Tau, atrophy, and domain‐specific cognitive impairment in typical Alzheimer's disease

Abstract

INTRODUCTION

A granular understanding of the mechanisms linking tau pathology to cognitive decline in Alzheimer's disease is crucial. We investigate mediating effects of medial temporal lobe (MTL) and neocortical neurodegeneration on tau‐induced domain‐specific cognitive impairment in amyloid‐beta (Aβ) positive cognitively normal and impaired adults.

METHODS

We assessed magnetic resonance imaging–derived MTL and neocortical volume/thickness and ¹⁸F‐Flortaucipir positron emission tomography in 319 Aβ‐positive individuals. Cognitive functions across six domains were isolated by adjusting for other cognitive measures.

RESULTS

MTL tau correlated with memory subdomains, neocortical tau with executive function, and both with semantic fluency. Specific structural measures partially mediated these tau–cognition associations: Brodmann area 35 mediated tau–immediate and tau–delayed recall, posterior hippocampus tau–recognition, and inferior temporal cortex tau–semantic fluency associations.

DISCUSSION

Our findings provide a nuanced understanding of region‐specific macrostructural atrophy as one pathway of tau‐induced cognitive changes, aligning with known tau spread patterns. Additionally, isolating cognitive functions is a promising approach for future research.

Highlights

• Medial temporal lobe tau was related to memory domains; neocortical tau to executive function.

• Both tau positron emission tomography measures were associated with semantic fluency.

• Specific regional atrophy partially mediated tau‐induced cognitive changes.

• Other mechanistic links between tau and cognitive subdomains require investigation.

• Isolated cognitive domains should be explored as future avenues of research.

1. BACKGROUND

Alzheimer's disease (AD) is characterized by cognitive decline and the accumulation of amyloid‐beta (Aβ) plaques and neurofibrillary tau tangles (NFTs). ¹ Tau tangles are closely linked to cognitive decline. ² , ³ , ⁴ , ⁵ Cortical tau accumulation occurs first in the medial temporal lobe (MTL), prior to spreading across the neocortex, ⁶ , ⁷ hypothesized to be driven by Aβ plaque accumulation. ⁸ , ⁹ This pathological progression parallels the clinical course of AD from episodic memory loss, linked to the MTL, to more widespread cognitive impairment, in semantic knowledge, executive function, language, and visuospatial domains, ¹⁰ linked to neocortical tau. ³ Tau pathology may negatively impact cognition through several mechanisms, including neuronal loss and synaptic dysfunction. ¹¹ Given the MTL's critical role in AD, it is important to establish how neurodegeneration in fine‐grained MTL regions may drive tau‐induced changes in cognitive subdomains to better understand early pathophysiological changes.

MTL tau positron emission tomography (PET) uptake has been related to decreased episodic memory performance ¹² , ¹³ , ¹⁴ while neocortical tau‐PET uptake has been associated with decreased global cognition. ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ Additionally, tau‐PET uptake in specific regions has been associated with different cognitive domains: uptake in temporal parietal regions was associated with episodic memory, uptake in fronto‐temporal regions with executive function, and uptake in temporal‐occipital regions with semantic fluency. ¹⁶ , ¹⁸ , ¹⁹ This indicates the utility of tau‐PET to study more regional directional relationships with cognitive change in AD.

Extant work has investigated neurodegeneration, measured with morphometric magnetic resonance imaging (MRI), as a mechanistic link between tau pathology and cognitive impairment. One study found that MTL structural measures mediated the tau–memory decline association in Aβ– and Aβ+ individuals, ²⁰ although a study including various AD phenotypes did not replicate this. ¹⁸ Moreover, mediating effects of distinctive regional atrophy were shown for the association between tau‐PET and other cognitive domains. ¹⁸ For example, full mediations were reported for occipital and temporal lobe volume in tau‐PET–visuospatial functioning and tau‐PET–fluency associations, respectively. ¹⁸ While these findings support neurodegeneration as a potential mechanistic link, the contrasting findings highlight the complexity of the tau pathology, neurodegeneration, and cognitive impairment relationship.

Past work investigating mediating effects on cognitive decline have generally used composite measures, and thereby potentially overlooked distinct neurobiological substrates underlying cognitive subdomain impairments. For instance, episodic memory subdomains involve different regions in the brain, ¹⁶ , ²¹ , ²² , ²³ , ²⁴ , ²⁵ for example, fronto‐parietal regions for immediate and hippocampus for delayed recall. ²² Thus, the interpretation of cognitive tests may be obscured by the grouping of different functions. Additionally, more specific isolated measures of cognitive domains are often lacking, because cognitive tests often assess multiple cognitive domains (e.g., a learning test requires attention). Of note, performance on the Trail‐Making Test Part B, which measures executive functioning, is commonly adjusted for Part A, measuring processing speed. ²⁶ This approach could be applied in other cognitive domains. Finally, past work investigated coarse MTL atrophy measures, despite evidence of differential vulnerability of MTL subregions to tau pathology ⁶ , ²⁷ , ²⁸ , ²⁹ and their mappings to memory subdomains. ²¹ , ²³ Combining isolated measures of cognitive subdomains with fine‐grained MTL structural measures will enable a more precise characterization of biological substrates and mechanisms of decline of cognitive subdomains in AD.

Here, we aim to gain a more granular understanding of how tau pathology leads to domain‐specific cognitive decline. Thus, we map associations between MTL and neocortical tau PET, fine‐grained MRI‐based atrophy in MTL subregions and neocortical regions implicated in AD, and measures of isolated cognitive (sub)domains in typical AD across the disease spectrum. We focus on two questions: (1) What regional MRI‐based structural measures mediate the association between tau on domain‐specific cognitive impairments? and (2) Which regions found in (1) are the primary mediators? Investigated cognitive subdomains include episodic memory (immediate, late, delayed recall, and recognition), semantic fluency, and executive function. Importantly, the different cognitive subdomains are adjusted for specific subdomains aiming to isolate their function. Ultimately, this may enhance the utility of psychometric testing to improve tracking of sensitive and specific outcome measures in clinical trials.

