Tau-atrophy variability reveals phenotypic heterogeneity in Alzheimer’s disease

Abstract

Objective:

Tau neurofibrillary tangles (T) are the primary driver of downstream neurodegeneration (N) and subsequent cognitive impairment in Alzheimer’s disease (AD). However, there is substantial variability in the T-N relationship – manifested in higher or lower atrophy than expected for level of tau in a given brain region. The goal of this study was to determine if region-based quantitation of this variability allows for identification of underlying modulatory factors, including polypathology.

Methods:

Cortical thickness (N) and ¹⁸F-Flortaucipir SUVR (T) were computed in 104 gray matter regions from a cohort of cognitively-impaired, amyloid-positive (A+) individuals. Region-specific residuals from a robust linear fit between SUVR and cortical thickness were computed as a surrogate for T-N mismatch. A summary T-N mismatch metric defined using residuals were correlated with demographic and imaging-based modulatory factors, and to partition the cohort into data-driven subgroups.

Results:

The summary T-N mismatch metric correlated with underlying factors such as age and burden of white matter hyperintensity lesions. Data-driven subgroups based on clustering of residuals appear to represent different biologically relevant phenotypes, with groups showing distinct spatial patterns of higher or lower atrophy than expected.

Interpretation:

These data support the notion that a measure of deviation from a normative relationship between tau burden and neurodegeneration across brain regions in individuals on the AD continuum captures variability due to multiple underlying factors, and can reveal phenotypes, which if validated, may help identify possible contributors to neurodegeneration in addition to tau, which may ultimately be useful for cohort selection in clinical trials.

Introduction

In a shift towards a biological, rather than clinical, definition of Alzheimer’s Disease (AD), the National Institute on Aging – Alzheimer’s Association (NIA-AA) recently formalized a biomarker driven classification referred to as the A/T/(N) framework¹. Following this system, individuals receive dichotomous designations for the presence (+) or absence (−) of Beta-amyloid plaques (A), tau-based neurofibrillary tangles (T), and neurodegeneration (N). To be classified as AD, one must be both A+ and T+ paralleling the pathological criteria for AD neuropathologic change (ADNC). Neurodegenerative markers are meant to support biological staging of disease.

Neurofibrillary tangles (NFT) are more tightly linked to downstream neurodegenerative change than amyloid plaques^2–6. As such, the A/T/(N) framework posits that neurodegeneration without NFTs (T-) is due to the presence of suspected non-AD pathophysiology (SNAP) regardless of the presence or absence of amyloid^7,8.

While the dichotomous designation of these biomarkers provides a simplicity in interpretation, it may blunt the information provided about an individual’s underlying disease status. In particular, T and N imaging biomarkers provide continuous and spatially varying information that are likely important for understanding phenotype. As NFTs are the putative primary driver of neurodegeneration in AD, T is expected to be tightly linked to N. Thus, discordance between T and N suggests additional non-AD modulators. While the extreme of T− and N+ supports SNAP, even in a T+ individual, a relatively larger magnitude of N than the magnitude of T also may support concomitant SNAP. Further, the spatial pattern of this discordance may support different underlying non-AD pathologies, given the differential loci of greatest neurodegeneration with these conditions⁹. For example, greater anterior temporal atrophy than expected for the amount of local NFT pathology may suggest concomitant limbic-predominant age-associated TDP-43 encephalopathy (LATE)^10,11 which often co-occurs with AD and primarily affects this region. Alternatively, less neurodegeneration than expected for a given degree of T, may suggest resilience to AD pathology, or brain reserve¹².

Given that co-pathology with AD is common¹³ and has implications for clinical interventions and prognosis, and that definitive biomarkers for non-AD molecular pathology are lacking¹⁴, approaches that can operationalize the presence of co-pathology are critical, as are measures that capture resilience. The current study attempts to provide greater precision to the A/T/(N) framework by treating T and N as continuous variables and then determining their relationship in a way that is “spatially aware”. We first establish the normative relationships between T and N at regions of interest (ROIs) across the cortical mantle in people on the AD continuum (A+). Deviation from this relationship at each ROI constitutes mismatch. We predicted that a quantitative measure of this T-N mismatch would be associated with factors that may reflect non-AD neurodegeneration, such as the presence of cerebrovascular disease or age, as well as with cognitive performance. Further, we examined whether using data-driven clustering based on spatial pattern of these T-N mismatch metrics would produce distinct groups that would offer insight into the presence of different concomitant pathologies or brain reserve/resilience.

