Uncovering Diverse Mechanistic Spreading Pathways in Disease Progression of Alzheimer’s Disease

Data demographics

Data used in the preparation of this article were obtained from the ADNI database (https://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 MRI, 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 https://www.adni-info.org.

ADNI enrolls participants between the ages of 55 and 90 who are recruited at 57 sites in the United States and Canada. After obtaining informed consent, participants undergo a series of initial tests that are repeated at intervals over subsequent years, including clinical evaluation, neuropsychological tests, genetic testing, lumbar puncture, and MRI and PET scans. There are four phases of the ADNI study (ADNI1, ADNI-GO, ADNI2, and ADNI3). Some participants were carried forward from previous phases for continued monitoring, while new participants were added with each phase to further investigate the progression of AD.

In this work, we put the spotlight on the effect of AD-related pathology on memory (MEM) decline and executive function (EF) decline, respectively. It is worth noting that we derived composite scores for MEM and EF using data from the ADNI neuropsychological battery using item response theory (IRT) methods. The formation of ADNI-MEM was complicated by the use of different word lists in the Rey Auditory Verbal Learning Test (RAVLT) and the ADAS-Cog, and by Logical Memory I data missing by design [29].

For each mechanic pathway, we select subjects from ADNI data based on the following two criteria: 1) each subject should have both two modalities and clinical outcome scores; 2) the data collection of upstream AD biomarker should be prior to the downstream biomarker in the mediation analysis. For example, the amyloid-PET scan is required to be earlier than FDG-PET scan in the A⟶N pathway, in order to characterize the mediation effect of amyloid on the metabolism level. We show the demographic information of participants for A-N pathway, A-T pathway, and T-N pathway in Table 1. The diagnostic information is shown at the bottom of Table 1, in which CN and MCI are the abbreviations for cognitive normal and mild cognitive impairment, respectively.

Image processing

Each subject has T1-weighted MRI, amyloid-PET, tau-PET, and FDG-PET scans. Destrieux atlas [30] is used to parcellate each brain into 160 regions of interest (ROIs), which consist of 148 cortical regions (frontal lobe, insula lobe, temporal lobe, occipital lobe, parietal lobe, and limbic lobe) and 12 sub-cortical regions (left and right hippocampus, caudate, thalamus, amygdala, globus pallidum, and putamen). The major image preprocessing steps include 1) skull strip and tissue segmentation on T1-MRI; 2) brain parcellation using the Destrieux atlas; 3) constructing cortical surface using free-surfer [31]; and 4) calculating the standardized uptake value ratio (SUVR) where the cerebellum is selected as the reference region.

Covariates and confounders in the statistical inference

The initial set of covariates consists of 160 regional SUVRs from amyloid, 160 regional SUVRs from tau, and 160 regional SUVRs from FDG-PET, respectively. The confounders include biological sex, education, and age. To remove the effect of those potential confounders, we regress the value of 480 ROIs and MEM/EF score on them separately and independently. The residuals of regional SUVRs and MEM/EF scores are utilized as covariates and outcomes in the downstream integrative factor regression model to select the key regional SUVRs associated with MEM/EF score. The selected SUVRs are used in a further mediation analysis to construct the pathways of how they affect the MEM/EF scores. Mediators found in mediation analysis and the corresponding exposures are further involved as covariates in the Cox’s proportional hazard model [32] to study how they affect the time to cognitive decline. APOE4 status, sex, and age are adjusted in the Cox’s model.

Statistical approach

In general, our statistical inference consists of three major steps. Firstly, we apply a multi-variate variable selection method to reduce the number of brain regions for region-to-region mediation analysis. Secondly, we apply a set of mediation analyses to all possible pathways between regional AT[N] biomarkers. To make the analysis computationally tractable, each mediation analysis consists of one upstream regional biomarker, one downstream regional biomarker, and clinical outcomes (shown in Fig. 2). Thirdly, we apply survival analysis to examine the hazard ratio of identified critical brain regions (i.e., manifesting significant direct/indirect effect on clinical outcome in the mediation analysis) to the cognitive decline in each pathway analysis.

OPEN IN VIEWER

Fig. 2

Three SEMs of AT[N] pathway in this paper.