2. METHODS

2.1. Participants

For this cross‐sectional study, data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu) were obtained. 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 mild cognitive impairment (MCI) and early AD. For up‐to‐date information, see www.adni‐info.org. Inclusion and diagnosis criteria for ADNI have been described previously. ³⁰ , ³¹ ADNI includes amnestic MCI and dementia cases due to AD that would generally be classified as “typical” by design.

There were four inclusion criteria specific to this study, aiming to include cognitively unimpaired (CU) and impaired individuals on the continuum of typical AD (CU, MCI, and AD dementia). First, individuals had to be Aβ+ based on a cut‐off for amyloid PET (Florbetapir or F‐lorbetaben PET) standardized uptake value ratio (SUVR) closest to baseline tau‐PET scan (see below for details). Second, the participants had undergone neuropsychological testing with no missing data for the cognitive tests of interest. Third, the participants had to have one tau‐PET scan and a T1‐weighted MRI. Fourth, the time difference between tau‐PET and MRI and between Aβ. and tau‐PET scan was limited to two standard deviations (SDs) above and below the mean (i.e., maximum 302 days between tau‐PET and MRI; and maximum 185 days between Aβ‐ and tau‐PET, but only if Aβ‐PET occurred after the tau‐PET scan), to reduce a mismatch between modalities that could be introduced by long intervals between scans. Using these criteria led to a sample size of 319 participants in the present study. All these individuals had an MCI and AD etiology “due to Alzheimer's disease.”

We also used data from 131 cognitively normal Aβ– controls (sex: female n = 73 [55.7%]; age: mean [SD] = 69.6 [5.45]; education: mean [SD] = 16.9 [2.24]) to compute age, sex, and education‐adjusted z scores for cognitive performance. Thus, no direct analyses with performed with these individuals.

2.2. Neuropsychological assessment

Cognitive assessment closest to the tau PET scan date was used for this study (median = 15 days). Five neuropsychological tests were selected: the Auditory Verbal Learning Test (AVLT), the Alzheimer's Disease Assessment Scale Cognitive subscale (ADAS‐Cog), category fluency, letter fluency, and the Trail‐Making Test (TMT).

RESEARCH‐IN‐CONTEXT

1. Systematic review: The authors reviewed the literature using traditional sources (e.g., PubMed, Scopus). While neurodegeneration has been investigated as a link between tau and cognition, most research used coarse measures of neurodegeneration and/or used composite scores of cognition. Thus, knowledge is limited for fine‐grained measures of neurodegeneration and cognitive subdomains.

2. Interpretation: Our findings provide a nuanced understanding of region‐specific macrostructural atrophy as one pathway of tau‐induced changes in isolated cognitive functions, implicating specific structural measures as partial mediators of different tau–cognition associations.

3. Future directions: This study increases our understanding of how fine‐grained macrostructural measures map to tau‐induced changes in cognitive functions, warranting further research to clarify its usefulness to track subtle changes in clinical trials. Additionally, an approach to isolate cognitive functions showed biologically plausible results, indicating it is a promising avenue for further research.

The AVLT is a 15‐word list learning task ³² , ³³ in which individuals learn a list of words over several trials. Different trials were used to calculate immediate recall and late immediate recall. Immediate recall is based on trial 1 + 2 and late immediate recall on trial 4 + 5.

Delayed recall was additionally calculated. This measure was partially based on the AVLT delayed recall after 30 minutes and on the 10‐item word list from the ADAS‐Cog. ³⁴ The latter involves three rounds of learning a list of 10 words and immediate recall, delayed recall (after a 5 minute interval) is completed. Number of errors was recorded and converted to number of correct words (10 – number of errors) to match the AVLT score. Number of words recalled on both of the measures were summarized and transformed into a z score based on the Aβ– controls.

Recognition memory was based on the AVLT. Fifteen studied words along with fifteen foils were determined to be either studied or not on the prior list. To account for incorrect “recognition” of non‐studied items (“false alarms”), a measure of discriminability, d prime (d’), was calculated using a standard formula to convert these values: Hits = (#Hits + 0.5)/(#studied items + 1) and False Alarms = (#FA + 0.5)/(#unstudied items + 1) to avoid undefined values of d’. ³⁵

The category fluency test is a test of semantic fluency and involves naming as many words of a given category (animal, vegetables) as possible within 1 minute. Letter fluency, a test of lexical fluency, involves naming as many words starting with a given letter (“f”) as possible within 1 minute. The total number of correct named words is recorded for fluency measures.

The TMT is a test of executive functioning, asking participants to either connect numbers on a page in sequential order (TMT Part A [TMT‐A]) or connect numbers and letters alternating in sequential order (TMT Part B [TMT‐B]) as fast as possible. ³⁶ The score of these tests is the time (in seconds) taken to complete the tasks, so shorter time indicates better performance. Thus, the scores were inverted to match the other cognitive tests on which higher scores correspond to better function.

Tests of visuospatial function are limited in the ADNI cohort and this domain is expected to be at most minimally affected in this cohort spanning the spectrum of preclinical to mild dementia due to AD; thus, this cognitive domain was not investigated in this study.

All cognitive scores were transformed to zscores based on the mean and SD of Aβ– controls.

2.2.1. Isolating cognitive functions

To further isolate the function within a specific cognitive domain, we adjusted each cognitive domain for a relevant other cognitive domain test score as shown in Table 1.

TABLE 1.

List of adjustments for each cognitive measure made to isolate a given function.

Cognitive domain	Cognitive test	Adjusted for	Removing effects of
Immediate recall	AVLT trial 1 + 2	/	/
Late recall	AVLT trial 4 + 5	Immediate recall	Working memory
Delayed recall	AVLT + ADAS‐Cog delayed	/	/
Recognition	AVLT recognition trial	Delayed recall	Episodic memory
Semantic fluency	Animal + vegetable fluency	Lexical fluency	Executive load
Executive function	TMT‐B	TMT‐A	Attention/visuomotor speed

Abbreviations: ADAS‐Cog, Alzheimer's Disease Assessment Scale Cognitive subscale; AVLT, Auditory Verbal Learning Test; TMT, Trail‐Making Test.