Materials and Methods:

Participants:

Data used in the preparation of this article were obtained from the ADNI database. The current study included 137 A+ cognitively impaired participants with diagnosis of MCI or dementia. The number of datasets were based on data available for download in January 2020. They included 80 MCI (age 76.1 ± 7.5 years, 38 Female, 15.8 ± 2.7 years of education, MMSE score of 27.1 ± 2.4), and 57 Dementia cases (age 77.7 ± 9.5 years, 27 Female, 15.9 ± 2.5 years of education, MMSE score of 22.5 ± 4.1).

Image acquisition:

ADNI MR imaging included a T1-weighted structural scan of resolution 1.0×1.0× 1.2 mm³, and a FLAIR MRI acquired in the same session with variable spatial resolution as prescribed in the ADNI protocol. Tau PET imaging consisted of a continuous 30-minute brain scan (six 5-minute frames) started 75 minutes following injection of approximately 10 mCi of ¹⁸F-Flortaucipir injection. Most patients (N=106) had an amyloid PET scan, with a 20 min brain scan (four 5-minute frames) performed 50 min after an approximately 10 mCi injection of the radiotracer ¹⁸F-Florbetapir. The rest of the participants received an ¹⁸F-Florbetaben Amyloid PET scan. After an injection of approximately 8.1 mCi of ¹⁸F-Florbetaben and a 90 min uptake phase, a 20 min brain scan (four 5-minute frames) was performed. The respective PET scans acquired closest in time to the structural MRI were analyzed. PET images were downloaded from the ADNI data archive in the most fully pre-processed format with the image description of “Coreg, Avg, Std Img and Vox Siz, Uniform Resolution”.

Image processing:

The pre-processed PET images were aligned with the anatomical MRI via rigid registration using the ANTs normalization software¹⁵ with a mutual information metric. The anatomical MRI was parcellated into ROIs including cortical, subcortical and cerebellar ROIs using a multi-atlas segmentation method¹⁶. The parcellation scheme is described elsewhere¹⁷. Mean tracer uptake in cerebellar gray matter and gray+white matter was used as reference to generate an SUVR map for the entire brain for ¹⁸F-Flortaucipir and ¹⁸F-Florbetapir/¹⁸F-Florbetaben, respectively. A composite ROI consisting of middle frontal, anterior cingulate, posterior cingulate, inferior parietal, precuneus, supra marginal, middle temporal and superior temporal cortex was used to compute a global SUVR for Amyloid PET scans. Thresholds of SUVR ≥ 1.11 and 1.08 for ¹⁸F-Florbetapir and ¹⁸F-Florbetaben respectively, were then used to determine Amyloid-β status, as previously described¹⁸. Composite SUVR provided by ADNI are now available for all but one participant (variable SUMMARYSUVR_WHOLECEREBNORM) and had a correlation of r=0.97 with those used here. Amyloid-β status was identical in all participants.

The T1-weighted structural MRI was processed using the ANTs cortical thickness pipeline¹⁹ that implements a diffeomorphic registration-based thickness estimation method²⁰. ROI-based measures of tau (T) and neurodegeneration (N) were then computed as average ¹⁸F-Flortaucipir SUVR and cortical thickness respectively within 104 bilateral ROIs (52 from each hemisphere).

White matter hyperintensity masks were computed from FLAIR images using a previously validated deep learning-based method that was the top performer in the WMH segmentation challenge²¹. The masks were visually examined by a trained rater and edited as necessary before computing volumes.

