Multicenter Resting State Functional Connectivity in Prodromal and Dementia Stages of Alzheimer’s Disease

INTRODUCTION

Today, early diagnosis of Alzheimer’s disease (AD) in prodromal and dementia stages is supported by the detection of disease characteristic biomarkers, such as amyloid accumulation or tau pathology [1–3]. In addition, people in prodromal or clinical at risk stages of AD such as people with mild cognitive impairment (MCI) [4] or subjective cognitive decline (SCD) [5] show characteristic pattern of metabolic decline in cortical areas belonging to cortical functional networks, particularly the default mode network [6], in ¹⁸FDG-PET studies [7–10]. In comparison to ¹⁸FDG-PET, functional MRI during resting state (rs-fMRI) is more widely available, less costly and has no radiation exposure. A large range of studies has shown that seed based as well as data driven analyses can reveal distinct intrinsic connectivity networks, including the default network, based on low frequency resting state fluctuations of the blood oxygenation level dependent signal (BOLD) (for review, see [11]). The BOLD signal is considered a measure of neuronal activity, similar to FDG-PET metabolism, and intrinsic connectivity networks from rs-fMRI have been suggested to carry diagnostically useful information in MCI and AD dementia [12]. However, the coupling between neuronal metabolism and activity and the oxygen extraction fraction, the physiological driver of the BIOLD signal, is complex, varying with blood flow, cerebral pathology, and activity [13]. Thus, findings on the accuracy of group discrimination based on rs-fMRI between prodromal and dementia stages of AD and healthy controls vary even between monocenter studies [14–17].

A step further in the evaluation of rs-fMRI as a diagnostic tool is the assessment of the multicenter stability of this measure across different scanners. In healthy people the pattern of functional connectivity was found to be affected by different scanners [18–21]. Studies on the effect of multi scanner variability on the diagnostic value of rs-fMRI in AD are still scarce. In a previous study using retrospectively collected multicenter rs-fMRI data [22] we found a medium sized effect of scanner on between group differences. The discriminatory accuracy between AD dementia and MCI patients and healthy controls after cross validation was about 80 to 70%, respectively, rendering multicenter rs-fMRI inferior in diagnostic accuracy compared with hippocampus volume [22], a well-established structural MRI marker of AD [23].

We expected that the retrospective collection of fMRI data without controlling for the acquisition parameters increased multicenter variability and decreased group discrimination compared with harmonized rs-fMRI acquisitions across scanners. Therefore, in the present study we determined the diagnostic accuracy of two frequently employed rs-fMRI indices, functional connectivity [16] and amplitude of low-frequency fluctuation (ALFF) [24], from a prospective cohort study with a harmonized acquisition protocol and prospectively employed scan quality control. Extending previous studies, we included not only people with AD dementia, MCI or healthy controls, but also a group of people with SCD which represent a preMCI at risk stage of AD [5]. We hypothesized that multicenter variability of functional connectivity and ALFF would be lower compared with previous multicenter studies using harmonized retrospective data collection. In addition, we hypothesized that the cross-validated diagnostic accuracy would be higher in the prospective cohort than in previous multicenter studies which were based on retrospective data. To address the issue of multicollinearity of functional imaging features, we employed the technique of elastic net regression [25] which was found superior for feature selection in several previous studies, including fMRI, FDG-PET, structural MRI and amyloid PET data of people with AD and MCI, compared with classical stepwise selection [26–29].

MATERIALS AND METHODS

RESULTS

Between-center variability

The temporal signal to noise ratio and the standardized DVARS across the six sites included in the analysis are shown in Fig. 1. Both indices showed wide overlap between centers and less variation in comparison to the corresponding values from the retrospectively collected psymri data (shown as boxplot on the left for both indices), suggesting a low between center variability in these key measures of scan quality.

OPEN IN VIEWER

Fig.1

Multicenter data quality indices. a) Mean whole brain temporal Signal to Noise Ratio (tSNR). b) Standardized DVARS. Box plots for the quality indices with open circles indicating outlying values for the six centers (denoted center I to center VI). For comparison, the boxplots to the left indicate the corresponding values from the retrospectively collected fMRI data of the psymri cohort.