Late recall was adjusted for immediate recall to account for working memory function and thereby isolating learning ability. Recognition has been conceptualized as being dependent on two types of memory representations: recollection and familiarity. ²¹ , ³⁷ Recollection is partially captured by the delayed recall test. We therefore aimed to capture familiarity by adjusting recognition memory for delayed recall to remove the effects of recollection. ²¹ , ³⁷ Semantic fluency was adjusted for a measure of lexical fluency to account for the executive load of the task and isolate the fluency around semantic knowledge. ³⁷ Finally, the TMT‐B, a measure of executive function, was adjusted for the TMT‐A to remove effects of attention/visuomotor speed and isolate cognitive flexibility/shifting as a function. ³⁷ In all primary analyses, the cognitive measures of interest are adjusted by including the specified cognitive measure as a covariate in the models.

Two subdomains were not adjusted. First, immediate recall was not adjusted for another cognitive measure. While trials 1 + 2 of the AVLT tap into attention in addition to short‐term/working memory, it is difficult to disentangle from existing measures. Second, delayed recall was not adjusted for working memory/attention measures because consolidation is assumed to have taken place by the time of the delayed recall.

2.3. Imaging protocols

2.3.1. MRI

T1‐weighted structural MRI (voxel size: 1.0 × 1.0 × 1.2 mm³) scans were used. The MRI scans were chosen to be the MRI scan closest to the tau‐PET scan of the participant (with a tau‐PET done a median of 33 days after MRI, range: –302 to 176 days). The MRI imaging protocols of the ADNI study that were used to acquire the T1‐weighted MRI scans are described here: https://adni.loni.usc.edu/data‐samples/adni‐data/neuroimaging/mri/mri‐scanner‐protocols/.

2.3.2. Structural MRI processing and analysis

T1‐weighted MRI scans were segmented using the automated multi‐atlas segmentation pipeline Automatic Segmentation of Hippocampal Subfields (ASHS), ³⁸ providing estimates for MTL subregions anterior/posterior hippocampus, entorhinal cortex (ERC), Brodmann areas 35 and 36 (BA35 and BA36 respectively), and parahippocampal cortex (PHC). Volumetric measurements of bilateral anterior/posterior hippocampus were generated. For the MTL cortical subregions (ERC, BA35, BA36, and PHC), a graph‐based multi‐template thickness analysis pipeline ³⁹ was applied to derive the thickness of each subregion. In short, each ASHS segmentation was matched to a 3D template using diffeomorphic deformable registration and the template was mapped back to subject space to obtain a smooth approximation of each MTL subregion. ³⁹ Then the 3D skeleton of each smooth approximation was computed using the Voronoi skeletonization algorithm ⁴⁰ and the median distance from the surface of each subregion to the closest point on the skeleton was determined. ³⁹ T1‐weighted images were used for this study because specific hippocampal subfield associations beyond anterior versus posterior hippocampus require more specialized cognitive tasks.

Intracranial volume (ICV) was measured from T1‐weighted MRI using FreeSurfer and was obtained from the ADNI database. The anterior and posterior hippocampal volumes were adjusted for ICV prior to all analyses according to Voevodskaya et al., ⁴¹ in which ICV is regressed out.

In addition to the MTL subfields, we measured cortical thickness outside the MTL using the DiReCT method; ⁴² see also Klein and Tourville ⁴³ and Tustison et al., ⁴⁴ implemented in the Advanced Normalization Tools measurement pipeline. ⁴⁴ MRI scans were parcellated using a multi‐atlas segmentation method. ⁴⁵ These regions of interest (ROIs) were used to generate mean cortical thickness estimates averaged across bilateral regions.

In addition to the ROIs in the MTL, we selected the following neocortical ROIs for analysis: angular gyrus (AG), inferior temporal gyrus (ITG), middle frontal gyrus (MFG), superior frontal gyrus (SFG), precuneus (PRC), supramarginal gyrus (SMG), superior parietal lobule (SPL), and temporal pole (TP). These regions were chosen due to their common involvement in AD ⁴⁶ and putative role in the cognitive domains investigated. ²² , ²⁵ , ⁴⁷ , ⁴⁸

2.3.3. ¹⁸F‐Flortaucipir PET

The first ¹⁸F‐Flortaucipir PET scan per participant was used for estimating tau burden. The ¹⁸F‐Flortaucipir protocol used a 10 mCi tracer injection and a 30 minute scan (6 frames of 5 minute duration) after a 75 minute uptake phase.

2.3.4. Tau PET processing and analysis

Fully preprocessed images were downloaded from the ADNI archive (“Coreg, Avg, Std Img and Vox Size, Uniform Resolution”) and processed as described here. ⁴⁹ The inferior cerebellar gray matter was used as a reference region for creating SUVRs. Partial volume correction (PVC) was applied to the tau‐PET data, to reduce effects of off‐target binding in the MTL regions. A summary measure of MTL tau burden was computed as the average ¹⁸F‐Flortaucipir SUVR in a composite MTL cortical ROI consisting of ERC and BA35, from the ASHS‐T1 segmentations. Other MTL subregions were excluded from this measure because we aimed to capture early tau uptake in the MTL, which by definition includes only ERC and BA35. ⁶ , ⁷ Mean SUVR for ¹⁸F‐Flortaucipir uptake was also calculated for each bilateral neocortical region selected for cortical thickness estimates (see above; not including MTL) and summarized into one neocortical tau‐PET composite.

