Prediction of Alzheimer’s Disease in Amnestic Mild Cognitive Impairment Subtypes: Stratification Based on Imaging Biomarkers

Abstract

Background: Prediction of progression to Alzheimer’s disease (AD) in amnestic mild cognitive impairment (MCI) is challenging because of its heterogeneity.

Objective: To evaluate a stratification method on different cohorts and to investigate whether stratification in amnestic MCI could improve prediction accuracy.

Methods: We identified 80 and 79 patients with amnestic MCI from different cohorts, respectively. They underwent baseline magnetic resonance imaging (MRI) and ¹⁸F-fluorodeoxyglucose (FDG) positron emission tomography (PET) scans. We performed hierarchical clustering with three imaging biomarkers: Brain volume on MRI, left hippocampus grey matter loss on MRI, and left inferior temporal gyrus glucose hypometabolism on FDG-PET. Regions-of-interest for biomarkers were defined by the Automated Anatomical Labeling atlas. We performed voxel-wise statistical parametric mapping to explore differences between clusters in patterns of grey matter loss and glucose hypometabolism. We compared time to progression using an interval-censored parametric model. We evaluated predictive performance using logistic regression.

Results: Similar clusters were found in different cohorts. MCI1 had the healthiest biomarker profile of cognitive performance and imaging biomarkers. MCI2 had cognitive performance and MRI measures intermediate between those of nonconverters and converters. MCI3 showed the severest reduction in brain volume and left hippocampal atrophy. MCI4 showed remarkable glucose hypometabolism in the left inferior temporal gyrus, and also demonstrated significant decreases in most cognitive scores, including non-memory functions. MCI4 showed the highest risk for progression. The prediction of progression of MCI2 especially benefited from the stratification.

Conclusion: Stratification with imaging biomarkers in amnestic MCI can be a good approach for improving predictive performance.

Keywords

Alzheimer’s disease clustering heterogeneity magnetic resonance imaging mild cognitive impairment positron-emission tomography

INTRODUCTION

Mild cognitive impairment (MCI) [1] is a heterogeneous entity due to various pathological substrates and characterized by different outcomes [2 –4]. Different research groups use various definitions [4]. MCI may lead to dementia resulting from various diseases such as Alzheimer’s disease (AD), vascular dementia, dementia with Lewy bodies, frontotemporal lobar degeneration, and primary progressive aphasia [5]. Of these, AD is the primary cause of dementia. Heterogeneity in MCI makes it difficult to predict progression (conversion) of MCI to AD. In general, it is more difficult to distinguish MCI converters from nonconverters based on their baseline biomarkers than to distinguish AD from normal controls or to distinguish MCI from normal controls [6 –8].

Amnestic MCI, characterized by memory impairment, is also heterogeneous [5, 9]. A recent study reported biological heterogeneity in amnestic MCI [10] in which hierarchical clustering of the Alzheimer’s Disease Neuroimaging Initiative (ADNI) amnestic MCI subjects on the basis of 11 biomarkers identified four clusters. One of these clusters had a higher risk of progression. However, unfortunately, differences in predictive performance among these clusters were not investigated. In addition, use of less invasive biomarkers for clustering rather than cerebrospinal fluid (CSF) biomarkers would be preferable. Thus, we focused on clustering based on imaging-based biomarkers. We have previously investigated an atlas-based feature extraction approach [11, 12]. This method has the advantage of methodological simplicity and availability of an anatomical framework to interpret results.

In this study, we performed agglomerative hierarchical clustering based on three imaging-based biomarkers to stratify patients with amnestic MCI from two different cohorts. We then examined whether stratification into subgroups can improve the prediction accuracy for progression of MCI relative to that for no stratification.

MATERIALS AND METHODS

Datasets

Data used in the preparation of this article were obtained from the Research group of the Studies on Diagnosis of Early Alzheimer’s Disease-Japan (SEAD-J) [13] and the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu/). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early AD. For up-to-date information, see http://www.adni-info.org.

SEAD-J participants

We identified 80 patients from 114 participants with amnestic MCI in the SEAD-J Study [13]. This study was approved by the local Institutional Review Boards of all participating institutions. Diagnosis of MCI was based on an interview with neurologists as described previously [12]. “Progression (conversion)” was defined as a change in diagnosis from amnestic MCI to AD during a 3-year follow-up period. Forty participants (50%) progressed to AD within 3 years. We excluded 34 participants from the analysis for the following reasons: Two participants did not undergo baseline three-dimensional T1-weighted MRI scans, three participants converted to non-Alzheimer’s dementia (vascular dementia, dementia with Lewy bodies, and frontotemporal dementia), 23 participants withdrew from the study within 3 years, and six participants were excluded because of the lack of whole-brain coverage in their baseline T1-weighted MRI scans. Tables 1 and 2 show more details on the participants’ baseline characteristics, including the Mini-Mental State Examination (MMSE) [14], the Alzheimer’s Disease Assessment Scale-Cognitive Subscale, Japanese version (ADAS-Jcog) [15], the Wechsler Memory Scale-Revised Logical Memory test (WMS-R LM) [16], and the Geriatric Depression Scale (GDS) [17, 18]. The follow-up period was defined as the period during which a participant was diagnosed as having AD. The nonconverters’ follow-up period was 3 years. The converters’ follow-up periods included 1, 2, and 3 years. Each patient was classified as C1, C2, C3, or NC according to his or her follow-up period and diagnosis. NC represents a nonconverter.

ADNI participants

To evaluate our stratification method on a different dataset, we performed an additional analysis with the ADNI dataset. We identified 79 subjects with amnestic MCI from participants in ADNI1. All subjects underwent screening (baseline) T1-weighted MRI scans and FDG-PET scans, and underwent lumbar puncture as described in the ADNI procedures manual (http://www.adni-info.org/). “Conversion” was defined as a change in diagnosis from aMCI to AD during a 36-month follow-up period. The differences in baseline characteristics between ADNI and SEAD-J subjects were summarized in Table 1.

