Neuroimaging markers of global cognition in early Alzheimer's disease: A magnetic resonance imaging–electroencephalography study

Abstract Introduction Magnetic resonance imaging (MRI) and electroencephalography (EEG) are a promising means to an objectified assessment of cognitive impairment in Alzheimer's disease (AD). Individually, however, these modalities tend to lack precision in both AD diagnosis and AD staging. A joint MRI–EEG approach that combines structural with functional information has the potential to overcome these limitations. Materials and Methods This cross‐sectional study systematically investigated the link between MRI and EEG markers and the global cognitive status in early AD. We hypothesized that the joint modalities would identify cognitive deficits with higher accuracy than the individual modalities. In a cohort of 111 AD patients, we combined MRI measures of cortical thickness and regional brain volume with EEG measures of rhythmic activity, information processing and functional coupling in a generalized multiple regression model. Machine learning classification was used to evaluate the markers’ utility in accurately separating the subjects according to their cognitive score. Results We found that joint measures of temporal volume, cortical thickness, and EEG slowing were well associated with the cognitive status and explained 38.2% of ifs variation. The inclusion of the covariates age, sex, and education considerably improved the model. The joint markers separated the subjects with an accuracy of 84.7%, which was considerably higher than by using individual modalities. Conclusions These results suggest that including joint MRI–EEG markers may be beneficial in the diagnostic workup, thus allowing for adequate treatment. Further studies in larger populations, with a longitudinal design and validated against functional‐metabolic imaging are warranted to confirm the results.


| INTRODUC TI ON
Alzheimer's disease (AD) is a fatal disorder that is associated with the accumulation of β-amyloid plaques and neurofibrillary tau tangles causing progressive neurodegeneration in certain cortical and subcortical regions (Hyman et al., 2012). AD invariably affects episodic memory and other complex cognitive processes, but the perceived onset and early course of the symptoms are highly subjective and depend on the individual cognitive reserve (Stern, 2012). An accurate clinical assessment of the cognitive deficits is crucial for disease staging and thus for optimal pharmacological treatment and therapy planning. Typically, cognitive impairment is assessed in a doctor-patient/-caregiver interview and neuropsychological screening tests such as the mini-mental state examination (MMSE; Folstein, Folstein, & McHugh, 1975). These tests, however, are susceptible to daily variations, and their outcome is affected by sociodemographic factors and the individual cognitive reserve (Crum, Anthony, Bassett, & Folstein, 1993). Consequently, there is a need for accurate alternatives to measure the progression of cognitive impairment in AD.
Magnetic resonance imaging (MRI) and electroencephalography (EEG) are potential surrogate in vivo measures of AD progression that are noninvasive, inexpensive, and widely available. MRI measures of regional brain volumes and cortical thickness are promising markers of both AD neuropathology and cognitive decline (Babiloni et al., 2015;Dubois et al., 2014;Scheltens, Fox, Barkhof, & De Carli, 2002).
Especially, the limbic system in the medial temporal lobe has been identified to be vulnerable to AD: Typically, the entorhinal cortex is among the regions affected in the earliest disease stages (Killiany et al., 2002). The hippocampal volume appears to be a good marker of both disease onset and its progression (Bateman et al., 2012;Den Heijer et al., 2010;Dickerson et al., 2001;Jack et al., 2010;Lo et al., 2011). Other affected brain regions include amygdala, posterior association cortex, and the cholinergic basal forebrain (Bottino et al., 2002;Braak & Braak, 1991;Dubois et al., 2014). The cortical atrophic topography is often in line with the clinical phenotype; subjects with verbal memory impairment, for instance, frequently exhibit early atrophy in the left temporal lobe (Johnson, Fox, Sperling, & Klunk, 2012).
Electroencephalography, on the other hand, is a promising tool to assess the AD-related functional disintegration of large scale brain networks such as the default mode network during resting state (Dillen et al., 2017;Horn, Ostwald, Reisert, & Blankenburg, 2014).
Despite their evident potential as noninvasive assistive tools in AD diagnose, MRI is routinely used solely for the purpose of excluding other possible causes such as vascular lesions, strategic lunar infarcts, or cerebral haemorrhages (Dubois et al., 2014;Schmidt et al., 2010), and EEG is not included in the standard diagnostic workup at all (American Psychiatric Association, 2013;Dubois et al., 2007Dubois et al., , 2010Dubois et al., , 2014Hyman et al., 2012;McKhann et al., 1984McKhann et al., , 2011Schmidt et al., 2010). A reason is that, individually, these modalities tend to lack precision in both AD diagnosis and staging. However, combining the structural and functional information from MRI and EEG in a multimodal approach has the potential to overcome the limitations of the individual modalities. This cross-sectional study systematically investigates the usefulness of different MRI and EEG measures as symptom-independent markers of the cognitive deficits in a cohort of early AD patients. We hypothesize that (a) these markers are significantly related to global cognition as measured by MMSE in early AD and that (b) joint MRI-EEG markers allow for an accurate identification of significant cognitive deficits.

