The bivariate distribution of amyloid-β and tau: relationship with established neurocognitive clinical syndromes.

Large phenotypically diverse research cohorts with both amyloid and tau PET have only recently come into existence. Our objective was to determine relationships between the bivariate distribution of amyloid-β and tau on PET and established clinical syndromes that are relevant to cognitive ageing and dementia. All individuals in this study were enrolled in the Mayo Clinic Study of Aging, a longitudinal population-based study of cognitive ageing, or the Mayo Alzheimer Disease Research Center, a longitudinal study of individuals recruited from clinical practice. We studied 1343 participants who had amyloid PET and tau PET from 2 April 2015 to 3 May 2019, and met criteria for membership in one of five clinical diagnostic groups: cognitively unimpaired, mild cognitive impairment, frontotemporal dementia, probable dementia with Lewy bodies, and Alzheimer clinical syndrome. We examined these clinical groups in relation to the bivariate distribution of amyloid and tau PET values. Individuals were grouped into amyloid (A)/tau (T) quadrants based on previously established abnormality cut points of standardized uptake value ratio 1.48 (A) and 1.33 (T). Individual participants largely fell into one of three amyloid/tau quadrants: low amyloid and low tau (A-T-), high amyloid and low tau (A+T-), or high amyloid and high tau (A+T+). Seventy per cent of cognitively unimpaired and 74% of FTD participants fell into the A-T- quadrant. Participants with mild cognitive impairment spanned the A-T- (42%), A+T- (28%), and A+T+ (27%) quadrants. Probable dementia with Lewy body participants spanned the A-T- (38%) and A+T- (44%) quadrants. Most (89%) participants with Alzheimer clinical syndrome fell into the A+T+ quadrant. These data support several conclusions. First, among 1343 participants, abnormal tau PET rarely occurred in the absence of abnormal amyloid PET, but the reverse was common. Thus, with rare exceptions, amyloidosis appears to be required for high levels of 3R/4R tau deposition. Second, abnormal amyloid PET is compatible with normal cognition but highly abnormal tau PET is not. These two conclusions support a dynamic biomarker model in which Alzheimer's disease is characterized first by the appearance of amyloidosis and later by tauopathy, with tauopathy being the proteinopathy associated with clinical symptoms. Third, bivariate amyloid and tau PET relationships differed across clinical groups and thus have a role for clarifying the aetiologies underlying neurocognitive clinical syndromes.

See Gordon and Tijms (doi:10.1093/brain/awz278) for a scientific commentary on this article.

Introduction

The first successful amyloid PET ligand was reported 15 years ago (Klunk ), and over the past decade and a half investigators in the ageing and dementia field have thoroughly examined relationships between amyloid PET and various clinical presentations (Rabinovici ; Nordberg ; Villemagne ; Ossenkoppele ; Bilgel ; Buckley ; Gordon ; Jansen ; Lim ; Lopez ; Leuzy ; Timmers ). More recently, relationships between tau PET and various clinical presentations have also been reported (Cho ; Johnson ; Ossenkoppele , 2018; Scholl ; Chiotis ; Maass ; Gordon ; Sperling ). Because tau PET was only recently introduced into the academic community (Chien ), very large phenotypically diverse research cohorts with both amyloid and tau PET have only recently come into existence. Our main objective was to determine relationships between the bivariate distribution of amyloid-β and tau on PET and established clinical syndromes relevant to cognitive ageing and dementia: cognitively unimpaired, mild cognitive impairment (MCI), frontotemporal dementia (FTD), probable dementia with Lewy bodies (DLB), and Alzheimer clinical syndrome (AlzCS).

Materials and methods

Enrolment and clinical characterization

This study was approved by the Mayo Clinic and Olmsted Medical Center Institutional Review Boards. Written informed consent was obtained from all participants and in the case of persons with cognitive impairment sufficient to interfere with capacity, from a close family member.