Group discrimination

As shown in Fig. 2, the cross-validated discrimination accuracy for ALFF and functional connectivity ranged between 56% and 81% in the people with AD dementia compared to healthy controls, was 57% in the MCI cases and ranged between 56% and 57% in the people with SCD. Using functional connectivity measures based on other atlases than the Greicius atlas [44] yielded essentially identical results with on average slightly lower AUC for the Craddock [45] (– 3%) and Schaefer-Yeo [46] (– 4%) atlases (Supplementary Figure 1). Notably, the Schaefer-Yeo atlas [46] performed substantially worse (– 15% AUC) for amyloid-positive AD dementia cases versus amyloid-negative controls, which might result from the missing subcortical regions in that specific atlas. With exception of the AD dementia cases, the bootstrapped 95% confidence intervals (CI) of the cross validated accuracy levels included the value of 50%, suggesting that the degree of diagnostic accuracy did not significantly exceed random guessing accuracy in all groups except the AD dementia cases. For comparison, the diagnostic accuracy based on the hippocampus volume was only 54% in people with SCD, but reached 75% in people with MCI and 88% in people with AD dementia. When we repeated the analysis with a dummy coded center covariate forced into the model, the diagnostic accuracy remained essentially unchanged.

OPEN IN VIEWER

Fig.2

Cross-validated areas under the ROC curves. Bar diagrams for the group discrimination accuracy as measured by cross-validated AUC with 95% confidence intervals. SCD, subjective cognitive decline; HC, healthy controls; MCI, mild cognitive impairment; AD, Alzheimer’s disease; ALFF, amplitude of low-frequency fluctuation; FC, functional connectivity; HPvol, hippocampal volume.

When we assessed the diagnostic accuracy in the MCI subgroup of 13 cases with an abnormal Aβ₄₂/Aβ₄₀ ratio in CSF and the 25 healthy controls with a normal Aβ₄₂/Aβ₄₀ ratio, we found a cross-validated AUC for ALFF and functional connectivity between 69% and 73%, with the lower level of the 95% CI including 50%. For comparison in this subsample, hippocampus volume reached a cross-validated AUIC of 78%, where the 95% CI included 50% as well.

When we determined the 50% most frequently selected functional connectivity networks using elastic net regression for the AD dementia cases versus controls comparison, these networks involved functional connectivity between regions from the dorsal DMN, the anterior and posterior salience network, the language network, the left executive network, the visuo-spatial and the sensori-motor network (Table 2 and Fig. 3). Since the other comparisons did not significantly exceed random guessing accuracy, we did not assess the regional distribution of discriminatory features for these comparisons.

OPEN IN VIEWER

Fig.3

Resting state network components contributing to the discrimination between AD patients and controls. Resting state network components whose functional connectivity was contributing to the discrimination between AD patients and controls projected on a structural MRI scan in MNI space. Axial sections go through MNI z coordinates – 38, – 26, – 14, – 2, 10, 22, 34, and 46 mm. Colors code the sum of the beta weights for each region.

OPEN IN VIEWER

Table 2

Functional connectivity on regional and network level discriminating AD patients from controls

Beta	network 1	network 2	Corresponding AAL region 1	Corresponding AAL region 2	mean FC controls	mean FC AD
– 1.90	dDMN	sensorimotor	l. hippocampus + parahippocampal gyrus	l. + r. cerebellum	0.01	– 0.19
– 1.25	dDMN	post. salience	r. angular gyrus	l. insula	0.23	0.04
– 1.01	language	precuneus	l. middle temporal gyrus	l. inferior parietal lobule + angular gyrus	0.35	0.21
– 0.89	dDMN	dDMN	l. + r. thalamus	l. hippocampus + parahippocampal gyrus	0.50	0.32
– 0.85	language	post. salience	r. middle and sup. temporal gyrus	r. mid. cingulum	0.30	0.11
– 0.78	dDMN	dDMN	l. + r. superior medial frontal gyrus + ant. cingulum	l. angular gyrus	0.22	0.10
– 0.47	dDMN	precuneus	l. + r. thalamus	l. post. cingulum	0.21	0.03
0.15	ant. salience	dDMN	l. insula	Angular R	– 0.25	– 0.06
0.38	dDMN	visuospatial	l. + r. thalamus	r. angular, supramarginal, and postcentral gyrus, and parietal inferior lobule	0.08	0.23
0.52	language	LECN	l. inf. orbitofrontal	L superior and inferior frontal gyrus	– 0.17	– 0.03
0.52	dDMN	sensorimotor	l. + r. superior medial frontal gyrus + ant. cingulum	l. thalamus	– 0.14	0.01
0.55	dDMN	post. salience	r. angular gyrus	l. precuneus	– 0.18	– 0.03
0.83	auditory	sensorimotor	r. thalamus	r. thalamus	– 0.17	– 0.04
0.92	dDMN	post. salience	r. angular gyrus	r. mid. cingulum	– 0.28	– 0.14
1.00	dDMN	visuospatial	l.+ r. precuneus + post. cingulum	r. cerebellum	– 0.17	– 0.04
1.01	ant. salience	visuospatial	r. middle frontal gyrus	r. angular, supramarginal, and postcentral gyrus, and parietal inferior lobule	– 0.05	0.08
1.25	dDMN	RECN	l. + r. thalamus	r. middle frontal gyrus	0.33	0.45
1.90	dDMN	post. salience	l. + r. thalamus	r. middle frontal gyrus	– 0.09	0.04
2.40	dDMN	post. salience	l. + r. superior medial frontal gyrus + ant. cingulum	l. thalamus	– 0.10	0.02

r., right; l., left; ant., anterior; post., posterior; FC, functional connectivity; dDMN, dorsal default mode network; RECN, right executive network; LECN, left executive network.