Modeling of Tau (T) – Neurodegeneration (N) relationship: a T-N mismatch metric

Tau-Neurodegeneration relationship was modeled using average gray matter thickness (N) and ¹⁸F-Flortaucipir SUVR (T) for 104 bilateral ROIs in each subject. A robust linear regression²² was performed at each ROI. To mitigate effects of a skewed distribution of SUVR, the common logarithmic transform was applied to T. Additionally, a bi-square weighting function which can further minimize the influence of outliers was used for robust regression. An example is shown in Figure 1 (a).

Figure 1.

Scatter plots showing (a) Tau tracer SUVR vs. cortical thickness in the right inferior temporal gyrus ROI along with robust regression fit and ± 1.5 standard deviation lines, (b) T-N mismatch metric vs. age, (c) T-N mismatch metric vs. MMSE scores, and (d) T-N mismatch metric vs. WMH volume.

For each individual, a summary measure, T-N mismatch metric, was calculated by taking the difference between the number of ROIs that had residuals below 1.5 × standard deviation from the regression line (indicating greater than expected N for the given level of T load) and number of ROIs that had residuals above 1.5 standard deviation from the regression line (indicating less than expected N for the given level of T load). Thus, a positive T-N mismatch metric suggests generally more atrophy than expected for tau (i.e more vulnerable) while a negative T-N mismatch metric suggests greater cortical thickness than expected for tau burden (i.e. resilience or brain reserve). The relationship of this metric with demographic variables and cognitive performance scores were analyzed.

Clustering for phenotype discovery

The regression residual for each ROI was discretized into a two-element binary vector based on whether it was farther than 1.5 standard deviation away from the regression line and if the residual was negative or positive. These binary vectors were then entered into Ward’s hierarchical agglomerative clustering method²³ to generate data-driven groupings of subjects. The number of clusters (6) was specified empirically by the elbow method proposed by Thorndike²⁴ that uses the percentage of variance explained as a function of number of clusters and optimizes within-cluster similarity.

Results

Pearson correlation between the summary T-N mismatch metric, defined as the difference in number of ROIs with negative versus positive residuals of greater than 1.5 × standard deviation, and age and WMH volume were calculated (Table 1, Figure 1 (b),(d)). As expected, age (r=0.38, P<10⁻⁵) and WMH volume (r=0.35, p<0.0001) were strongly correlated with the mismatch metric, such that older age and greater WMH volume were associated with a higher T-N mismatch metric. In other words, these factors were associated with more atrophy than expected for a given level of tau pathology within this study population. Moreover, when the relationship between WMH volume and T-N mismatch was covaried for age, the correlation remained significant (r = 0.31, p<0.001).

Table 1:

Summary of demographic and imaging variables within tertiles of T-N mismatch metric and its correlation

Variable	Lower tertile (resilient)	Middle tertile	Upper tertile (vulnerable)	Pearson r/chi-square	p-value
Age (Years)	73.3	76.4	80.9	0.38	< 10−5
Gender (F/M)	26/20	24/22	15/30	5.53	0.06
ApoE4 status (0/1/2 allele)	12/14/12	10/18/8	16/18/5	4.88	0.29
MMSE	26.2	25.9	23.5	−0.30	0.0005
WMH volume (log mm3)	1.30	1.21	2.25	0.35	< 0.0001
ERC tau (18F-Flortaucipir SUVR)	1.35	1.34	1.31	0.03	0.69
IT tau (18F-Flortaucipir SUVR)	1.43	1.45	1.46	0.003	0.96
Precuneus tau (18F-Flortaucipir SUVR)	1.25	1.26	1.32	0.04	0.63
Angular gyrus tau (18F-Flortaucipir SUVR)	1.38	1.34	1.39	0.01	0.88

Mean (age, MMSE, WMH volume and regional tau burden) or categorical counts (gender and ApoE4 status) for participants trichotomized into tertiles of T-N mismatch metric. Upper tertile has the highest relative atrophy than expected for tau (vulnerable). Rightmost columns show Pearson correlation of age, MMSE and the imaging measures with the T-N mismatch metric is also shown. For the two categorical variables gender and ApoE4 status, the two result of chi-square test between the three tertiles are shown instead.