All individuals in this study were enrolled in one of two studies. The Mayo Clinic Study of Aging (MCSA) is a longitudinal population-based study of cognitive ageing among a stratified random sample of a geographically- defined population (Roberts ). Residents of Olmsted County, Minnesota, USA aged 30–89 years old were enumerated using the medical records-linkage system of the Rochester Epidemiology Project (St Sauver ). From this sampling frame, individuals were randomly selected by 10-year age and sex strata such that males and females were represented equally. All individuals without a medical contraindication were invited to participate in imaging studies. The Mayo Alzheimer Disease Research Center (ADRC) is a longitudinal research study of individuals recruited from clinical practice.

Evaluations included a medical history review and interview with the participant and a study partner, a neurological examination by a physician; and a neuropsychological examination (Roberts ). Participants were assigned a diagnosis of cognitively unimpaired (defined as not MCI or dementia), MCI (Petersen, 2004), or dementia (American Psychiatric Association, 1994) using established criteria. Among individuals with dementia, syndromic diagnoses followed established criteria for FTD syndromes (Gorno-Tempini ; Rascovsky ), probable DLB (McKeith ), and AlzCS (McKhann , 2011; Jack ). The AlzCS group included individuals with a classic amnestic phenotype who would meet clinical criteria for probable Alzheimer’s disease (McKhann , 2011), as well as established atypical Alzheimer’s disease phenotypes (i.e. language, visuospatial, and dysexecutive) (Wolk ; Jones ). Clinical diagnoses were always made blinded to PET results in the MCSA; however, this was not always the case for ADRC participants.

The study sample consisted of 1343 participants who met the following criteria (Supplementary Fig. 1). For inclusion an individual must have been a participant in the MCSA or ADRC, had amyloid PET, tau PET, and MRI (MR is used for PET quantification) from 2 April 2015 to 3 May 2019, and have met criteria above for membership in one of the five clinical diagnostic groups described above. We excluded individuals who: (i) were cognitively impaired but did not meet established criteria for one of the five syndromic diagnostic groups above; (ii) could not be confidently labelled either cognitively impaired or unimpaired; (iii) had imaging studies that were inadequate for technical reasons; and (iv) were members of families with known mutations. We used data from the first imaging session for individuals with serial imaging.

Imaging methods

Amyloid PET imaging was performed with Pittsburgh Compound B (PIB) (Klunk ) and tau PET with flortaucipir (Chien ). Amyloid and tau PET standardized uptake value ratios (SUVRs) were formed by normalizing composite multi-region target regions of interest to the cerebellar crus grey matter (Jack ). The amyloid PET target meta-region of interest included the prefrontal, orbitofrontal, parietal, temporal, anterior and posterior cingulate, and the precuneus (Jack ). The tau PET target meta-region of interest used in our primary analysis included the amygdala, entorhinal cortex, fusiform, parahippocampal, and inferior temporal and middle temporal gyri (Jack ). As a post hoc sensitivity analysis, we examined three alternative tau PET reporter regions of interest highlighted in recent publications (Cho ; Johnson ; Scholl ; Maass ; Mishra ; Pontecorvo ; Lowe ). They were: (i) entorhinal cortex; (ii) inferior temporal gyrus; and (ii) a lateral parietal meta-region of interest composed of the angular, supramarginal, and inferior parietal regions of interest. PET data were not partial volume corrected. MRI was performed at 3 T and was used in the PET data processing pipeline, described in previous work (Schwarz ).

Statistical methods

Our main objective was to determine relationships between the bivariate distribution of amyloid-β and tau on PET and established clinical syndromes. To clearly illustrate this, the bivariate distribution of tau and amyloid PET SUVR values over the study sample was segmented into quadrants using cut points based on previously published research (Jack ). The amyloid PET cut point was the SUVR value 1.48 (centiloid 22; Klunk ), beyond which rates of amyloid PET reliably increased (Jack ). The cut point for tau PET (SUVR 1.33) was the value that most accurately discriminated between cognitively impaired individuals with abnormal amyloid and age-matched cognitively unimpaired individuals with normal amyloid (Jack ). Previously, we have described both lenient and conservative (i.e. labels fewer people abnormal) cut points for tau PET (Jack ) and used the latter here because the questions pertain to aetiological diagnoses in impaired individuals.