MRI acquisition and preprocessing

Three-dimensional structural MRI scans at baseline from SEAD-J subjects were collected using T1-weighted gradient echo sequences on a variety of 1.5-T MRI scanners at eight sites and a 3.0-T MRI scanner at one site. Of 80 participants, 77 (96%) participants were scanned on the 1.5-T scanners and 3 (4%) on the 3-T scanner. Details about MRI acquisition parameters are provided in Supplementary Table 2. Baseline (screening) ADNI MRI dataset was downloaded from the Laboratory of Neuro Imaging (LONI) Image Data Archive (https://ida.loni.usc.edu/). These three-dimensional structural MRI scans were acquired from 1.5-T MRI scanners. We performed the following voxel-based morphometry (VBM) preprocessing procedures using the SPM8 software package (http://www.fil.ion.ucl.ac.uk/spm/software/spm8) and VBM8 Toolbox (http://dbm.neuro.uni-jena.de/vbm) in MATLAB 7.12. The images were first segmented into grey matter, white matter, and cerebrospinal fluid using Unified Segmentation [19] implemented in SPM8 and a technique based on the maximum a posteriori (MAP) estimation [20] and the Partial Volume Estimation [21] implemented in VBM8 with standard parameters. The segmented images were spatially normalized using the Diffeomorphic Anatomical Registration using the Exponentiated Lie Algebra (DARTEL) algorithm [22]. Jacobian modulation was applied to compensate for the effect of spatial normalization and to restore the original absolute grey matter density in the segmented grey matter images. We performed standard smoothing using an 8-mm full-width at half-maximum isotropic Gaussian kernel to obtain smoothed modulated normalized data.

¹⁸F-FDG-PET image acquisition and preprocessing

FDG-PET scans from SEAD-J subjects were collected in a resting state in a dark room 40–60 min after venous injection of FDG [23]. Baseline FDG-PET scans from 79 ADNI subjects were downloaded from the LONI Image Data Archive. The whole brain was used as the reference region for intensity normalization to compare the effects of two different brain atlases on the ability of prediction because of its higher signal-to-noise ratio than intensity normalization to the cerebellum [24]. The average whole brain uptake of the converters was 10143±10289 and that of nonconverters was 8451±10912, (t-test, p = 0.48, Cohen’s d = 0.11). The images were normalized using an in-house FDG-PET template created from FDG-PET and MRI brain scans of 23 normal elderly individuals [age (mean±SD) = 71±3.5 years, male = 11 (47.8%)]. Partial volume effects were not corrected according to a previous finding that partial volume effect correction did not affect detection of hippocampal hypometabolism in amnestic MCI when using the global mean for scaling [25]. The resulting images were smoothed using an 8 mm isotropic Gaussian filter to blur individual variations in gyral anatomy and to increase the signal-to-noise ratio [25].

Feature extraction

To extract feature vectors from the preprocessed structural MRI and FDG-PET data above, we used an atlas-based approach as described previously [12]. In this study we used regions-of-interest (ROIs) defined by the Automated Anatomical Labeling (AAL) atlas [26]. The AAL atlas has 116 ROIs, including 90 cerebral regions and 26 cerebellar regions. We extracted feature vectors from the preprocessed baseline MRI, which was modulated with the Jacobian, and from the FDG-PET data to obtain MR-based features representing grey matter density in each ROI and PET-based feature representing the relative cerebral metabolic rate for glucose in each ROI.

Measures for clustering

To cluster patients with amnestic MCI, we selected three imaging-based measures at baseline on the basis of prior knowledge as described in detail in the Discussion section: (1) brain volume, the ratio of grey matter volume (GMV) and white matter volume (WMV) to intracranial volume (ICV); (2) left hippocampus (HIP) atrophy, grey matter density in the left HIP on MRI; and (3) left inferior temporal gyrus (ITG) hypometabolism, relative cerebral metabolic rate of glucose in the ITG on FDG-PET.

The first measure, brain volume, was defined as the sum of GMV and WMV divided by each subject’s ICV. ICV was the sum of the raw volume estimates (grey matter, white matter, and CSF) from VBM8 outputs.

The other two measures were ROI-based measures. Of the 116 regions defined by AAL, we selected the left hippocampus for MRI and the left ITG for FDG-PET. The MRI-based measure reflects grey matter density in the left HIP. A meta-analysis of voxel-based morphometry studies reported that left medial temporal lobe atrophy is the most consistent structural biomarker to predict progression of amnestic MCI to AD [27].

The FDG-PET-based measure reflects relative glucose metabolism in the left ITG. Hypometabolism in the left temporal cortex is important for prediction of progression to AD [28]. Previous work using autopsies reported that amnestic MCI had significant synaptic loss in the left ITG relative to that in a group with no cognitive impairment and that synaptic numbers correlated highly with MMSE scores and a test of category verbal fluency [29].

Clustering

To investigate whether a heterogeneous MCI cohort could be stratified into more homogeneous subgroups with the image-based features above, we used agglomerative hierarchical clustering with Ward’s method and the Euclidean distance measure [10]. To estimate the optimal number of clusters, we used the gap statistical method [30]. To compute the gap statistic, we implemented the algorithm in R that was based on source code available at https://github.com/echen/gap-statistic. We determined the optimal number of clusters as four from the plot of the gap statistic.

Statistical parametric mapping

To clarify the differences in MRI and FDG-PET data among MCI clusters, we performed voxel-based comparison of grey matter maps and of FDG-PET images between each cluster and a group of 23 normal controls [mean age±SD = 71.0±3.5, 11 males (47.8%)] using the SPM8 software package. We used analysis of covariance models with age and sex as covariates to obtain statistical maps. To visualize differences between clusters, the statistical maps were mapped onto the surface using Caret version 5.65 [31]. The image threshold was set to T > 2 (approximately uncorrected p < 0.03) for the purpose of visualization.