| Study cohort
The study cohort consisted of 111 subjects diagnosed with probable (N = 77) or possible (N = 34) AD according to NINCDS-ADRDA 1 criteria (McKhann et al., 1984) who participated in the prospective dementia (PRODEM) study of the Austrian Alzheimer Society (Seiler et al., 2012). Enrollment criteria included the availability of a caregiver, written informed consent of participant and caregiver, no need for 24-hr care, and the absence of other physical or neurological causes of dementialike symptoms. All patients underwent a routine laboratory assessment, measurement of serum vitamin B12 and folic acid levels, as well as serologic (HIV, Lues) and thyroid testing. AD-inconsistent patterns of cerebral atrophy and lesion were excluded in T1-weighted MRI (3D MPRAGE) sequences. The dementia severity was assessed by clinical dementia rating (CDR) on a scale of 0 (no dementia) to 3 (severe dementia) (Hughes, Berg, Danziger, Coben, & Martin, 1982;Morris, 1993). Our cohort consisted of subjects with CDR scores of 0.5 (very mild dementia) or 1 (mild dementia). The global cognition was assessed by MMSE scores on a scale of 0-30 where scores below 24 indicate significant cognitive deficits (Folstein et al., 1975;Mungas, 1991). The participants' age, sex, and completed years of education were included as potential confounders (Crum et al., 1993;Kittner et al., 1986;O'Connor, Pollitt, Treasure, Brook, & Reiss, 1989).
MRI scans were processed with the FreeSurfer neuroimaging software (Version 5.3, Massachusetts General Hospital, Boston MA, USA) that included automated methods for volume segmentation and measurement of cortical thickness of various brain regions.
Briefly, the processing flow consisted of skull stripping followed by segmentation of gray/white matter and mapping of different brain structures in Talairach space (Desikan et al., 2006;Fischl et al., 2002). A detailed technical description can be found, for example, in Fischl et al., 2002 andReuter, Schmansky, Rosas, &Fischl, 2012. In previous studies, these morphometric FreeSurfer procedures were validated against manual mapping (Fischl et al., 2002;Kuperberg et al., 2003;Salat et al., 2004) and proved to be reliable across scanner types and field strengths (Han et al., 2006;Reuter et al., 2012).
The regional volume and mean cortical thickness of the frontal, parietal, left and right temporal, and occipital lobe of the cerebral cortex -the major control of higher cognitive function -were employed as potential markers of cognitive impairment. The left and right temporal lobes were hereby assessed separately since the left temporal lobe has been described to be more affected in early AD stages (Johnson et al., 2012). In addition, we included the volume of entorhinal cortex, hippocampus, and amygdala due to their vulnerability to atrophy in early AD (Juottonen, Laakso, Partanen, & Soininen, 1999;Poulin, Dautoff, Morris, Barrett, & Dickerson, 2011). All MRI measures were normalized by the individual total intracranial volume to account for anatomical differences between the subjects (Fischl, 2012).
More specifically, each recording was downsampled to 128 Hz to reduce computational cost and bandpass filtered in the range of 1-30 Hz (type 1 FIR filter, 60 dB). Eye artifacts were removed by constrained independent component analysis using the EOG (Lu & Rajapakse, 2006). Cardiac artifacts were corrected by a modified Pan-Tompkins algorithm using the ECG (Waser & Garn, 2013). The resulting EEG samples were divided in 2-s epochs with 1-s overlap (Blanco, Garcia, Quiroga, Romanelli, & Rosso, 1995). Epochs with residual artifacts were identified by a thresholding algorithm (Waser et al., 2017), visually validated by an EEG expert and omitted from further analyses.
The EEG measures listed in Table 1 were computed epoch-and channel-wise. Global markers were derived separately for REC and ENC by taking the average over all channels and epochs (cf. in a nonlinear way. Figure 1 illustrates the topographic logic of the marker assessment.

| Diagnostic utility assessment
The same marker subset was then used to distinguish subjects with MMSE ≥ 24 from those with MMSE < 24 by a machine learning classification approach (see Supporting information Material Section B.3). We used a support vector machine (SVM) with radial basis function kernel to separate the two groups (Cortes & Vapnik, 1995).
Parameters were hereby tuned using a grid search over defined parameter ranges. Classification was performed with leave-one-out cross-validation and evaluated by its sensitivity (true positive rate), specificity (true negative rate), and accuracy. Group differences regarding age, sex, and years of education were tested by χ²-and Kruskal-Wallis tests.  Table 3 lists the MRI and EEG markers and their respective (non-normalized) mean values ± standard deviation. In relating individual markers with MMSE scores, we found significant results for the volumes of the parietal and the left temporal lobe, the spectral power in theta, alpha1, beta1, and beta2, as well as aMI. More specifically, reduced lobar volume, increased portions of spectral power in a low frequency range, and increased aMI were all associated with lower MMSE scores. Figure 2 shows the between-marker correlation (Pearson's r) color-coded from blue (−1) to red (1) and tagged with a (*) in case of a high statistical significance (p < 0.01). We observed a widespread positive correlation between the MRI markers that were most pronounced in the cortical thickness markers. The relative spectral power in the delta and theta frequency bands was negatively correlated with the power in the higher beta1 and beta2 bands. The aMI showed high correlations with beta1 and beta2 power, indicating that information processing mechanisms were mainly reflected by a large portion of high-frequency EEG oscillations. As for MRI-EEG correlations, reduced volume and cortical thickness of the parietal lobe were significantly associated with high delta-power, the parietal volume was positively correlated with beta1 power as well as the left temporal volume with interhemispheric coherence.