To determine whether the T-N mismatch metric was associated with the degree of tau pathology, we examined correlations with ¹⁸F-Flortaucipir uptake in areas typically affected by AD: entorhinal cortex, inferior temporal cortex, precuneus, and angular gyrus. We found no significant relationship of the T-N mismatch metric to tau deposition in these regions, supporting the notion that T-N mismatch is not a reflection of AD severity, as determined by tau PET.

We also examined the relationship of this summary T-N mismatch metric with performance in various cognitive tasks, with age, gender, years of education and IT tau SUVR as covariates. The latter, an early region of tau pathology, was used as a surrogate measure of tau load to determine the degree to which mismatch influenced performance beyond AD NFT pathology. Cognitive tasks included a global measure (MMSE), measures of memory (Logical Memory and AVLT delayed recall), language (category fluency), visuomotor speed (Trailmaking Test A) and executive function (Trailmaking Test B). These results are summarized in Table 2. Significant correlations were found with MMSE, AVLT, category fluency scores and Trailmaking Test A while Trailmaking Test B was borderline.

Table 2:

Correlation of cognitive scores with T-N mismatch metric with age, gender, years of education and IT tau uptake as covariates

Cognitive test	Pearson r	p-value
MMSE	−0.30	0.0005
Logical Memory Test (Delayed Recall)	−0.15	0.09
AVLT 5-min Delayed Recall	−0.24	0.006
Category fluency	−0.23	0.01
Trail making Test A	0.19	0.03
Trail making Test B	0.17	0.06

Correlational analyses between T-N mismatch metric and demographic, imaging, and cognitive performance measures reported above were repeated with an additional covariate of composite amyloid PET SUVR. Amyloid SUVR was not a significant predictor for any of the analyses.

Data driven clustering of participants based on T-N mismatch metric

Participants were grouped into six clusters using hierarchical clustering. They appear to represent different imaging phenotypes with distinct spatial topography, which may be a proxy for underlying disease phenotypes, as summarized in Table 3. Overall, the groups differed in age, burden of white matter hyperintensity (WMH), clinical status, and MMSE. However, they did not differ in tau burden in key AD ROIs (e.g. inferotemporal cortex, precuneus, angular gyrus) based on ¹⁸F-Flortaucipir uptake.

Table 3.

Characteristics of groups defined by data-driven clustering of residuals from T-N. Bottom row indicates variables that differ significantly between groups.

Group (N)	Description	Diagnosis MCI/Dementia	Age	Gender (F/M)	Log White Matter Hyper-intensity	MMSE	IT Tau SUVR	T-N Mismatch Metric (median)
Group 1 (61)	Canonical (low residuals)	38/23	76.8±8.0	32/29	8.38	25.8±3.2	1.43±0.4	−2
Group 2 (24)	High temporal/limbic atrophy (pattern suggestive of TDP-43)	9/15	79.9±7.5	10/14	9.47*	24.2±3.7	1.46±0.4	10
Group 3 (13)	Resilient (less atrophy relative to tau in lateral cortical)	10/3	69.7±8.8*	9/4	8.02	27.0±3.4	1.46±0.4	−14
Group 4 (3)	High temporal/limbic and diffuse atrophy (oldest group; ?TDP + vascular + age-related change)	1/2	87.2±6.1	0/3	7.88*	23.0±5.6	1.46±0.2	43
Group 5 (25)	Resilient Temporal	18/7	75.1±6.8	11/14	7.70	25.1±4.4	1.46±0.3	−14
Group 6 (11)	High diffuse atrophy (lowest cognition, high WMH)	4/7	79.3±10.2	3/8	9.73*	22.6±6.0	1.45±0.5	27
Group Difference		p < 0.05	p< 0.01	p=0.15	p = 0.06	p<0.05	p> 0.05	-

^*indicates significant group difference compared with Group 1 (two-sided t-test).