To further clarify the differences in laterality, we computed and visualized the asymmetry index (AI) defined by the following formula: AI = (left –right)/(left + right). Positive AI values (colored in red) represent left-dominant asymmetry, and negative AI values (colored in blue) represent right-dominant asymmetry.

Survival analysis

To assess time to progression, we performed interval-censored survival analysis with a Weibull AFT model [32] as described in a prior study [10]. We used the Package “survival” for R and the R Package “SurvRegCensCov”.

Prediction of progression to AD

To investigate whether stratification of amnestic MCI patients improves prediction of progression relative to that of the whole group of subjects, we used logistic regression for classification. MCI3 was excluded because of its small sample size. We built models using L2-regularized logistic regression (dual) implemented in LiblineaR package, R interface to LIBLINEAR [33]. We performed a grid-search using five-fold cross-validation to select the best parameter C for L2-regularization.

We performed 100 trials of leave-one-out cross-validation (LOOCV). In each trial, we generated vertically averaged receiver operator characteristic (ROC) curve [34] and computed the area under the curve (AUC) as a measure of the performance of a classifier [35, 36] using the pROC package for R [37]. To compare the ability of prediction among the groups, we formed an averaged ROC curve with 95% confidence intervals for each group on the basis of the results of 100 LOOCV tests using Fawcett’s vertical averaging algorithm [35].

Correlation analysis

For several biomarkers selected from the ADNI1 dataset, Pearson’s correlation coefficients between each variable and the others were computed. Correlation matrices for the overall MCI subjects, MCI1, MCI2, MCI3, and MCI4 were displayed with corrplot package for R.

Statistical analysis

To test for differences between the distributions of categorical variables, χ² test with adjusted residuals was performed. To estimate effect sizes, we computed Cramer’s V for χ² test. Student’s t-tests were performed to compare the means of two groups. To estimate effect sizes, we computed r from t-tests. One-way ANOVAs were performed to test for differences between more than two groups. In case of significant group differences, the ANOVAs were followed by the Tukey–Kramer post hoc multiple comparisons. To estimate effect sizes, we computed η ² as the ratio of the sum of squares for an effect to the total sum of squares [38]. Values of p < 0.05 were considered to indicate statistical significance. All analyses except voxel-based comparisons were performed using the R Statistical Computing Environment, version 3.0.2 [39].

RESULTS

Baseline differences between the SEAD-J and the ADNI cohorts

MCI subjects in the two cohorts significantly differed in baseline characteristics, including age, gender, education, MMSE, WMS-R LM test II, GDS, and left hippocampal gray matter density (p < 0.05) (Table 1).

Stratification in the SEAD-J cohort

SEAD-J MCI subjects were stratified into four clusters (Fig. 1A). The proportion of C1, C2, C3, and NC differed significantly between the clusters (χ² test, p = 0.0016, Cramer’s V = 0.33). C1 and NC in MCI1, C3 and NC in MCI2, and C1 and NC in MCI4 were significant by adjusted residual analysis (p < 0.05) (Fig. 1B). The proportion of converters was also significantly different (χ² test, p = 0.0016, Cramer’s V = 0.49) (Table 2).

These clusters were differentiated primarily on the basis of the MRI measure reflecting grey matter density in the left HIP as seen from the heatmap in Fig. 1A and Fig. 2A. Left HIP atrophy differed significantly [F(3, 76) = 81.1, p = 8×10^–10, η ² = 0.76]. Post hoc multiple comparisons resulted in a significant difference between any pair of clusters [p < 0.001, except between MCI2 and MCI4 (p = 0.001), and between MCI3 and MCI4 (p = 0.003)]. Brain volume was also significantly different [F(3, 76) = 6.01, p = 0.001, η ² = 0.19]. Post hoc testing revealed significant differences between MCI1 and MCI3 (p < 0.001), between MCI2 and MCI3 (p = 0.015), and between MCI3 and MCI4 (p = 0.004). Left ITG hypometabolism also differed significantly [F(3, 76) = 35.7, p = 2×10^–5, η ² = 0.58]. Multiple comparisons indicated significant differences between MCI4 and any of the other clusters (p < 10^–4).

In summary, MCI1 had the healthiest profile of imaging biomarkers. Its left hippocampal atrophy was less severe than the mean of the nonconverters. MCI2 showed an intermediate grey matter reduction. The mean FDG-PET measure for MCI2 was nearly the same as those for nonconverters and MCI1. MCI3 was characterized by a severe reduction in brain volume and the severest left hippocampal atrophy among all groups, whereas the mean FDG-PET measure for MCI3 was the best among the clusters. MCI4 was quite distinct from the other clusters in the FDG-PET biomarker profile and only showed a remarkable decrease in glucose metabolism in the left ITG. MCI4 left hippocampal atrophy was less severe than MCI3 left hippocampal atrophy.

Follow-up was also significantly different between clusters (Table 2). Post hoc multiple comparisons showed significant differences between MCI1 and MCI2 (p = 0.008) and between MCI1 and MCI4 (p < 0.001). All neuropsychological test scores except GDS were significantly different between clusters. Post hoc tests revealed significant differences between MCI1 and MCI4 in MMSE (p = 0.008), Logical Memory I (p = 0.001), Logical Memory II (p = 0.035), and the ADAS-Jcog (p = 0.058) and also between MCI3 and MCI4 in Logical Memory I (p = 0.045).

Stratification in the ADNI cohort

Four clusters were found in the ADNI cohort (Table 3), similar to those from the SEAD-J dataset (Table 2). The proportion of converters was significantly different between clusters (p < 0.001). Follow-up also differed significantly (p = 0.005). MCI1 had a significantly longer follow-up than MCI3 (p = 0.03) and MCI4 (p = 0.006).

All imaging biomarkers for clustering showed significant differences (p < 0.001) (see also Fig. 2B). MCI1 had a significantly greater mean normalized brain volume than any of the other clusters (p < 0.002). Left HIP atrophy was significantly different between any two groups except between MCI2 and MCI4 (p < 0.001). MCI4 had a significantly lower mean ITG glucose metabolism rate than any of the other clusters (p < 0.001).