| Marker combinations
The following marker subset was selected to be included in the regression model: the left temporal volume, the cortical thickness of frontal, parietal and occipital lobe, and the spectral power in theta and alpha2. The regression was significant (p < 0.001) and the combined regressors explained 38.2% of the variation in MMSE scores.  Table 4 summarizes the results of discriminating the two groups (MMSE ≥ 24 and MMSE < 24) using the same marker subset from MRI and EEG individually, as well as in a modality-combined manner. Using the latter, individuals were classified with an accuracy of 84.7% (sensitivity 92.1%, specificity 75.0%). Figure

| D ISCUSS I ON
The first study question sought to determine whether there was a significant relation between the MRI-EEG markers and the severity of cognitive impairment in early AD. Prior studies have reported that individual MRI measures of regional volume and cortical thickness as well as abnormal EEG patterns are related with the cognitive status. Here, only few individual markers were significantly correlated with MMSE scores. However, the combined MRI-EEG markers were well associated with the subjects' cognitive status. By measuring the left temporal volume, the cortical thickness of frontal, parietal and occipital lobe, and the spectral power in the theta and alpha2 frequency bands, 38.2% of MMSE variation was explained. In line with previous studies, the inclusion of age, sex, and education improved these results (Crum et al., 1993;Kittner et al., 1986). O'Connor et al. (1989) suggested that also the sociocultural background, though  (Bottino et al., 2002), and 87%-99% for measures of cortical volume and thickness (Du et al., 2007). However, considering the small samples of these studies, the results need to be interpreted with caution. Similar results have been found in EEG studies, with EEG markers yielding accuracies from 76% to 85% (Jelic & Kowalski, 2009;Triggiani et al., 2016). Thereby, it seems that the MRI and EEG modalities might be complementary due to an incomplete overlap in subjects regarding MRI and EEG abnormalities (Strijers et al., 1997). This view got further support by a recent study showing superior AD-control classification accuracy (90%) for combined MRI and EEG markers as compared to the individual modalities as well. These data suggest that MRI-EEG markers have potential as accurate cognitive staging tools.
One interesting finding is that four of the six selected markers were measures of cortical lobe atrophy. Especially, atrophy patterns in the left temporal, frontal, and parietal lobe are in accord with previous studies (Du et al., 2007;Hwang et al., 2016). Hartikainen et al. (2012) also found occipito-parietal cortical thinning in AD. Caution is advised; however, for the cortical thickness, measures are highly inter-correlated and any conclusion on topographic patterns needs further validation.
In contrast to earlier findings, the medial temporal lobe structures hippocampus, amygdala, and entorhinal cortex were not included in the TA B L E 3 Original magnetic resonance imaging (MRI) and electroencephalography (EEG) marker values (mean ± standard deviation) and linear regression analysis: The slope β refers to the linear regression coefficient of the normalized markers as regressors and log-normalized MMSE scores as outcome, while correcting for age, sex, and completed years of education as covariates F I G U R E 2 Analysis of marker intercorrelation: The Pearson's correlation is shown color-coded and a (*) indicates a significant intercorrelation as tested by two-tailed Student's t test (α = 0.01) final model. A possible explanation might be the strong innervation of these structures and the cerebral cortex and, consequently, that limbic atrophy was implicitly included through the coarser cortical measures. Another reason might be that the included EEG markers reflected the changes in the limbic structures. It also seems possible that these relatively small brain structures were subject to measurement bias in contrast to the greater cortical lobes. The most plausible explanation, however, is that these brain structures have already been affected long before AD was diagnosed and that, during early AD, they do not vary as much.
Among the abnormal EEG patterns, increased theta and decreased alpha rhythms, commonly referred to as EEG slowing, were the most significant markers. These findings are consistent with previous studies and the role of theta and alpha rhythms in cognitive performance: Klimesch (1999)  A limitation of this study is the sole use of MMSE scores as measure of the global cognitive status. AD typically impairs cognitive complex domains, the sequence and severity of impairment varies from patient to patient, even more so in the earliest disease stages.
Another well-described issue of the MMSE is its susceptibility to demographic factors such as age and education. In the current study, we tried to overcome this limitation by including demographic information as covariates. However, a single MMSE cutoff to distinguish stages of cognitive impairment is thus problematic, and it is important to bear in mind the possible resulting bias. By using more elaborate cognitive tests that are less sensitive to demographic factors, further research should be undertaken to investigate which markers reflect deficits in which cognitive domain.

| CON CLUS ION
The purpose of this study was a systematic assessment of the use-