted in significant correlations between the average index and average MMSE (only AD-like CTL Spearman's r = −0.465, p = 0.039), average ADAS1 (only AD-like CTL Spearman's r = 0.504, p = 0.023) and average CDR-sum (only CTL-like CTL Spearman's r = 0.351, p = 0.006).

Carrying the APOE e4 allele is the strongest genetic risk factor associated with AD. This was used to stratify individuals as carriers vs. non-carriers and look for an association in cognitive decline consistent with AD pathology. Carriers showed a higher index compared to non-carriers at baseline and follow-up (baseline P = 0.041; 1-year P = 0.014) indicative of an AD-like atrophy pattern. Moreover, carriers showed an increase in ADAS1 (P = 0.017) and CDR-sum (P = 0.013) score compared to non-carriers at 1-year follow-up (Figure 7). Therefore, MCI carriers of the APOE e4 allele seemed to be more atrophied and presented a faster cognitive decline compared to non-carriers.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.