Tau/Aα_1 - 42 ratio and P-tau_181P/Aα_1 - 42 ratio differed significantly (p < 0.05). Post hoc multiple comparisons showed no significant difference between any of two groups.

Among neuropsychological test batteries, significant differences were found in MMSE, ADAS-cog, Logical memory I and II, RAVLT immediate, % forgetting, 30 minute delay, recognition, Boston Naming Test, Digit Symbol Test, and Trail Making Test Part B (p < 0.05). Post hoc multiple comparisons demonstrated the following significant differences (p < 0.05): Between MCI1 and MCI3, between MCI1 and MCI4, between MCI2 and MCI3, and between MCI2 and MCI4 in ADAS-Jcog; between MCI1 and MCI3, between MCI1 and MCI4, between MCI2 and MCI3 in LM II; between MCI1 and MCI3 and between MCI1 and MCI4 in MMSE, LM I, and RAVLT immediate; between MCI1 and MCI3, between MCI1 and MCI4, and between MCI1 and MCI3 in RAVLT % forgetting, 30 min delay, and recognition; between MCI1 and MCI4, between MCI2 and MCI4, and between MCI3 and MCI4 in BNT; and between MCI1 and MCI4 in Digit symbol and Trail B.

Heterogeneity in patterns of atrophy and hypometabolism

Regarding atrophy, MCI1 from the SEAD-J dataset, similar to the nonconverters, had no remarkable hippocampal atrophy relative to those of the other clusters (Fig. 3A). MCI2, MCI3, and MCI4 showed remarkable grey matter loss in the left medial temporal region. In MCI3, grey matter loss diffusely spread throughout the brain and was rather mild relative to those of MCI2 and MCI4, which suggested that the decreased brain volume in MCI3 may result from ventricular enlargement rather than cortical atrophy.

On the other hand, glucose hypometabolism in the inferior occipital lobe was found in MCI1 and MCI2, although below the significance threshold. MCI2 had severe hypometabolism in the posterior cingulate gyrus/precuneus. MCI3 showed no distinct hypometabolism in the posterior cingulate gyrus and the left ITG. MCI4 had severe hypometabolism in the left ITG and mild hypometabolism in the posterior cingulate gyrus/precuneus.

Left-dominant asymmetries on MRI were found in MCI1, MCI2, and particularly MCI4. MCI3 had no remarkable asymmetry on MRI compared to the other clusters. On the other hand, the FDG-PET patterns differed from those observed on MRI. MCI4 showed a strong left-dominant asymmetry in hypometabolism on FDG-PET. In contrast, the other three clusters had right-dominant asymmetrical patterns, particularly in the lateral temporal region to the parietal region on FDG-PET. All clusters showed left-dominant mild hypometabolism in the inferior frontal region as compared with normal controls.

ADNI clusters also showed different patterns of atrophy and hypometabolism (Fig. 3B) similar to those found in SEAD-J MCI clusters (Fig. 3A). ADNI-MCI4 also had left-dominant asymmetrical patterns of atrophy and hypometabolism.

Time to progression

The estimated median survival time for each cluster from the SEAD-J dataset was 9.72 years for MCI1, 1.69 years for MCI2, 1.88 years for MCI3, and 0.81 years for MCI4 (Fig. 4A). The predicted survival probability at the end of the 3-year follow-up for each cluster was 0.77 (95% confidence interval (CI): 0.53–0.90) for MCI1, 0.33 (0.05–0.66) for MCI2, 0.36 (0.01–0.78) for MCI3, and 0.13 (0.001–0.53) for MCI4 (Fig. 4A). The hazard ratio was 4.34 (95% CI: 1.89–9.96) for MCI2, 3.99 (1.05–15.14) for MCI3, and 8.05 (3.11–20.84) for MCI4 (Fig. 4C).

The estimated median survival time for each cluster from the ADNI1 dataset was 9.61 years for MCI1, 2.29 years for MCI2, 2.27 years for MCI3, and 1.49 years for MCI4 (Fig. 4B). The predicted survival probability at the end of the 3-year follow- up for each cluster was 0.87 (95% confidence interval (CI): 0.55–0.97) for MCI1, 0.36 (0.009–0.80) for MCI2, 0.36 (0.004–0.82) for MCI3, and 0.15 (0.00006–0.70) for MCI4 (Fig. 4B). Hazard ratio for progression to AD relative to progression from MCI1 as a reference was 7.68 (95% CI: 1.82–32.4) for MCI2, 7.78 (1.61–37.5) for MCI3, and 14.2 (3.04–66.1) for MCI4 (Fig. 4D).

Effect of stratification on prediction

Figure 5A shows averaged ROC curves. The mean AUC values for six groups (MCI1+2+3+4, MCI1, MCI2+3+4, MCI2, MCI3+4, and MCI4) were 0.692, 0.739, 0.743, 0.970, 0.693, and 0.758, respectively (Fig. 5B). Prediction for MCI2 showed particularly outstanding performance relative to that of the other groups. The mean AUC values were significantly different between six groups [F(5, 594) = 207, p < 2×10^–16, η ² = 0.64]. Post hoc multiple comparisons revealed significant differences in the mean AUC values between any of the group combinations (p < 0.001) except between MCI1+2+3+4 and MCI3+4 (p = 1.0), between MCI1 and MCI2+3+4 (p = 1.0), between MCI1 and MCI4 (p = 0.47), and between MCI2+3+4 and MCI4 (p = 0.68).

Correlation analysis

The CSF Aα_1 - 42 concentrations were inversely correlated with tau and with P-tau_181P in the overall MCI subjects from the ADNI1 dataset (Fig. 6A), while no correlations were found between CSF Aα_1 - 42 and any of the other CSF measures in MCI4. ADNI subjects in MCI1 had moderate correlations were found between left inferior temporal gyrus hypometabolism and CSF biomarkers as well as a negative correlation between left hippocampal atrophy and tau. Left hippocampal atrophy and FDG-PET metabolism in the left inferior temporal gyrus were not correlated with any of CSF measures in the other groups. MCI4 had a negative correlation between brain volume and P-tau_181P, whereas MCI3 showed a trend toward a negative correlation between brain volume and Aα_1 - 42.