Variable selection

We use the integrative factor regression model [33] to select potentially influential variables (regional SUVRs) to be included in the second-stage mediation analysis. For each modality (amyloid, tau, and FDG-PET), we control the confounding effects on the observed regional SURVs by regressing raw values of ROIs on the confounders. Each two of them will be used as the exposure and mediator variables of the outcome (MEM/EF) in the mediation analysis. We first regress the outcome on every two of the three modalities (A&F, A&T, or T&F) and select regional SUVRs in these modalities that are associated with the outcome. Let x _m be the vector of variables from the mth modality for m = 1,  2, and y be the residual of regressing MEM/EF on confounders. We consider a linear regression model of y = ∑ m = 1 2 x m ′ β m * + ∈ , where β m * is the effects of variables from the mth modality, and ∈ is the error term. To account for the within- and across-modality correlations, the integrative factor model assumes that x _m has a decomposition of x _m  = Λ _m f _m  + u _m , where f _m is the vector of latent factors in the mth modality with their loadings as Λ _m , and u _m is the idiosyncratic error. Then, we follow [33] to use Bai and Ng’s method [34] to estimate the number of latent factors and a principal component analysis (PCA) method to obtain the estimators of f _m and u _m . After that, we utilize the integrative factor model to regress y on the estimators of f _m  and u _m with a SCAD (smooth clipped absolute deviation) penalty [35] to select non-zero components of β m * , which identifies the regional SUVRs in the mth modality associated with the MEM/EF.

Mediation analysis

After variable selection, we design the following region-to-region structural equation models (SEMs) to elucidate the mechanistic pathways between AT[N] biomarkers, as shown in Fig. 2.

SEM 1: Test whether FDG acts as a mediator of amyloid on MEM/EF score (A→F→MEM/EF).

The input to each SEM consists of selected regional SUVRs and the clinical phenotype. Each SEM has an exposure variable, a mediator, and an outcome. To infer the causal relation, we need these three components to be ordered by time i.e., T _exposure  < T _mediator  < T _outcome . In SEM 1 and SEM 2, we use the latest FDG/tau measurement (potential mediator) after amyloid (exposure) measurement but prior to MEM/EF measurement (outcome). Besides, subjects without records on APOE4, education, sex, and age are removed. This gives sample sizes of 669 and 137 in SEM 1 and SEM 2, respectively. Since the temporal changes of regional amyloid/FDG/tau SUVRs are slow, when performing SEM 3, we relax the time restriction by allowing FDG measurement to be no more than one year later than tau measurement, instead of being strictly earlier than tau measurement. This gives us a sample size of 41 in SEM 3.

To illustrate the mediation analysis, let’s take SEM 1 as an example. We let x be the exposure variable (i.e., the selected regional amyloid SUVR), y be the independent variable (i.e., the composite score of memory), m be the mediator (i.e., the selected regional FDG SUVR). Three regression models are constructed to evaluate the mediation effect:

Model 1: y = γ₁ + τ₁x + ∈ ₁

Model 2: m = γ₂ + αx + ∈ ₂

Model 3: y = γ₃ + τ₂x + βm + ∈ ₃

where γ₁, γ₂, and γ₃ are the intercept terms, and ∈₁, ∈₂, and ∈₃ are the error terms. In Model 1, τ₁ measures the total exposure-outcome effect. In Model 2, α measures the exposure-mediator effect. In Model 3, τ₂ measures the direct exposure-outcome effect after adjusting for the mediator effect and β measures mediator-outcome effect after adjusting for the exposure effect.

In the cliché of mediation analysis, αβ = τ₁ - τ₂ measures the mediation/indirect effect of exposure on the outcome and we perform the Sobel test to assess its significance. The null hypothesis of the non-existence of indirect effect is αβ = 0. The Sobel test statistic is Z = α ˆ β ˆ / SE ˆ , where α ˆ and β ˆ are the Least Squares estimators of α and β with its standard error as SE ˆ = α ˆ 2 σ ˆ β 2 + β ˆ 2 σ ˆ α 2 , σ ˆ α 2 and σ ˆ β 2 are the estimated variances of α ˆ and β ˆ . The Z-statistics is then compared with the standard normal distribution to determine p-value. These p-values are then adjusted by the Storey’s method [29] to control the False Discovery Rate (FDR). The FDR-controlled p-values are reported in Tables 2 and 3.