Group 1 was the largest group (n=61) and had atrophy and tau load measures closest to the regression line. We labeled this group “canonical” since it defines the expected relationship between neurodegeneration and tau burden. Group 2 (n=24) displayed greater atrophy than expected for ¹⁸F-Flortaucipir uptake particularly in anterior temporal/temporal limbic regions (Figure 2). The group was slightly older than the canonical group with significantly greater WMH burden (p=0.01) and slightly lower MMSE. A number of cases in Group 2 had particularly severe anterior hippocampal/temporal polar atrophy suggestive of temporal lobe pathology such as concomitant TDP-43^10,11 or hippocampal sclerosis²⁵, as described in autopsy studies. Alternatively, Group 3 (n=13) appeared to be a resilient imaging phenotype with less atrophy relative to tau, particularly in lateral cortical regions (Figure 2). While Group 3 was significantly younger than the canonical group (p=0.01), this group did include older individuals who appeared to have minimal atrophy despite significant ¹⁸F-Flortaucipir uptake as illustrated by the case in Figure 2. This group also had higher mean MMSE scores than the canonical group. Group 4 (n=3) was quite small and had severe, diffuse atrophy greater than expected for tau in both limbic and cortical regions. This group was marked by being the oldest (age range of 82–94). Group 5 (n=25) displayed resilience, particularly in temporal and frontal regions. This group was similar in age to the canonical group but had less WMH. However, Group 5 did not have significant difference in MMSE from the canonical group. Finally, Group 6 (n=11) had widespread higher atrophy relative to tau. This group was marginally older than the canonical group, but displayed significantly higher WMH volume (p=0.01) and lower MMSE. Some individuals within this group had quite severe WMH, as exemplified by the case in Figure 2. The entirety of the T-N regression residual data used for clustering is shown in Figure 3, along with a summary of measure of percentage of participants that had a suprathreshold residual for each ROI.

Figure 2:

Median ROI-wise group residual maps (columns 1–3). More saturated color represents higher residual, indicating higher (red) or lower (blue) atrophy relative to tau burden. Column 4 shows representative examples with arrows indicating features characteristic of the group.

Figure 3.

T-N regression residuals for each ROI and each participant visualized on a heatmap. ROIs on the x-axis are sorted by lobes, participants on the y-axis are sorted by clusters. Shades of red show positive residuals (more atrophy) and shades of blue show negative residuals (more tau). Stacked bar graph in the top panel shows percentage of ROIs showing “mismatch”, as defined by the residual being greater than 1.5 standard deviation away from the regression line.

We also performed an analysis of longitudinal changes in MMSE across the groups using linear mixed effects model. Followup period of cognitive data used was between 1–4 years and a maximum of 4 time points. Time*cluster interaction was entered as a predictor and participants as random effect; covariates included age, gender and years of education. Time*cluster interaction was significant (p = 0.04) indicating the groups differed in their rate of decline.

Sensitivity analyses

We conducted the following sensitivity analyses to assess the robustness of the proposed residual-based approach: 1) When T and N measures were averaged across left and right hemispheres, the T-N mismatch metric had a correlation of r=0.99 with that using unilateral measures, and clustering produced groups with similar characteristics, with the vast majority of subjects continuing to be clustered together. This is pictorially shown in a tanglegram display in Figure 5(top), 2) When non-discretized residuals were directly fed into hierarchical clustering, again qualitatively similar groups were obtained, as shown in the tanglegram display in Figure 5(bottom), 3) When we weighted the contribution of each ROI (# of suprathreshold residuals) with ROI size, the resulting T-N mismatch metric was still highly correlated (r=0.97) with those without such weighting.

Figure 5.

Tanglegram displays comparing cluster memberships. Top panel: comparison memberships between using left and right hemispheric ROI measures separately (left) vs. using averaged measures (right). Bottom panel: comparison using discretized regression residuals (left) vs. the raw continuous regression residuals (right). In both panels, color for each cluster on the right was chosen to be similar to the cluster on the left that had the most overlap in membership for ease of visualization. Cluster labels (1–6) are arbitrary, as output by the clustering algorithm.