2.3.5. ¹⁸F‐Florbetapir/¹⁸F‐Florbetaben PET

¹⁸F‐Florbetapir and ¹⁸F‐Florbetaben PET scans acquired closest in time to the baseline tau‐PET scan (median interval between Aβ‐ and tau‐PET scans 5 days after Aβ‐PET, range: –161 to 796 days). The time interval between Aβ and tau‐PET was restricted in cases in which negative Aβ‐PET scans occurred before the tau‐PET scan but not if a positive Aβ‐PET scan occurred prior to tau‐PET. Aβ‐PET scans were used to determine participants’ Aβ status (Aβ+ vs. Aβ–). Images were acquired in a 20 minute scan (4 frames of 5 minute duration) after a 90 minute uptake phase after a ≈ 370 MBq tracer injection of ¹⁸F‐Florbetaben or 50 minute uptake phase of ≈ 300 MBq ¹⁸F‐Florbetapir. Image processing was performed by the ADNI PET core and values extracted from the ADNI database.

2.3.6. Amyloid PET processing and analysis

For ¹⁸F‐Florbetapir and ¹⁸F‐Florbetaben the whole cerebellum was used as reference region to calculate SUVRs. Thresholds to define Aβ positivity were previously established by the ADNI PET core ⁵⁰ and defined using a global composite cortical region. ⁵¹ For florbetapir the cut‐off was ≥ 1.11, for florbetaben ≥ 1.08, which correspond to Centiloid values of 20 and 18, respectively. ⁵²

2.4. Statistical analyses

All analyses were performed using R. All P values were controlled for the false discovery rate (FDR, Benjamini–Hochberg procedure) and considered statistically significant at P < 0.05.

2.4.1. Main analyses

The statistical analyses are divided into three parts (see Figure 1). Because our aim was to investigate mediation models, we first investigated the associations among the three aspects of the mediation (tau PET, regional structural MRI measure, and cognitive measure) to identify for which associations it is appropriate to fit mediation models. First, we investigated the associations between MTL or neocortical tau PET SUVRs with all cognitive (sub‐)domains using simple linear regression models covarying the models for age, sex, education, and cognitive measure chosen to isolate the cognitive function of interest (specified in Table 1). Additionally, the associations between tau PET SUVRs and structural measures were investigated, adjusting for age and sex. Furthermore, the association between structural measures and cognitive subdomains, adjusting for age, sex, education, and cognitive measure (specified in Table 1), were assessed. Further analyses were performed for the regions that were significantly associated with tau PET uptake and a given cognitive subdomain in this first step.

FIGURE 1.

Overview of the statistical analyses. Step 1: Linear regression models were fit to investigate the associations between the tau PET measures, regional thickness/volume, and cognitive subdomains. Step 2: Simple mediation models were fit per structural measure and cognitive subdomain. Step 3: Complex mediation models with the significant mediators from step 2 were fit per cognitive subdomain. Note that all models were adjusted for age and sex. Models including cognitive subdomains were additionally adjusted for education and the respective cognitive measure specified in Table 1. MTL, medial temporal lobe; n.s., not statistically significant; PET, positron emission tomography; SUVR, standardized uptake value ratio.

Second, we fit separate simple mediation models for cognitive domain and ROI thickness/volume using the lavaan R package. ⁵³ For MTL subregions, MTL‐tau PET SUVR was used, while neocortical tau‐PET SUVR was used for models investigating mediating effects of regional neocortical thickness measures on all cognitive domains. Models included age, sex, education, and cognitive measure chosen to isolate the cognitive function of interest (specified in Table 1) as covariates.

Third, to pinpoint the region primarily involved in mediating tau–cognition associations, we performed complex mediation models including all significant mediators from the simple mediation models. Before fitting these models, we investigated if both MTL and neocortical tau‐PET SUVRs were associated with all cognitive measures when included in a single model. This was done to investigate whether both tau‐PET measures were uniquely associated with a given cognitive domain. If this step of the analyses showed that both were associated with a cognitive domain when included in one model, the complex mediation was fitted with both tau‐PET composites and the mediators were fit with the associated regions (i.e., MTL thickness with MTL‐tau PET and neocortical thickness with neocortical tau‐PET). However, if analyses showed that only MTL‐ or neocortical tau‐PET was uniquely associated with a cognitive domain when included in one model, the complex mediation was only fit with the significant tauPET measure and the associated structural measures (MTL vs. neocortical). Complex mediation models were performed using the lavaan R package ⁵³ and included age, sex, education, and cognitive measure chosen to isolate the cognitive function of interest (specified in Table 1) as covariates. The reported standardized coefficients are extracted from the models.

3. RESULTS

3.1. Demographics

The study included of 319 Aβ+ older adults (51% female; mean age 72.8; mean education 16.3 years; 53.6% apolipoprotein E ɛ4 carriers; 91.2% White, 4.4% Hispanic). Most of the participants were CU (46.7%) or had MCI (42%). The demographics of the whole sample can be found in Table 2. Note that for the cognitive subdomains, raw values are shown in Table 2 for ease of interpretation, while z scored values were used for all analyses.

TABLE 2.

Overview of demographic information of the whole sample.