CSF biomarkers showed no correlations with cognitive test scores in the overall MCI subjects from the ADNI1 dataset (Fig. 6B). In ADNI-MCI3, ADAS-cog negatively correlated with tau and with P-tau_181P. ADNI-MCI4 also showed a correlation between ADAS-cog and tau.

DISCUSSION

The present stratification in amnestic MCI patients identified four distinct clusters. The key to these findings is the choice of biomarkers. We selected imaging biomarkers on the basis of prior knowledge. Downstream topographical markers such as hippocampal atrophy on MRI and cortical hypometabolism on FDG-PET may be valuable for detection andquantification of disease progression or for determination of disease stages [40]. Why did this combination of three biomarkers work?

First, grey matter loss in the left hippocampus is primarily important for clustering amnestic MCI patients. A meta-analysis reported that left hippocampal atrophy is a robust biomarker for distinguishing converters from nonconverters [27]. Second, hypometabolism in the left inferior temporal gyrus on FDG-PET was an essential biomarker for discrimination of MCI4 from the other clusters. MCI4 and ADNI-MCI4 showed left-dominant asymmetrical patterns of atrophy and hypometabolism (Fig. 3), similar to those reported in recent work [41, 42]. MCI4 also had significant cognitive decline, suggesting association between the left ITG and cognitive impairment [29, 43]. Laterality also seems to be important for prediction of progression. Third, decrease in brain volume, which possibly reflects ventricular enlargement, is also important for distinguishing MCI3. Previous studies showed relationship between ventricular enlargement and AD progression of MCI [44 –48].

Hypometabolism in the posterior cingulate gyrus is considered to be a signature pattern of early AD/MCI [49, 50]. MCI2 showed greater hypometabolism in the posterior cingulate gyrus. MCI3 (n = 5) had smaller reduction in glucose metabolism in the posterior cingulate gyrus than MCI2 (n = 28) (Fig. 3A). Although this might be because the number of subjects in MCI3 was small, ADNI-MCI3 (n = 12) similarly showed less hypometabolism in the posterior cingulate gyrus than ADNI-MCI2 (n = 40) and ADNI-MCI4 (n = 11) (Fig. 3B).

Although below the statistical significance threshold, MCI1 and MCI2 particularly showed a tendency toward occipital hypometabolism (Fig. 3A). Dementia with Lewy bodies (DLB) typically shows occipital lobe hypometabolism [51]. MCI1 as well as the other groups included no DLB cases. One participant in the SEAD-J study developed DLB after inclusion, but this subject was excluded from this analysis on the basis of our exclusion criteria as mentioned above. One possible explanation for this occipital hypometabolism might be that cognitively normal elderly controls from a single-center study were used for comparison.

The relationship between these imaging biomarkers and CSF biomarkers, such as CSF amyloid-α_1 - 42 (Aα_1 - 42), CSF total tau (tau), and CSF tau phosphorylated at the threonine 181 (P-tau_181P) in this stratification, is also intriguing. To support our findings, we carried out an additional analysis with the ADNI dataset. Stratification in 79 ADNI MCI subjects resulted in four clusters (Table 3) similar to those resulted from the SEAD-J dataset (Table 2). Of these subgroups, ADNI-MCI4 was at risk for the most rapid progression to AD (Fig. 4B). This cluster had the lowest mean CSF Aα_1 - 42 level. All subjects in this subgroup were “CSF Aα_1 - 42 positive”, which was defined as a CSF Aα_1 - 42 level below a cutoff level (192 pg/mL) [52] (Fig. 2C). ADNI-MCI4 had a correlation between brain volume and P-tau_181P, whereas ADNI-MCI3 showed a tendency toward a correlation between brain volume and Aα_1 - 42 (Fig. 6A), suggesting that the underlying CSF pathophysiology for each group might be different. In ADNI-MCI1, left ITG metabolism correlated with CSF biomarkers (Fig. 6A). Recent work showed significant direct effects of alterations in CSF Aα_1 - 42 levels on hypometabolism in the middle/inferior temporal gyrus in MCI [53].

ADNI-MCI4 had a significantly lower mean score in Boston Naming Test than any of the other clusters. Poorer spontaneous naming was associated with the left ITG [54]. Asymmetrical patterns of atrophy and hypometabolism in ADNI-MCI4 (Fig. 3B) were similar to those found in logopenic variant of primary progressive aphasia (PPA) rather than bilateral patterns of reduced metabolism and grey matter volume in AD [55], although none of the subjects developed PPA. This finding from our cross-sectional study suggests that longitudinal progression of temporal lobe hypometabolism in the AD spectrum might initially occur in the left hemisphere followed by the right side with disease progression. Longitudinal studies on this subgroup might be needed for further evidence.

ADNI-MCI4 also showed significantly worse scores in Digit Symbol and Trail Making Test Part B than ADNI-MCI1. Recent work reported that orientation was associated with temporal lobe glucose metabolism and episodic memory function including performance on Digit Symbol test, and also associated with rapid functional decline and worse prognosis [56].

ADNI-MCI3, which had the greatest left hippocampal atrophy, showed negative correlations between ADAS-cog and injury markers tau and P-tau_181P (Fig. 6B), suggesting that injury markers such as CSF tau and P-tau_181p and hippocampal atrophy were associated with cognitive decline [57].

Correlation profiles in the overall subjects and in each cluster were different (Fig. 6). Regardless of the type of biomarker for stratification, an inconsistent subgroup was found in both previous work [10] and our study, and therefore stratification is important.

SEAD-J subjects showed worse cognitive performance than ADNI (Table 1), consistent with previous work [23], and had greater atrophy on MRI (Fig. 3A). Most SEAD-J participants were classified as late MCI [11]. These findings may be associated with different inclusion criteria [23]. In spite of these differences, the present stratification demonstrated similar clusters in different cohorts. Stratification with a smaller number of biomarkers might result in better generalization performance.