Survival analysis

We further study how the exposure and mediator variables affect the time from baseline to the time of Mini-Mental State Examination (MMSE) score falling below 28 using the Cox’s proportional hazard model. Subjects with MMSE score lower than 28 at baseline are excluded from the analysis and subjects with MMSE score greater than 28 during the whole study are treated as right-censored. All exposure variables and significant mediators are included in the Cox’s model. APOE4 status, sex, and age are also adjusted in the model.

Effect of AT[N] pathways on memory decline

Variable selection

The variable selection results are shown in Fig. 3. For amyloid SUVRs, five regions (left inferior temporal gyrus, left/right putamen, right pallidum, and right inferior angular parietal gyrus) have been identified with significant association to the clinical outcome. For FDG-PET images, we have identified six brain regions after variable selection, which include the right superior parietal gyrus, right subparietal sulcus, right occipital temporal gyrus, left posterior cingular gyrus, left inferior frontal lobe, and left hippocampus. Due to the small sample size, we have not found any brain regions showing a significant statistical association between tau SUVRs and outcome scores. In this regard, we only apply SEM 1 for memory and executive function.

OPEN IN VIEWER

Fig. 3

Selected brain regions from amyloid-PET (left) and FDG-PET (right) w.r.t. memory score.

A-[N]-MEM mechanistic pathway analysis

Since five brain regions in the amyloid modality and six brain regions in the FDG modality have been identified in the region selection step, there are 30 region-to-region A⟶N⟶MEM pathways. We summarize the mediation analysis results in Tables 2 and 3. First, 5 out of 30 possible region-to-region pathways (highlighted in red in Table 2) manifest a direct effect of amyloid accumulation on the decrease of memory composite score at the significance level of p = 0.05. Second, we have found 10 region-to-region pathways show mediated effects on memory decline, that is, the concentration of amyloid deposition affects memory status through the other brain region that shows changes of metabolism level. In Table 3, we show the overall indirect effect sizes (1st column) and the effect sizes (with the adjusted p-values) on A⟶N link (2nd column) and N⟶MEM link (3rd column). It is worth noting that the effect of increased amyloid burden at the left inferior temporal gyrus and right angular gyrus on memory decline is mediated by reduced metabolism levels.

We have included APOE4 status in SEM1, where ‘1’ stands for APOE4 carrier (with at least one ɛ4 allele) and ‘0’ otherwise. APOE4 shows a strong effect on ten mediated pathways shown in Table 3 at a significance level p = 0.05. In Fig. 4, we show three mediated region-to-region pathways associated with the inferior temporal gyrus (left panel) and another three mediated pathways associated with angular gyrus, where the average adverse effect size of APOE4 carriers is –3.406.

OPEN IN VIEWER

Fig. 4

Graphic demonstration of mediated pathways in structural equation model A⟶F⟶MEM.

Effect of AT[N] pathways on the decline of executive function

Variable selection

The variable selection results with respect to EF score are shown in Fig. 5, where we list the selected region names for amyloid and FDT-PET in the left and right panels of Table 5. There are 14 brain regions showing the association between amyloid burden and EF score. Similarly, there are 24 brain regions showing the association between metabolism level (measured by FDG-PET) and EF score. Again, we have not found any brain regions showing association in the tau∼EF relationship, due to the small sample size. In the following, we apply mediation analysis of SEM 1 to 14 × 24 = 336 A⟶N⟶EF pathways.

OPEN IN VIEWER

Fig. 5

Selected brain regions from amyloid-PET (left) and FDG-PET (right) images with respect to EF score.

OPEN IN VIEWER

Table 5

Selected brain regions with significant amyloid∼EF relationship (left) and significant metabolism∼EF relationship (right)

IOGS, inferior occipital gyrus and sulcus; dPCC, posterior-dorsal part of cingulate gyrus; SIPJ, sulcus intermedius primus of Jensenl MOS&LS, middle occipital sulcus and lunatus sulcus; LIG&CSI, long insular gyrus and central sulcus of the insula; IS&TPS, intraparietal sulcus and transverse parietal sulci; SOS&TOS, superior occipital sulcus and transverse occipital sulcus; TFGS, transverse frontopolar gyri and sulci; pMCC, middle-posterior part of the cingulate gyrus and sulcus; PRLS, posterior ramus of the lateral sulcus; CSI, circular sulcus of the insula; MO-TS&LS, medial occipito-temporal sulcus and lingual sulcus.