	Overall	CU	MCI	AD dementia
	319	149	134	36
Age	72.8 ± 6.86	72.7 ± 6.25	72.4 ± 7.07	74.9 ± 8.17
Median [min, max]	72.6 [55.0, 90.1]	71.9 [56.5, 90.1]	73.9 [55.0, 88.1]	75.1 [55.3, 89.2]
Sex (female) n (%)	164 (51.4)	93 (62.4)	58 (43.3)	13 (36.1)
Education (years)	16.3 ± 2.44	16.7 ± 2.29	16.1 ± 2.49	15.4 ± 2.61
Ethnicity n (%)
Hispanic/Latino	14 (4.4)	6 (4.0)	6 (4.5)	2 (5.6)
Non‐Hispanic/Latino	303 (95.0)	141 (94.6)	128 (95.5)	34 (94.4)
Unknown	2 (0.6)	2 (1.3)	0 (0)	0 (0)
Race n (%)
Asian	4 (1.3)	2 (1.3)	0 (0)	2 (5.6)
Black	17 (5.3)	10 (6.7)	5 (3.7)	2 (5.6)
White	291 (91.2)	135 (90.6)	124 (92.5)	32 (88.9)
Several apply	6 (1.9)	2 (1.3)	4 (3.0)	0 (0)
Unknown	1 (0.3)	0 (0)	1 (0.7)	0 (0)
APOE ɛ4 carrier n (%)	171 (53.6)	70 (46.9)	75 (55.9)	26 (72.2)
MMSE total score	27.2 ± 3.13	28.8 ± 1.34	26.5 ± 3.47	23.4 ± 2.79
Immediate recall # correct	5.76 ± 2.08	6.72 ± 1.99	5.16 ± 1.78	3.99 ± 1.41
Late recall # correct	8.95 ± 3.20	10.9 ± 2.64	7.65 ± 2.60	5.51 ± 1.55
Delayed recall # correct	4.96 ± 3.50	7.27 ± 2.78	3.44 ± 2.76	0.97 ± 1.34
Recognition # correct	9.20 ± 4.70	11.90 ± 3.09	7.56 ± 4.50	4.00 ± 3.68
Semantic fluency # words	18.3 ± 6.07	20.6 ± 5.36	17.0 ± 5.89	13.3 ± 5.20
Phonemic fluency # words	13.9 ± 4.78	14.5 ± 4.45	13.4 ± 5.03	12.9 ± 4.85
TMT‐A seconds *	39.8 ± 20.6	32.7 ± 10.2	43.6 ± 21.7	55.2 ± 33.7
TMT‐B seconds *	115 ± 74.7	85.5 ± 45.6	129 ± 79.2	183 ± 93.8

Note: Continuous variables are displayed as mean ± SD. Categorical variables are displayed as n (%). If the time difference variables are negative it indicates that the first named date was earlier than the second.

Abbreviations: AD, Alzheimer's disease; APOE, apolipoprotein E; CU, cognitively unimpaired; MCI, mild cognitive impairment; MMSE, Mini‐Mental State Examination; SD, standard deviation; TMT, Trail‐Making Test.

^*A higher score indicates worse performance, for all analyses this is reversed to match the other cognitive tests.

3.2. Associations among MTL and neocortical tau PET, cognitive subdomains, and structural thickness

As a first step, we looked at the associations between the two tau‐PET composites (MTL and neocortical; Figure 2A) with cognitive domains. In independent models, both MTL and neocortical tau‐PET uptake were significantly associated with all cognitive domains, except for neocortical tau‐PET uptake and recognition (P_FDR  = 0.364, all other P_FDR  < 0.001; Figure 2B; Table S1 in supporting information).

FIGURE 2.

Associations among tau‐PET uptake, cognitive domains, and structural measures of different regions‐of‐interest in individuals on the AD continuum. A: Overview of the regions of interest within the MTL and the neocortex and the definition of the two tau‐PET composite measures (MTL vs. neocortical). B: Results of separate regression models predicting cognitive function with one of the two tau‐PET composite measures. C: Association of MTL tau‐PET with MTL subregional volume/thickness and neocortical tau‐PET with neocortical thickness. D: Association between the structural thickness/volume measures in all regions of interest and the cognitive domains. Note that the dots in the graph represent the standardized beta coefficient of the mediation effect and the bars correspond to the 95% confidence interval. All models were adjusted for age, sex, and education and the models including cognition for the respective cognitive measure specified in Table 1. Gray lines are not statistically significant at P _FDR < 0.05. AD, Alzheimer's disease; BA, Brodmann area; FDR, false discovery rate; MTL, medial temporal lobe; PET, positron emission tomography; std. β, standardized β coefficient; SUVR, standardized uptake value ratio.

Next, the associations of MTL tau‐PET uptake with MTL subregional measures and neocortical tau PET uptake with neocortical thickness were investigated, adjusting for age and sex. Structural measures of all regions were associated with respective tau‐PET uptake, except for middle and superior frontal thickness (Figure 2C; Table S2 in supporting information). Numerically, the largest associations with MTL tau‐PET uptake were posterior hippocampus (std. β = –0.50, P_FDR  < 0.001) and entorhinal cortex (std. β = –0.42, P_FDR  < 0.001), while strongest associations with neocortical tau‐PET uptake were inferior temporal and angular gyrus (std. β = –0.30, P_FDR  < 0.001 and std. β = –0.25, P_FDR  < 0.001, respectively).

Finally, the associations between the structural thickness/volume and the cognitive subdomains were investigated, adjusting for age, sex, education, and the cognitive measure specified in Table 1 (Figure 2D). For immediate and late recall, fluency and executive functioning, structural measures of almost all regions were associated with each cognitive test score. For recognition and delayed recall associations with thickness were more limited and only a few regions within and outside the MTL were associated with these cognitive subdomains (Figure 2D; Table S3 in supporting information). None of the results were significantly altered when accounting for the time difference between tau‐PET and MRI/cognition.

3.3. Different MRI‐based regional structural measures are mediators of tau–cognition associations

In the second step of the analysis, simple mediation models for the tau–cognition associations were fitted, as specified in Figure 3, per cognitive domain and for the thickness/volume measures of ROI separately. Only regional thickness/volume that showed associations with tau‐PET uptake and cognitive subdomains (see Figure 2) were investigated as potential mediators. The results of the mediation models are shown in Figure 3 (see Tables S4–S9 in supporting information for detailed results).

FIGURE 3.

Simple mediation models show differential patterns of single partial thickness mediators per model for the tau and cognition associations in individuals on the AD continuum. The results of the indirect mediating effect per cognitive measure and region are shown both projected on the surface rendering of the MTL and brain as well as in the graph next to the renderings. The black outline on the surface projections and green bars on the graphs indicate P _FDR < 0.05. Mediation models were only fit for regions that were associated to both tau‐PET and cognitive measures. Note that the dots in the graph represent the standardized beta coefficient of the mediation effect and the bars correspond to the 95% confidence interval. All mediators are partial mediators. All mediations were adjusted for the respective cognitive measure (see Table 1), age, sex, and education. The whole sample was used for all analyses (CU/MCI/AD). a*b, standardized β coefficient of the mediating effect; AD, Alzheimer's disease; AG, angular gyrus; aH, anterior hippocampus; BA, Brodmann area; CU, cognitively unimpaired; ERC, entorhinal cortex; FDR, false discovery rate; IT, inferior temporal gyrus; MCI, mild cognitive impairment; MTL, medial temporal lobe; PET, positron emission tomography; PC, precuneus; pH, posterior hippocampus; PHC, parahippocampal cortex; SM, supramarginal gyrus; SP, superior parietal gyrus; std. β of a*b, standardized β coefficient of the mediating effect.