Our results also suggest that stratification can improve prediction accuracy (Fig. 5). MCI2, which showed a typical early AD/MCI pattern of hypometabolism, especially benefited from the stratification. One possible limitation of this study was the small sample size of our datasets.

In summary, the present stratification in amnestic MCI identified a subgroup at risk for rapid progression, with left-dominant asymmetrical patterns of atrophy and hypometabolism. Hypometabolism in the left ITG on FDG-PET can benefit this MCI subtype. Our findings have implications for refinement of the concept of MCI and can contribute to subject selection in clinical trials. Our imaging results also suggest that stratification is useful for elucidation of the pathophysiological mechanisms of disease progression. Although group analysis is useful, intersubject variability should be taken into account for interpretation of group means. Intersubject variability is not a type of noise [58]. Unsupervised learning approaches can reveal statistically invisible minorities among MCI subjects that may contribute to understanding of the syndrome.

Footnotes

ACKNOWLEDGMENTS

The authors appreciate the efforts and contributions of those who engaged in the subjects’ care and the collection of FDG-PET and MRI images and clinical reports.

This work was supported by a Health Labor Sciences Research Grant from the Ministry of Health, Labor, and Welfare of Japan (H17-Tyojyu-023) and the Research Funding for Longevity Sciences from the National Center for Geriatrics and Gerontology, Japan (20-1). This work was partly supported by the following: Grant-in-Aid for Scientific Research (C) 15K09920 from the Japan Society for the Promotion of Science (JSPS) (to N.O.), Grant-in-Aid for Scientific Research on Innovative Areas, “Glial assembly: A new regulatory machinery of brain function and disorders” (to K.O., N.O., and H.F.), and Development of BMI Technologies for Clinical Application under the Strategic Research Program for Brain Sciences by the Ministry of Education, Culture, Sports, Science, and Technology of Japan (to N.O.).

Authors’ disclosures available online ().

Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Disease Cooperative Study at the University of California, San Diego. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of SouthernCalifornia.

Appendix

References

Petersen

, Doody

, Kurz

, Mohs

, Morris

, Rabins

, Ritchie

, Rossor

, Thal

, Winblad

(2001) Current concepts in mild cognitive impairment. Arch Neurol 58, 1985–1992.

Yaffe

, Petersen

, Lindquist

, Kramer

, Miller

(2006) Subtype of mild cognitive impairment and progression to dementia and death. Dement Geriatr Cogn Disord 22, 312–319.

Fischer

, Jungwirth

, Zehetmayer

, Weissgram

, Hoenigschnabl

, Gelpi

, Krampla

, Tragl

(2007) Conversion from subtypes of mild cognitive impairment to Alzheimer dementia. Neurology 68, 288–291.

Ganguli

, Snitz

, Saxton

, Chang

C-CH

, Lee

C-W

, Vander Bilt

, Hughes

, Loewenstein

, Unverzagt

, Petersen

(2011) Outcomes of mild cognitive impairment by definition: A population study. Arch Neurol 68, 761–767.

Dubois

, Albert

(2004) Amnestic MCI or prodromal Alzheimer’s disease?. Lancet Neurol 3, 246–248.

Cuingnet

, Gerardin

, Tessieras

, Auzias

, Lehéricy

, Habert

M-O

, Chupin

, Benali

, Colliot

(2011) Automatic classification of patients with Alzheimer’s disease from structural MRI: A comparison of ten methods using the ADNI database. Neuroimage 56, 766–781.

Zhang

, Wang

, Zhou

, Yuan

, Shen

(2011) Multimodal classification of Alzheimer’s disease and mild cognitive impairment. Neuroimage 55, 856–867.

Young

, Modat

, Cardoso

, Mendelson

, Cash

, Ourselin

, Alzheimer’s Disease Neuroimaging Initiative (2013) Accurate multimodal probabilistic prediction of conversion to Alzheimer’s disease in patients with mild cognitive impairment. Neuroimage Clin 2, 735–745.

Ganguli

, Dodge

, Shen

, DeKosky

(2004) Mild cognitive impairment, amnestic type: An epidemiologic study. Neurology 63, 115–121.

10.

Nettiksimmons

, DeCarli

, Landau

, Beckett

, Alzheimer’s Disease Neuroimaging Initiative (2014) Biological heterogeneity in ADNI amnestic mild cognitive impairment. Alzheimers Dement 10, 511–521.

11.

Ota

, Oishi

, Ito

, Fukuyama

, SEAD-J Study Group (2014) A comparison of three brain atlases for MCI prediction. J Neurosci Methods 221, 139–150.

12.

Ota

, Oishi

, Ito

, Fukuyama

, SEAD-J Study Group (2015) Alzheimer’s Disease Neuroimaging, Initiative Effects of imaging modalities, brain atlases and feature selection on prediction of Alzheimer’s disease. J Neurosci Methods 256, 168–183.

13.

Ito

, Fukuyama

, Senda

, Ishii

, Maeda

, Yamamoto

, Ouchi

, Ishii

, Okumura

, Fujiwara

, Kato

, Arahata

, Washimi

, Mitsuyama

, Meguro

, Ikeda

, SEAD-J Study Group (2015) Prediction of outcomes in mild cognitive impairment by using 18F-FDG-PET: A multicenter study. J Alzheimers Dis 45, 543–552.

14.

Folstein

, Folstein

, McHugh

(1975) “Mini-mental state.” A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 12, 189–198.

15.

Homma

, Fukuzawa

, Tsukada

, Ishii

, Hasegawa

, Mohs

(1992) Development of a Japanese version of Alzheimer’s Disease Assessment Scale (ADAS). Jpn J Geriatr Psychiary 3, 647–655.

16.

Sullivan

(1996) Estimates of interrater reliability for the Logical Memory subtest of the Wechsler Memory Scale-Revised. J Clin Exp Neuropsychol 18, 707–712.