Discussion:

Explicitly modeling the variability in the relationship between tau burden and neurodegeneration can be a useful tool for understanding phenotype and may provide important clues to additional factors that modulate neurodegeneration and cognitive decline. Determining such factors, e.g., the presence of co-pathology, is essential in both clinical research and practice to achieve a precision-based approach to diagnosis, prognosis, and management. While the dichotomous designation of biomarkers in the A/T/(N) framework has proven to be a powerful concept for AD classification, it fails to utilize the continuous and spatially varying information that in-vivo multimodal imaging can provide. Here we demonstrate that a measure of deviation from a normative relationship between tau burden and neurodegeneration across brain regions in individuals on the AD continuum captures variability due to multiple underlying factors.

The basic conceptual principle motivating the approach described here is that NFT pathology is tightly linked to AD-related neurodegeneration, as has been demonstrated in numerous in vivo and ex vivo studies^26–30. In contrast, measures of beta-amyloid plaque pathology do not strongly correlate with neurodegeneration³¹. Thus, the load of NFT pathology in the setting of AD (i.e. presence of amyloid and tau pathology as defined by the A/T/(N) framework) should provide an estimation of the degree to which brain structure would be expected to be affected specifically by AD. Deviation from this expectation would suggest that additional factors such as co-pathology or brain reserve are modulating this relationship. While this assumption may only approximate reality and more complex relationships may be present, the basic notion is likely to be correct and is consistent with the findings of this study.

T-N mismatch metric has meaningful associations

As presented in Tables 1 and 2, a simple scalar measure summarizing the degree of mismatch between the amount of tau pathology and neurodegeneration across the whole cerebrum is a useful metric to study underlying factors that may contribute to the heterogeneity of AD. This scalar measure used spatial information to account for the extent of this mismatch across the brain (i.e. the number of ROIs), but is insensitive to the spatial pattern. Nonetheless, we found that the T-N mismatch metric was significantly correlated with two factors that are likely to be drivers of atrophy beyond that of AD-related pathology; age³² and cerebrovascular disease^33,34, as measured by WMH burden³⁵. Moreover, the extent of this mismatch was associated with the degree of cognitive impairment across several domains, even after controlling for age and ¹⁸F-Flortaucipir uptake in the inferior temporal cortex, a surrogate for NFT burden. As one would predict, more atrophy than expected for a given level of tau was associated with poorer cognition while less atrophy was associated with better performance in the face of the same tau burden. The fact that the T-N mismatch metric did not correlate with ¹⁸F-Flortaucipir uptake in key AD-related regions supports the notion that our measure of mismatch is not confounded by disease severity in a non-linear relationship with structural neurodegeneration.

The current result supports a prior study by Ossenkoppele et al.³⁶ which used a global measure of T and N, whole cortical ¹⁸F-Flortaucipir uptake and mean cortical thickness, respectively, and a global cognitive measure, MMSE, to calculate a brain and cognitive reserve metric. Consistent with our findings, Ossenkoppele et al. reported that brain reserve was associated with age, such that younger individuals had more brain reserve, and that cognitive reserve was related to cortical thickness. Additionally, they found that female sex was associated with higher brain reserve. We also found a trend towards sex related to our continuous measure of mismatch. Thus, the findings here provide convergence in an independent dataset, but also demonstrate that these relationships occur independent of overall tau burden. Further, the sensitivity analyses using non-discretized residuals and bilaterally averaged ROI measures points to the robustness of the general approach of using T-N residuals.

T-N mismatch metric can potentially reveal phenotypes

The scalar T-N mismatch metric captures the extent of, but not the spatial pattern of mismatch. This pattern may provide additional insights, particularly related to the presence of co-pathology in cases with more neurodegeneration than expected for a level of tau pathology. To explore this, we used a data-driven approach to cluster individuals based on the spatial pattern of T-N mismatch. The clustering yielded six groups with the largest, not surprisingly, demonstrating little T-N mismatch that we referred to as a canonical group. Three groups (2,4, and 6) displayed greater neurodegeneration than expected for the level of tau, but differed in spatial patterns. One group in particular (Group 2) had excessive atrophy in temporal-limbic structures, particularly anterior hippocampus and temporal pole. While requiring pathologic confirmation in future work, this group had a pattern of atrophy suggestive of concomitant TDP-43 pathology consistent with descriptions of limbic-predominant age-associated TDP-43 encephalopathy^10,11, or TDP-43 Type C pathology which is associated with semantic dementia³⁷. The two other groups with higher levels of neurodegeneration had differing degrees of severity, but were generally diffuse with one group associated with high levels of WMHs. Post-hoc analysis of longitudinal changes in MMSE also found a significantly steeper rate of decline in these groups compared to the canonical group, further supporting the notion that they represent vulnerable phenotypes.