Focusing first on the memory subdomains, the associations between tau and immediate recall as well as tau and delayed recall were mediated by structural measures from a large set of regions including MTL and neocortical regions, while the mediators of the tau–late recall and tau–recognition associations were restricted to the MTL. ,  

The MTL tau–delayed recall association was partially mediated by MTL structural measures (hippocampus, entorhinal cortex, BA35, variance explained: 13.9%–22.7%), while the neocortical tau–delayed recall association was partially mediated by thickness in neocortical regions inferior temporal cortex, supramarginal cortex, and temporal pole (variance explained: 5.1%–8.8%; Figure 3iii). The MTL mediating effects were numerically larger than the mediating effects for the other cognitive subdomains. The MTL tau–immediate recall association was partially mediated by posterior hippocampal volume, entorhinal cortex, and BA35 thickness (variance explained: 9.1%–18.1%), while neocortical tau–immediate recall was partially mediated by angular and inferior temporal, precuneus and temporal pole cortex thickness (variance explained: 8.1%–20.8%; Figure 3i). The MTL tau–late recall association was partially mediated by MTL subregional volume/thickness of anterior and posterior hippocampus and BA35 (std. β = –0.06; –0.09; –0.05, respectively; variance explained: 11.7%–23.0%; Figure 3ii). No significant mediations for the neocortical tau–late recall association were observed.

The MTL tau–recognition association was partially mediated by posterior hippocampal volume only (std. β = –0.06; variance explained: 9.6%; Figure 3iv). No significant mediations for the neocortical tau–recognition association were observed. As specified in Table 1 and explained above, recognition was adjusted for delayed recall to account for recollection effects and investigating the role of familiarity in recognition. However, as a sensitivity analysis, the mediation models were performed without adjusting the mediation models for delayed recall (see Table S10 and Figure S1 in supporting information). These sensitivity analyses provided support for the posterior hippocampus as the strongest, albeit not unique, partial mediator (std. β = –0.15, variance explained: 6.7%–23.9%).

The association between neocortical tau and semantic fluency was partially mediated by thickness in inferior temporal gyrus and precuneus (std. β = –0.05; –0.04; variance explained: 9.2%–14.3%; Figure 3v). No significant mediators for the MTL tau–semantic fluency association were observed. Last, the association between MTL tau and executive function was partially mediated by entorhinal cortex thickness (std. β = –0.07, variance explained: 34.9%; Figure 3vi). No significant mediators were observed for the neocortical tau–executive function association.

3.4. Complex mediation analyses support differential mediating effects of specific tau–cognition associations

Before fitting complex mediation models, we investigated if both tau‐PET measures should be included in the model, or, in other words, if both measures are uniquely associated with the cognitive subdomain under investigation. When including both tau‐PET composites in one model (Figure 4A), only MTL tau–PET uptake was associated with the episodic memory subdomains (immediate: std. β = –0.23, P_FDR  = 0.004, late: std. β = –0.25, P_FDR  < 0.001, delayed recall: std. β = –0.50, P_FDR  < 0.001, recognition: std. β = –0.24, P_FDR  < 0.001). Only neocortical tau‐PET uptake was associated with executive functioning (std. β = –0.21, P_FDR  = 0.002). Both MTL and neocortical tau‐PET uptake were associated with semantic fluency (MTL tau: std. β = –0.20, P_FDR  = 0.006; neocortical tau: std. β = –0.19, P_FDR  = 0.009). Therefore, only MTL tau‐PET was included in the mediation models for the memory domains, only neocortical tau‐PET for executive functioning, and both tau‐PET measures for semantic fluency.

FIGURE 4.

Complex mediation models of partial thickness mediators for the tau and cognition associations in individuals on the AD continuum. A: Results of regression models predicting cognitive function using both tau PET composites in one model. B: Results of the surviving mediators of the complex mediation models in which all significant mediators per cognitive domain from Figure 3 have been fitted into one complex mediation model (Bi–Bvi). The mediators of the model for the tau–immediate recall (Bi) association are not significant, while the overall model is, indicating that both regions may be involved but share a considerable portion of the variance explained. Note that for a in the complex mediation, inferior temporal thickness was p = 0.06. All mediators are partial mediators. All mediations were adjusted for the respective cognitive measure (see Table 1), age, sex, and education. The whole sample was used for all analyses (CU/MCI/AD). For improved readability, only the significant mediators of the complex mediations are shown together with the percentage of variance explained compared to the base model without mediators. a*b, standardized β coefficient of the mediating effect; AD, Alzheimer's disease; BA, Brodmann area; CU, cognitively unimpaired; MCI, mild cognitive impairment; MTL, medial temporal lobe; PET, positron emission tomography; std. β of a*b, standardized β coefficient of the mediating effect; SUVR, standardized uptake value ratio.

To fit the complex mediation models, we only included the significant mediators from the simple mediation models from Figure 3 in one complex mediation per cognitive subdomain. The results are shown in Figure 4.

The results indicate a significant complex mediation for immediate recall (Figure 4Bi, variance explained = 9%) with posterior hippocampus, entorhinal cortex, and BA35 being partial mediators. However, while the overall model is significant, the individual mediators are not, potentially suggesting that while all three regions may be relevant, they share much of the variance explained. For all other memory domains (Figure 4Bii‐vi), specific structural mediators remained significant: anterior and posterior hippocampus for late recall (variance explained = 27%), posterior hippocampus and BA35 for delayed recall (variance explained = 27%), and posterior hippocampus for recognition (variance explained = 12%).