17.

Yesavage

, Brink

, Rose

, Lum

, Huang

, Adey

, Leirer

(1982) Development and validation of a geriatric depression screening scale: A preliminary report. J Psychiatr Res 17, 37–49.

18.

Nyunt

MSZ

, Fones

, Niti

, Ng

T-P

(2009) Criterion-based validity and reliability of the Geriatric Depression Screening Scale (GDS-15) in a large validation sample of community-living Asian older adults. Aging Ment Health 13, 376–382.

19.

Ashburner

, Friston

(2005) Unified segmentation. Neuroimage 26, 839–851.

20.

Rajapakse

, Giedd

, Rapoport

(1997) Statistical approach to segmentation of single-channel cerebral MR images. IEEE Trans Med Imaging 16, 176–186.

21.

Tohka

, Zijdenbos

, Evans

(2004) Fast and robust parameter estimation for statistical partial volume models in brain MRI. Neuroimage 23, 84–97.

22.

Ashburner

(2007) A fast diffeomorphic image registration algorithm. Neuroimage 38, 95–113.

23.

Kawashima

, Ito

, Kato

, SEAD-J Study Group (2012) Inclusion criteria provide heterogeneity in baseline profiles of patients with mild cognitive impairment: Comparison of two prospective cohort studies. BMJ Open 2, e000773.

24.

Dukart

, Mueller

, Horstmann

, Vogt

, Frisch

, Barthel

, Becker

, Möller

, Villringer

, Sabri

, Schroeter

(2010) Differential effects of global and cerebellar normalization on detection and differentiation of dementia in FDG-PET studies. Neuroimage 49, 1490–1495.

25.

Mevel

, Desgranges

, Baron

J-C

, Landeau

, la Sayette de

, Viader

, Eustache

, Chételat

(2007) Detecting hippocampal hypometabolism in mild cognitive impairment using automatic voxel-based approaches. Neuroimage 37, 18–25.

26.

Tzourio-Mazoyer

, Landeau

, Papathanassiou

, Crivello

, Etard

, Delcroix

, Mazoyer

, Joliot

(2002) Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage 15, 273–289.

27.

Ferreira

, Diniz

, Forlenza

, Busatto

, Zanetti

(2011) Neurostructural predictors of Alzheimer’s disease: A meta-analysis of VBM studies. Neurobiol Aging 32, 1733–1741.

28.

Morbelli

, Piccardo

, Villavecchia

, Dessi

, Brugnolo

, Piccini

, Caroli

, Frisoni

, Rodriguez

, Nobili

(2010) Mapping brain morphological and functional conversion patterns in amnestic MCI: A voxel-based MRI and FDG-PET study. Eur J Nucl Med Mol Imaging 37, 36–45.

29.

Scheff

, Price

, Schmitt

, Scheff

, Mufson

(2011) Synaptic loss in the inferior temporal gyrus in mild cognitive impairment and Alzheimer’s disease. J Alzheimers Dis 24, 547–557.

30.

Tibshirani

, Walther

, Hastie

(2001) Estimating the number of clusters in a data set via the gap statistic. J R Stat Soc Series B Stat Methodol 63, 411–423.

31.

Van Essen

(2012) Cortical cartography and Caret software. Neuroimage 62, 757–764.

32.

Lindsey

, Ryan

(1998) Tutorial in biostatistics methods for interval-censored data. Stat Med 17, 219–238.

33.

Fan

R-E

, Chang

K-W

, Hsieh

C-J

, Wang

X-R

, Lin

C-J

(2008) LIBLINEAR: A library for large linear classification. J Mach Learn Res 9, 1871–1874.

34.

Metz

(1978) Basic principles of ROC analysis. Semin Nucl Med 8, 283–298.

35.

Fawcett

(2004) ROC graphs: Notes and practical considerations for researchers. Mach Learn 31, 1–38.

36.

Huang

, Ling

(2005) Using AUC and accuracy in evaluating learning algorithms. IEEE Trans Knowl Data Eng 17, 299–310.

37.

Robin

, Turck

, Hainard

, Tiberti

, Lisacek

, Sanchez

J-C

, Müller

(2011) pROC: An open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 12, 77–.

38.

Cohen

(1973) Eta-squared and partial eta-squared in fixed factor Anova designs. Educ Psychol Meas 33, 107–112.

39.

R Development Core Team (2013) R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing, Vienna.

40.

Dubois

, Feldman

, Jacova

, Hampel

, Molinuevo

, Blennow

, DeKosky

, Gauthier

, Selkoe

, Bateman

, Cappa

, Crutch

, Engelborghs

, Frisoni

, Fox

, Galasko

, Habert

M-O

, Jicha

, Nordberg

, Pasquier

, Rabinovici

, Robert

, Rowe

, Salloway

, Sarazin

, Epelbaum

, de Souza

, Vellas

, Visser

, Schneider

, Stern

, Scheltens

, Cummings

(2014) Advancing research diagnostic criteria for Alzheimer’s disease: The IWG-2 criteria. Lancet Neurol 13, 614–629.

41.

Murayama

, Ota

, Kasanuki

, Kondo

, Fujishiro

, Fukase

, Tagaya

, Sato

, Iseki

(2015) Cognitive dysfunction in patients with very mild Alzheimer’s disease and amnestic mild cognitive impairment showing hemispheric asymmetries of hypometabolism on (18) F-FDG PET. Int J Geriat Psychiatry 27 Epub 2015 Mar 27.

42.

Cerami

, Rosa Della

, Magnani

, Santangelo

, Marcone

, Cappa

, Perani

(2015) Brain metabolic maps in mild cognitive impairment predict heterogeneity of progression to dementia. Neuroimage Clin 7, 187–194.

43.

Hirono

, Mori

, Ishii

, Imamura

, Tanimukai

, Kazui

, Hashimoto

, Takatsuki

, Kitagaki

, Sasaki

(2001) Neuronal substrates for semantic memory: A positron emission tomography study in Alzheimer’s disease. Dement Geriatr Cogn Disord 12, 15–21.