While we expected evidence of resilience as well, we found two different groups based on pattern of regions that had greater cortical thickness than expected. These groups may be explained by having higher brain reserve, or relative resistance to the presence of NFT pathology, perhaps through less neuropil or neuronal dropout³⁸. The separation of these two groups may reflect different mechanisms of resilience that if studied on a larger scale could provide important insights into environmental or genetic factors that could contribute to reducing risk or slowing disease progression.

This data-driven approach needs to be validated using post-mortem pathological analysis, to determine if the imaging phenotypes found indeed map on to disease phenotypes. Nonetheless, it has implications for clinical interventions and prognosis, particularly as a surrogate for neurodegenerative pathologies for which no in-vivo biomarkers exist³⁹. On the other hand, this approach can also help researchers identify and study resilience or brain reserve as a phenotype. If validated with longitudinal cognitive data, this also may be a valuable prognostic tool for clinicians.

Heterogeneity of underlying phenotypes is a significant factor that reduces power in clinical trials⁴⁰. As an increasing number of new therapies that specifically target pathological tau species are tested⁴¹, this approach may help exclude trial participants that might have significant non-tau co-morbidity (such as vascular disease or other proteinopathy) that might be driving their neurodegeneration. This could potentially reduce cost and provide a clearer understanding of the efficacy of a potential therapy.

Limitations and future work

The approach presented here for utilizing the richer spatial, continuously varying, information about T and N biomarkers to study variability is necessarily limited by the composition of the study cohort. The current composition of the ADNI research cohort has relatively strict inclusion criteria that limits the number of underlying factors that can be examined, including a relatively restricted range of disease severity. Future work will apply this framework to more heterogeneous cohorts that encompass a more diverse phenotypic landscape, including earlier onset disease⁴², non-amnestic presentations, and greater inclusion of groups with comorbidities, such as vascular risk factors such as diabetes.

Alternative approaches to defining a T-N mismatch metric could also be considered. The use of ROI-based regression residuals, as we have done, is easy to operationalize and implement. One could use a similar approach using voxel-level regressions instead. Another way to estimate a mismatch would be to first build a model of direct image-to-image (tau PET to MRI) translation using deep neural networks^43,44 instead of using biomarker measurements from both modalities as we have done here, and then use a metric of deviation from the predicted model output as a mismatch measure.

Any data-driven clustering approach is exploratory in nature, and the results will likely vary depending on cohort composition and the underlying disease phenotypes. Nonetheless, it is encouraging that the clusters that emerged based on T-N residuals appear to differ from each other based on underlying factors other than the pair of imaging measurements used to derive them. These factors included those likely related to vulnerability and resilience, e.g. a marker of vascular disease (white matter hyperintensity) and age respectively.

Finally, the relatively small cohort used here limits strong conclusions about the stability of these groups, but does provide a proof-of-principle that accounting for the relationship of T and N provides phenotypic groupings sensitive to biological modifiers of these relationships. In addition to more heterogeneous populations, future studies will need to apply this approach to larger cohorts and study longitudinal outcomes to determine the clinical significance of mismatch. Ultimately, the study of cross-modality relationships of different biological constructs, including A, T, and N, may allow for quantitative assessments linked to disease heterogeneity⁴⁵.

Figure 4.

Trajectory of cognitive decline for different groups of participants. Slopes shown for each group is estimated by linear mixed-effects model. Groups 2 and 6 are the vulnerable groups showing steeper decline than group 1 (canonical). Groups 3 and 5 are the resilient groups. Group 4 is not shown because of the small group size. Note that the variable maximum time periods for different groups reflect differences in duration of followup data available.