For the non‐memory domains, when including all significant mediators from the prior step in the complex model, only inferior temporal thickness survived as the partial mediator of the neocortical tau–semantic fluency association (variance explained = 14%, Figure 4Bv). For the neocortical tau–executive function association, no mediation model was significant.

It is of importance to note that these complex mediation models explain between 9% and 29% of the variance between respective tau‐PET and cognitive subdomain measures, leaving a considerable portion of variance unexplained.

3.5. Results are consistent in the cognitively impaired group

We additionally performed sensitivity analyses to assess the possibility of a disease stage effect by repeating the analyses for the CU and cognitively impaired groups separately. Focusing first on the cognitively impaired group, compared to the main analyses a lower number of significant mediators were observed (Figure S2 in supporting information). This may be due to reduced statistical power given a lower sample size (n = 170, see Table 2). For tau–immediate recall, precuneus was found as a partial mediator, while hippocampus, entorhinal cortex, and BA35 remained significant partial mediators for the tau–delayed recall association. Additionally, the MTL tau–executive function association was partially mediated by entorhinal cortex thickness. These results are consistent with the primary analyses.

For the cognitively unimpaired group, the tau‐PET composites were significantly associated with all cognitive subdomains (Figure S3 in supporting information). However, no significant associations between tau‐PET composites and regional thickness/volume as well as between cognition and regional thickness/volume were found and, thus, no mediation models were fit (Figure S3). This indicates limited neurodegeneration in this group when measured in a cross‐sectional manner.

4. DISCUSSION

Using tau‐PET, fine‐grained measures of cortical atrophy and measures of isolated domain‐specific cognitive impairment in early amnestic AD, we identified unique and granular information about the tau‐structure–cognition pathway. We find that MTL tau‐PET is uniquely associated with memory domains, MTL and neocortical tau‐PET with semantic fluency, and neocortical tau with executive functioning. These associations for the memory subdomains and semantic fluency were partially mediated by structural MRI measures in a region‐specific manner (9%–27% of explained variance), for example, tau–immediate recall by hippocampus and cortical extrahippocampal MTL regions, tau–recognition by hippocampus only, and tau–semantic fluency by inferior temporal gyrus. The association between neocortical tau‐PET and executive functioning was not mediated by structural MRI. These findings were also observed, albeit more restricted, in the cognitively impaired group, but not in the CU group. Overall, these results are consistent with the hypothesis that pathological tau partially leads to cognitive impairment through macrostructural changes. However, alternative pathways, such as synaptic dysfunction, are likely at play and the pathways may change throughout the course of the disease.

In this study we used complex mediation models to assess if local structural brain changes uniquely mediate effects of tau‐induced domain‐specific cognitive impairment, giving insights into the neurobiological substrates of tau‐induced cognitive changes. Importantly, BA35 emerged as a significant mediator for both tau–immediate and tau–delayed recall associations. For immediate recall, we anticipated a weaker influence of MTL structures on this measure due to the involvement of the verbal working memory system. ²⁵ , ⁵⁴ This is supported by the simple mediation models showing more widespread neocortical mediators. Yet, it must be kept in mind that we likely captured primarily early tau pathology with this cohort consisting of Aβ‐positive cognitively normal to primarily mild dementia cases (n = 4 with a Clinical Dementia Rating global score = 2), which likely resulted in limited tau pathology in neocortical regions. Additionally, off‐target binding within the MTL could have increased the tau‐PET signal slightly in this region. The finding that also entorhinal cortex and posterior hippocampus, central for integration of information and retrieval, are relevant mediators for immediate recall supports the idea that a larger memory system is important for this function. ⁵⁵ , ⁵⁶ For delayed recall previous work has highlighted the importance of the hippocampus, ²¹ , ²⁵ , ⁴⁸ , ⁵⁷ , ⁵⁸ corroborated by our observation of posterior hippocampus as mediator. Our finding of BA35 being a mediator for both immediate and delayed recall may be because we, in contrast to previous work, investigated mediating effects of the tau and cognition association, not associations between structural measures and these functions. Given that BA35 (≈ transentorhinal cortex) is the earliest cortical region to accumulate tau pathology, ⁶ , ²⁸ our findings suggest a significant impact of early tau‐induced changes within this critical subregion for both immediate and delayed episodic memory. Of additional interest is the emergence of the posterior hippocampus as a mediator of the tau–recognition association. This contrasts with research highlighting the importance of extrahippocampal MTL regions for recognition memory, ²¹ , ²⁵ , ⁴⁸ , ⁵⁷ , ⁵⁹ but note Diana et al. ⁶⁰ Variations in MTL subregion segmentation reliability, with higher reliability for hippocampal subregions, ³⁸ could at least partially explain these discrepancies. Moreover, adjusting recognition for delayed recall may have influenced our findings. However, the posterior hippocampus has been shown to be closely connected to entorhinal cortex, ⁶¹ a region accumulating tau early after BA35. ⁶ Thus, our findings could be explained by the hypothesis that early tau accumulation already affects the posterior hippocampus.

Interestingly, inferior temporal thickness was indicated to be a mediator for the tau–semantic fluency association. This finding aligns with the notion that the inferior temporal cortex serves as a transition zone between early (trans‐)entorhinal tau and regions accumulating tau later ⁶² , ⁶³ and has been linked to accelerated atrophy. ⁶⁴ Additionally, while involvement of the perirhinal cortex would be expected for semantic fluency, the perirhinal cortex exhibits connectivity with and is located adjacent to the inferior temporal cortex and both have been found to be affected in semantic dementia. ⁶⁵ Thus, our findings further emphasize the importance of the inferior temporal cortex to semantic processes in AD. Overall, we contribute to a more nuanced understanding of the regions involved in different tau‐induced changes in cognitive processes. Within the MTL, first accumulation of tau pathology is thought to occur in the transentorhinal cortex (corresponding to Brodmann area 35) and the entorhinal cortex after which the pathology spreads to other MTL subregions and the inferior temporal cortex. ⁶ , ⁷ The spread of tau outside the MTL is typically thought to involve temporal, parietal, and frontal regions but has been shown to be variable between individuals. ⁶ , ⁶⁶ In addition to the MTL's established role in episodic memory, ²³ corroborated by our results, we demonstrate a correspondence between these effects and known patterns of tau pathology spread within the MTL and inferior temporal cortex, implicating tau‐induced effects in specific subregions that map to distinct memory processes.