44.

Nestor

, Rupsingh

, Borrie

, Smith

, Accomazzi

, Wells

, Fogarty

, Bartha

(2008) Alzheimer’s Disease Neuroimaging, Initiative. Ventricular enlargement as a possible measure of Alzheimer’s disease progression validated using the Alzheimer’s disease neuroimaging initiative database. Brain 131, 2443–2454.

45.

Chou

Y-Y

, Lepore

, Avedissian

, Madsen

, Parikshak

, Hua

, Shaw

, Trojanowski

, Weiner

, Toga

, Thompson

, Alzheimer’s

Disease Neuroimaging Initiative

(2009) Mapping correlations between ventricular expansion and CSF amyloid and tau biomarkers in 240 subjects with Alzheimer’s disease, mild cognitive impairment and elderly controls. Neuroimage 46, 394–410.

46.

Evans

, Barnes

, Nielsen

, Kim

, Clegg

, Blair

, Leung

, Douiri

, Boyes

, Ourselin

, Fox

, Alzheimer’s

Disease Neuroimaging Initiative

(2010) Volume changes in Alzheimer’s disease and mild cognitive impairment: Cognitive associations. Eur Radiol 20, 674–682.

47.

Apostolova

, Green

, Babakchanian

, Hwang

, Chou

Y-Y

, Toga

, Thompson

(2012) Hippocampal atrophy and ventricular enlargement in normal aging, mild cognitive impairment (MCI), and Alzheimer disease. Alzheimer Dis Assoc Disord 26, 17–27.

48.

Apostolova

, Babakchanian

, Hwang

, Green

, Zlatev

, Chou

Y-Y

, DeCarli

, Jack

, Petersen

, Aisen

, Cummings

, Toga

, Thompson

(2013) Ventricular enlargement and its clinical correlates in the imaging cohort from the ADCS MCI donepezil/vitamin E study. Alzheimer Dis Assoc Disord 27, 174–181.

49.

Mosconi

, Tsui

, Herholz

, Pupi

, Drzezga

, Lucignani

, Reiman

, Holthoff

, Kalbe

, Sorbi

, Diehl-Schmid

, Perneczky

, Clerici

, Caselli

, Beuthien-Baumann

, Kurz

, Minoshima

, De

Leon MJ

(2008) Multicenter standardized 18F-FDG PET diagnosis of mild cognitive impairment, Alzheimer’s disease, and other dementias. J Nucl Med 49, 390–398.

50.

Minoshima

, Giordani

, Berent

, Frey

, Foster

, Kuhl

(1997) Metabolic reduction in the posterior cingulate cortex in very early Alzheimer’s disease. Ann Neurol 42, 85–94.

51.

Sinha

, Firbank

, O’Brien

(2012) Biomarkers in dementia with Lewy bodies: A review. Int J Geriat Psychiatry 27, 443–453.

52.

Shaw

, Vanderstichele

, Knapik-Czajka

, Clark

, Aisen

, Petersen

, Blennow

, Soares

, Simon

, Lewczuk

, Dean

, Siemers

, Potter

, Lee

VMY

, Trojanowski

, Alzheimer’s

Disease Neuroimaging Initiative

(2009) Cerebrospinal fluid biomarker signature in Alzheimer’s disease neuroimaging initiative subjects. Ann Neurol 65, 403–413.

53.

Dowling

, Johnson

, Gleason

, Jagust

, Alzheimer’s

Disease Neuroimaging Initiative

(2015) The mediational effects of FDG hypometabolism on the association between cerebrospinal fluid biomarkers and neurocognitive function. Neuroimage 105, 357–368.

54.

Melrose

, Campa

, Harwood

, Osato

, Mandelkern

, Sultzer

(2009) The neural correlates of naming and fluency deficits in Alzheimer’s disease: An FDG-PET study. Int J Geriat Psychiatry 24, 885–893.

55.

Madhavan

, Whitwell

, Weigand

, Duffy

, Strand

, Machulda

, Tosakulwong

, Senjem

, Gunter

, Lowe

, Petersen

, Jack

, Josephs

(2013) FDG PET and MRI in logopenic primary progressive aphasia versus dementia of the Alzheimer’s type. PLoS One 8, e62471.

56.

Sousa

, Gomar

, Goldberg

(2015) Neural and behavioral substrates of disorientation in mild cognitive impairment and Alzheimer’s disease. Alzheimers Dement (N Y) 1, 37–45.

57.

van Rossum

, Vos

SJB

, Burns

, Knol

, Scheltens

, Soininen

, Wahlund

L-O

, Hampel

, Tsolaki

, Minthon

, L’italien

, van der Flier

, Teunissen

, Blennow

, Barkhof

, Rueckert

, Wolz

, Verhey

, Visser

(2012) Injury markers predict time to dementia in subjects with MCI and amyloid pathology. Neurology 79, 1809–1816.

58.

Zilles

, Amunts

(2013) Individual variability is not noise. Trends Cogn Sci 17, 153–155.

Prediction of Alzheimer’s Disease in Amnestic Mild Cognitive Impairment Subtypes: Stratification Based on Imaging Biomarkers

Abstract

Keywords

INTRODUCTION

MATERIALS AND METHODS

Datasets

SEAD-J participants

ADNI participants

MRI acquisition and preprocessing

18F-FDG-PET image acquisition and preprocessing

Feature extraction

Measures for clustering

Clustering

Statistical parametric mapping

Survival analysis

Prediction of progression to AD

Correlation analysis

Statistical analysis

RESULTS

Baseline differences between the SEAD-J and the ADNI cohorts

Stratification in the SEAD-J cohort

Stratification in the ADNI cohort

Heterogeneity in patterns of atrophy and hypometabolism

Time to progression

Effect of stratification on prediction

Correlation analysis

DISCUSSION

Footnotes

ACKNOWLEDGMENTS

Appendix

References

¹⁸F-FDG-PET image acquisition and preprocessing