A-[N]-EF mechanistic pathway analysis

We summarize the mediation analysis result in Tables 6 and 7. First, 64% region-to-region pathways (highlighted in red in Table 6) manifest a direct effect of amyloid accumulation on the decrease of EF composite score at the significance level p = 0.05. Second, we have found 69 (20.5%) region-to-region pathways show mediated effects on memory decline, that is, the concentration of amyloid deposition affects executive function performance through the other brain region that shows changes of metabolism level. In Table 7, we show the overall indirect effect sizes (1st column) and the effect sizes on A⟶N link (2nd column) and N⟶EF link (3rd column), where we use color to indicate the effect sizes. The effect sizes marked in blue are negative effects, while red ones are positive effects. The darker the color is, the stronger the effect. Since eight brain regions with amyloid are involved in the 69 mediated pathways, we display the mediated pathways in Fig. 6, where the yellow dot represents each of these eight brain regions and red dots represent the mediated brain regions with significant metabolism∼EF relationship. We use ‘*’, ‘**’, and ‘***’ to denote the FDR-controlled p < 0.05, p < 0.001, and p < 0.0001 on A⟶N and N⟶EF links.

OPEN IN VIEWER

Table 6

Mediation analysis of A⟶N⟶EF. The selected region names for amyloid and FDT-PET correspond to the variable selection results in Fig. 5 and are listed in Table 5

“M” stands for the mediation effect of amyloid burden (row) on EF decline is mediated by reduced metabolism level (column). Otherwise, we display the direct effect size of the amyloid burden on EF decline, where the significant direct effect (p < 0.05) is marked in red.

OPEN IN VIEWER

Table 7

The effect sizes for 69 region-to-region A⟶N⟶EF pathways showing a mediation effect

dPCC, posterior-dorsal part of cingulate gyrus; SIPJ, sulcus intermedius primus of Jensen; IS&TPS, intraparietal sulcus and transverse parietal sulci; SOS&TOS, superior occipital sulcus and transverse occipital sulcus; MO-TS&LS, medial occipito-temporal sulcus and lingual sulcus.

OPEN IN VIEWER

Fig. 6

Graphic presentation of all 69 mediated A⟶N⟶EF region-to-region pathways.

Mechanistic roles of identified brain regions

We found a majority of identified critical brain regions exhibit a direct effect on cognitive decline. The identified brain regions on which the amyloid burdens exert an indirect effect on cognitive decline are oftentimes mediated by the reduction of metabolism levels at other brain regions. We conduct the same statistical analysis to the memory and executive function score separately. The absence of region overlap between memory-specific and EF-specific pathways that manifest the indirect effect on cognitive decline indicates that the pathophysiological pathways associated with memory and EF domain might have distinct spatial patterns. We have not identified any T⟶N and A⟶T mechanistic pathways that manifest an indirect effect on any clinical domain.

Role of non-modifiable risk factors in diverse mechanistic pathways

As frequently reported in the literature, we found APOE4 status plays a role not only in the indirect effect but also in the direct effect of amyloid burden on cognitive decline. The sex difference has not been detected in all mediation analyses. We have not discovered the contribution of lifestyle factors (such as education level) is significant in the identified A⟶N⟶MEM and A⟶N⟶EF pathways.

Survival analysis for cognitive decline on the mechanistic pathways

We found that having two APOE ɛ4 alleles significantly increases the risk of developing AD in our survival model, where the hazard ratio is 1.60 regarding memory problems and 1.75 for the decline of executive functions, both at the significant level of p < 0.05.

In our main result, we only use FDG-PET as the neurodegeneration biomarker. Since cortical thickness from structure MRI is also widely used as a [N] biomarker in many studies, it is worthwhile to investigate the A⟶N⟶MEM and A⟶N⟶EF pathways between regional amyloid burden and cortical thickness. Using the same data analysis workflow, we first apply our multi-variate feature selection method to screen critical brain regions. Figure 7 displays the most significant brain regions that manifest significance in the amyloid∼MEM relationship (top-left), cortical thickness∼MEM relationship (top-right), amyloid∼EF relationship (bottom∼left), and cortical thickness∼EF relationship (bottom right).

OPEN IN VIEWER

Fig. 7

Brain regions survival after the variable selection method in the context of amyloid∼MEM relationship (top-left), cortical thickness∼MEM relationship (top-right), amyloid∼EF relationship (bottom∼left), and cortical thickness∼EF relationship (bottom right). The color indicates the anatomical lobe parcellation.