When we used a support vector machine classifier, results were similar to the elastic net classification (see Supplementary Figure 2).

DISCUSSION

Compared with previous multicenter studies using retrospective data acquisition, in this prospective multicenter rs-fMRI analysis we found less pronounced variations of global scan quality parameters across sites, including measures of temporal signal to noise ratio that we had found to be particularly sensitive to between scanner variability in a previous study [22]. Despite the smaller inter-scanner variability, between group effects of resting state functional connectivity and ALFF reached only small effect size in prodromal and at risk stages of AD, such as SCD and MCI. Only in manifest stages of AD dementia, we found 80% accuracy of group discrimination, where the cross-validated level of accuracy was significantly above chance level. This level of accuracy is similar to the level of between group discrimination in previous monocenter studies comparing AD dementia and healthy controls using cross validated diagnostic accuracy. In contrast, the discrimination between MCI patients and controls and even between MCI patients with pronounced episodic memory impairment and controls did not significantly differ from chance level. The effect was even smaller for the people with SCD. These results were not sensitive to the atlas selection and were replicated using support vector machine as alternative classifier.

The level of accuracy found in our amnestic MCI cases is lower than the levels of accuracy found in previous mono-center studies when these did not consider the cross-validated accuracy [14, 15], and two studies using cross-validation [16, 17]. One recent study based on multicenter fMRI data from ADNI, found a level of 77% cross-validated accuracy for the discrimination between 54 MCI cases and 54 controls based on functional connectivity, without explicitly addressing the multicenter characteristics of the data [52]. Our findings agree with a previous multicenter study which was based on retrospective collected fMRI data sets, where we found cross validated accuracy of group discrimination between MCI patients and controls between 66% and 72% [22]. Together these data suggest that despite a reduction of multicenter variability through a prospectively harmonized acquisition protocol the usefulness of resting state functional connectivity for the discrimination between amnestic MCI cases and healthy controls is very limited. The level of accuracy was higher when comparing the subset of MCI cases who were amyloid positive with the controls who were amyloid negative. The level of accuracy here ranged between 66% and 69%; with the large confidence interval likely reflecting the small number of cases. One previous fMRI study showed steeper age related decline in DMN, posterior cingulate and frontoparietal resting state network connectivity in CSF amyloid positive versus amyloid negative cognitively healthy older people [53]. Consistently, previous rs-fMRI studies showed that resting state connectivity measures were related to CSF levels of amyloid across diagnostic groups, including cognitively healthy older people, people with MCI and AD dementia patients [54–58]. Taking these previous findings into account, our finding would indicate that rs-fMRI functional connectivity more closely reflects the underlying molecular pathology than the clinical phenotype of amnestic MCI; this is different to hippocampus volume that reached similar levels of accuracy both in amnestic MCI cases and the amyloid positive subgroup, suggesting that hippocampus volume reflects the amnestic phenotype of MCI, relatively independent from the underlying molecular pathology.

The number of previous monocenter studies in people with SCD is limited. One study used independent component analysis to decompose the rs-fMRI data of 13 healthy controls, 8 people with Alzheimer’s disease dementia, 21 with MCI and 16 with SCD into distinct functional connectivity networks and assessed the association of these networks with neuropsychological performance in different cognitive tests. Group differences in functional connectivity were not reported [59]. To our knowledge, the current study is the first multicenter study to report diagnostic accuracy of functional connectivity from rs-fMRI to discriminate people with SCD from healthy controls. Interestingly, two previous monocenter studies reported not only decreases, but also significant increases of functional connectivity measures in SCD people compared to controls, one study comprising 25 SCD cases and 60 controls [60] and the other study comprising 25 people with SCD and 29 controls [61].

The lack of a significant discrimination between SCD cases and controls in our study may be related to two factors: First, in the subset of SCD cases where CSF was available, the proportion of amyloid positive cases was low with only 19%, albeit comparable with previous studies in subjective memory complainers with a rate of 21% amyloid positive cases [62]. If we interpolate this proportion to the entire group of SCD cases this would indicate that the majority of the SCD cases was not in a preclinical state of AD so that the lack of discrimination may simply reflect the lack of underlying neuropathology. The few previous studies on resting state fMRI in SCD [59–61] did not control for amyloid status. Unfortunately, with only 7 amyloid positive SCD cases we were not able to determine a meaningful level of accuracy. Secondly, following the previous observation that SCD cases showed not only decreased, but also increased connectivity [60, 61], people with SCD may present with a mixture of increases and decreases of functional connectivity that together may lead to a poor diagnostic accuracy when searching for between group differences across the entire brain.