Our findings suggest that a portion of the tau–cognition association is mediated by macrostructural changes, particularly in brain regions exhibiting early tau accumulation. This supports the hypothesis that prolonged exposure to tau contributes to neurodegeneration, in turn contributing to cognitive impairment. It is important to note that the structural measures explained 9% to 27% of the tau–cognition associations. One possibility is that methodological limitations, such as reliability of MTL subfield measurements, statistical model limitations, noise in the cognition measures, or covariance between measures, may have resulted in a lower variance explained. Thus, more sensitive imaging measures should be investigated in this context. It is also possible that co‐pathologies are contributing to tau‐induced domain‐specific cognitive impairment, not captured by structural MRI measures. ⁶⁷ , ⁶⁸ Finally, this could implicate other processes in contributing to cognitive impairment, such as microstructural changes ¹¹ not captured (yet) by volume/thickness measures, functional changes such as synaptic dysfunction, ⁶⁹ or neuroinflammation. ⁷⁰ The absence of neocortical mediators for the tau‐PET and executive function association suggests that there may not have been enough macrostructural changes in the neocortex yet to mediate the tau effect or that other mechanisms play a more prominent role. These limited neocortical structural changes may be a cohort effect because the ADNI study population is comprised of earlier stages of typical AD and has been found to have lower neocortical tau‐PET uptake compared to other cohorts. ¹³ The mechanisms linking tau pathology to cognitive impairment may be different in a cohort with higher levels of neocortical tau, such as cohorts including more individuals with advanced typical AD dementia or with atypical presentations. Of note, even though atypical cases are excluded by design in the ADNI cohort, and all participants have memory deficits, it cannot be completely ruled out that some case with trends toward an atypical presentation have been included, although our results seem to speak against this possibility. Neocortical regions may then harbor more tau pathology and/or undergo more atrophy and decline in cognition, leading to stronger associations between tau and cognition in other domains than memory and potential mediating effects of macrostructural measures in neocortical regions. The considerations about alternative mechanisms also have implications for clinical trials, which are moving to enroll earlier stages of AD. ⁷¹ If tau effects on cognitive subdomains are driven by different fine‐grained regional changes, measuring larger regions with coarse measures (e.g., total hippocampus or cortical AD signature), may not be useful secondary endpoints. While our study examining granular MTL subregions provides more nuanced information, macrostructural measures of these granular regions explain associations to a limited extent. Future research should, thus, investigate whether microstructural changes within specific regions, such as cortical microstructure of MTL subfields, specifically posterior hippocampus or even hippocampal subfields and BA35, may be even more uniquely suited to detecting subtle early changes, ⁷² , ⁷³ , ⁷⁴ explaining a larger portion of tau‐induced cognitive changes, and could be used in clinical trials.

This study extends prior research by investigating the mediating effects of tau‐induced impairment of cognitive subdomains, using a framework in which adjusting cognitive functions for different subdomains is done to obtain isolated measures although definitive proof of this approach could not be provided in this study, as this would require further validation, for example against more advance cognitive tests (e.g., Berron et al., ⁷⁵ Gellersen et al.,⁷⁶ and Polk et al. ⁷⁷ ). However, the biological plausibility of the substrates observed in this study suggests that this approach could provide relevant or fine‐grained information. Nevertheless, as the research community emphasizes the need for improved cognitive measures to detect subtle and early cognitive changes, ⁷⁸ research using traditional cognitive testing should explore this approach further. Moreover, investigating this approach in the context of digital cognitive testing platforms presents a valuable avenue for future research, which is easily implementable.

4.1. Conclusion

In summary, the results suggest region‐specific atrophy functions as one pathway of tau‐induced impairments in isolated measures of different cognitive subdomains. However, the results also indicate the need to investigate other mechanistic pathways, such as synaptic function using synaptic PET, fluid markers, or microstructural changes using diffusion tensor imaging. Specifically, BA35 and inferior temporal gyrus are relevant regions that should be explored further in this context. Finally, the utility of the approach of adjusting cognitive domains to isolate specific cognitive functions was supported by biologically plausible structural substrates. This indicates that this approach is a promising starting point to investigate cognition and improve the isolation of cognitive functions, also in the context of digital cognitive testing. Combining fine‐grained regional atrophy measures with improved cognitive testing is of potential relevance for clinical trials as this may enhance monitoring and tracking of fine‐grained treatment responses.

CONFLICT OF INTERESTS STATEMENT

N.M.C. has received consultancy/speaker fees from Biogen, Eli Lilly, Owkin, and Merck. D.A.W. has served as a paid consultant to Eli Lilly and Beckman Coulter; serves on a DSMB for Functional Neuromodulation and GSK; and was a site investigator for a clinical trial sponsored by Biogen. L.X. received personal consulting fees from Galileo CDS, Inc. L.X. is a paid employee of Siemens Healthneers. S.R.D. received consultation fees from Rancho Bioscience and Nia Therapeutics. All other authors have nothing to declare. Author disclosures are available in the supporting information.

CONSENT STATEMENT

This study was conducted using the already collected data from the ADNI. Institutional review boards at all sites approved the ADNI study and written informed consent was obtained from all participants.

Supporting information

Supporting Information

Supporting Information

Wuestefeld A, Xie L, McGrew E, et al. Tau, atrophy, and domain‐specific cognitive impairment in typical Alzheimer's disease. Alzheimer's Dement. 2025;21:e70511. 10.1002/alz.70511