Next, we examine whether the effect of amyloid cascade on memory score is mediated by the reduced cortical thickness. Although most brain regions in Fig. 7 top-left exert a direct effect on the declined memory performance, we have identified eight mediated pathways at the significance level of 0.05, where the total indirect effect sizes, effect sizes on A⟶N pathway, and effect size on N⟶MEM are summarized in Table 9. Regarding the mediation analysis of region-to-region A⟶N⟶EF, all the brain regions manifest a direct effect on the decline of executive function performance. Comparing the mediation analysis between using the metabolism level and cortical thickness, it is apparent that the causal-like relationship between amyloid cascade and reduced metabolism level is stronger than that between amyloid and reduced cortical thickness.

Our work has several clinical benefits. First, the region-to-region mechanistic pathway between AT[N] biomarkers allows us to advance our understanding of the etiology for AD. One possible application is to stratify the heterogeneity of decline trajectories and discover latent subtypes associated with characteristic pathophysiological pathways. Second, the elucidated role of risk factors in the survival analysis provides useful information in developing personalized intervention plans for AD and AD-related dementias.

Related works

Understanding the causal relationship between AD biomarkers is an important topic in the AD field. Mediation analysis has been widely used to investigate whether these biomarkers act as mediators, linking the pathogenic processes involved in AD [36–39]. For example, mediation analysis has been used in investigating whether age-related changes in cognition are partially mediated by the presence of neuropathology and neurodegeneration (measured by CSF biomarker of Aβ and tau) [37]. Their study found age effects on cognitive decline are fully mediated by disease and neurodegeneration variables. However, the age effect on baseline cognition manifests differences between memory and executive function domains. In this regard, we stratify our mediation studies on regional imaging biomarkers for memory and EF domains. In our result, we found the A⟶[N]⟶MEM and A⟶[N]⟶EF exhibit different spatial patterns region-to-region mediation pathways, which supports previous studies that the relationship between AD biomarkers and cognitive decline is domain-specific.

Our hypothesis of the causal relationship between AT[N] biomarkers is rooted in the biological synergy between Aβ and tau in AD [38] and the NIA-AA research framework [16]. Furthermore, our work is inspired by the pioneering work on staging of Aβ, tau, regional atrophy rates, and cognitive change [40]. In their study, researchers discovered a direct correlation between amyloid accumulation, as measured by CSF biomarkers, and regional atrophy rates as well as memory loss in the cognitively normal group. Interestingly, this correlation was found to occur without any mediating factors. During the early stage of MCI, the presence of Aβ was observed to impact memory through the atrophy of the hippocampus. Subsequently, Aβ continued to affect memory, with the specific mediation of medial temporal atrophy, and without the involvement of intermediate tau. Our results (A⟶[N]⟶MEM medication pathways) align closely with these findings. As shown in Table 9, we have discovered that the amyloid burden at orbital gyrus, precuneus gyrus, temporal gyrus, and rectus gyrus manifest indirect effect on memory performance via hippocampus. In addition, the effect of Aβ accumulation at inferior temporal gyrus on memory decline is mediated by the atrophy at superior temoral sulcus. Comparing our results with the findings in [40], our results provide a more fine-grained brain mapping of A⟶[N]⟶MEM medication pathways since our data analysis is performed on regional SUVR level.

By examining the mediating role of amyloid, tau, and neurodegeneration biomarkers, researchers can better understand the causal pathways and identify potential targets for intervention. However, current work mainly uses global measurements such as CSF biomarkers and whole-brain concentration levels from PET scans. In this work, we investigate the causal relationship at the granularity of brain region, which provides much richer spatial information to link brain function and network organization. To address the new challenge of higher data dimensionality, we applied a multi-variate statistical approach [33] to select the most relevant brain regions.

Our current analysis has several limitations. First, the requirement of having complete paired imaging biomarkers in the mediation analysis raises the issue of a small sample size. In light of this, we have not found the region-to-region pathway related to tau pathology. One possible solution is to include other public datasets, such as OASIS [41]. Second, a well-designed structural equation model is needed to take the gene-by-environment interactions into account, which sets the stage for understanding the diversity in cognitive decline and health disparities in the incidence and prevalence of AD.