In addition, relevant for all comparisons multicenter acquisition may additionally reduce between group effects through sources of inter-scanner variance that remain even after prospective harmonization of acquisition protocols and strict image quality control. The negative finding in the SCD group is based on a relatively large sample of 90 cases compared to 80 controls with a strict matching in respect to age, sex distribution and MMSE score performance. These are the first data on the discriminatory power of rs-fMRI in SCD, however, due to the relatively high number of cases, the risk of a false negative finding is small.

When we assessed the networks and regions contributing to significant group discrimination in AD dementia, we found not only a strong involvement of default mode network (DMN) regions, including not only hippocampal and neocortical but also thalamic components of this network, but also language, executive, salience and visuo-spatial networks. This agrees with previous studies showing reduced within network connectivity in AD patients compared with controls in the DMN, ventral and dorsal salience network, executive network and a frontal-parietal network involving the precuneus [63] as well as altered between network connectivity involving the dDMN, executive, visuospatial, fronto-parietal and anterior and posterior salience networks [54, 65], including the thalamus. Indeed, we found a reduced strength of positive associations within the DMN or the DMN and the precuneus resting state network in AD patients compared with controls, and a reduced strength of negative associations between the DMN and the anterior salience network. Different to previous studies [54, 65], however, we did not a priori exclude resting state networks from the analyses. Using such a purely data driven approach, we found alterations also including networks that would be considered to be relatively spared by AD, such as the sensorimotor or auditory network. This does not imply decline of connectivity within these latter networks, but rather a change in coupling between relatively spared and more affected networks. So, for example, the negative association between sensorimotor and auditory network components became less strong in AD patients compared with controls or sensorimotor network and DMN components became more negatively associated in AD, suggesting a potential dysfunctional compensatory effect in people with AD.

A strength of our study is the large number of people with subjective cognitive decline. To our knowledge only the INSIGHT-preAD cohort [62] with 318 participants has a larger number of people with subjective cognitive decline and rs-fMRI, however, the INSIGHT-preAD cohort does not have a control group so that the analysis of the discriminatory accuracy of rs-fMRI has to await the longitudinal follow-up. A limitation of our study is that the number of available CSF samples was too small in the SCD group to allow for a meaningful comparison of the amyloid positive SCD cases with the amyloid negative controls. The DELCODE study is ongoing so that we will have access to a larger number of amyloid positive SCD cases to study the association of rs-fMRI with the presence or absence of biomarker-confirmed preclinical at risk stages of AD in the future.

A strength of the study in comparison to previous retrospective multicenter rs-fMRI studies is the strict harmonization of acquisition protocols as well as the strict implementation of scan quality controls throughout the study. Indeed, compared to our previous study based on retrospective data [22], key parameters of scan quality showed much less variability in the prospectively acquired data. Still, although the degree of variability between scanners was lower compared to the previous multicenter study, the achievable level of diagnostic accuracy was comparable to the previous study, suggesting that between scanner variability is not the main source of limited diagnostic accuracy from multicenter rs-fMRI data. Future follow-up data of this cohort will allow us to determine predictive accuracy of resting state functional connectivity for a clinically relevant functional outcome such as decline of cognitive function or conversion to dementia.

The prospective design of our study with a strict harmonization of acquisition protocols is a strength of the study when we interpret the findings in respect to future application of the technique in controlled multicenter trials. Since the outcome of our prospective study suggests a low diagnostic accuracy of rs-fMRI functional connectivity in a highly controlled setting, we expect the level of accuracy would even be lower in a less controlled routine care setting.

In summary, in a prospective multicenter resting state fMRI acquisition we found significant group discrimination between AD dementia patients and controls that was, however, inferior to the widely established measure of hippocampus volume. Functional connectivity or ALFF did not reach significant group discrimination above chance level neither in SCD nor in amnestic MCI cases. These findings suggest that measures of functional connectivity based on Pearson’s correlation or ALFF, independently of the selected atlas and the classifier, are not useful markers in prodromal stages of AD. Future multicenter studies should explore the diagnostic usefulness of alternative functional connectivity measures, such as partial correlation [66], regional homogeneity [67] or functional dynamics based on shorter time windows of the resting state scan [66] that are currently less widely used than ALFF and Pearson’s correlation metrics.