In addition to the above, a cluster analysis of the amyloid and tau PET SUVR values was performed using a model-based clustering approach (Banfield and Raftery, 1993; Scrucca ). The model treats the data as a mixture of Gaussian densities; it estimates the optimal number of clusters along with the centre and shape of each cluster by optimizing the Bayesian information criterion. An additional ‘diffuse’ cluster was added to the model that covers the entire range of amyloid and tau values, this serves to make the primary results resistant to any outlier points by assigning them to the random background (Banfield and Raftery, 1993).

Data availability

The Mayo Clinic Study of Aging and Alzheimers Disease Research Center make data available to qualified researchers upon reasonable request.

Results

Study participants

The majority of individuals in this study were cognitively unimpaired and most of these were participants in the MCSA (Table 1). MCI participants were split approximately evenly between the MCSA and ADRC. Nearly all individuals with dementia were participants in the ADRC. The probable DLB group had the highest proportion of males and the AlzCS group had the highest proportion of APOE ɛ4 carriers. Within the FTD group, 14 (61%) were diagnosed as behavioural variant FTD, seven (30%) as semantic dementia, one (4%) as progressive non-fluent aphasia, and one (4%) as progressive associative agnosia (Gorno-Tempini ; Rascovsky ).

Table 1

Demographic characteristics of study participants

Characteristic	CU	MCI	AlzCS	DLB	FTD
Number of subjects	976	182	123	39	23
Study, n (%)
MCSA	903 (93)	92 (51)	8 (7)	3 (8)	0 (0)
ADRCa	73 (7)	90 (49)	115 (93)	36 (92)	23 (100)
Age, years
Median (IQR)	68 (57, 77)	74 (67, 81)	69 (61, 77)	70 (66, 77)	65 (58, 69)
Min, Max	30, 98	44, 94	52, 89	51, 88	49, 73
Male sex, n (%)	526 (54)	117 (64)	51 (41)	31 (79)	14 (61)
Education, years
Median (IQR)	16 (13, 17)	16 (12, 18)	16 (12, 16)	15 (14, 18)	16 (15, 18)
Min, Max	6, 20	0, 24	8, 25	7, 20	12, 21
APOE ɛ4 carrier, n (%)	258 (28)	66 (39)	74 (69)	12 (44)	5 (29)
Short Test of Mental Status score
Median (IQR)	37 (35, 38)	33 (30, 35)	24 (19, 29)	29 (24, 32)	27 (23, 32)
Min, Max	26, 38	22, 38	4, 36	8, 35	20, 36
Amyloid PET
SUVR
Median (IQR)	1.38 (1.31, 1.51)	1.53 (1.36, 2.25)	2.47 (2.20, 2.71)	1.68 (1.37, 2.32)	1.36 (1.26, 1.45)
Min, Max	1.10, 3.63	1.15, 3.41	1.28, 3.41	1.25, 3.00	1.16, 2.16
Centiloid
Median (IQR)	13 (6, 24)	26 (11, 90)	110 (86, 131)	39 (12, 96)	11 (2, 19)
Min, Max	−13, 212	−8, 193	4, 193	1, 157	−7, 82
n (%) abnormalb	282 (29)	101 (55)	119 (97)	23 (59)	4 (17)
Tau PET, SUVR
Median (IQR)	1.18 (1.12, 1.23)	1.25 (1.18, 1.38)	1.97 (1.60, 2.29)	1.25 (1.17, 1.31)	1.23 (1.20, 1.31)
Min, Max	0.87, 1.75	0.91, 2.49	1.09, 3.27	1.05, 2.22	1.10, 1.40
n (%) abnormalc	54 (6)	55 (30)	109 (89)	7 (18)	4 (17)

CU = cognitively unimpaired.

aADRC includes 12 participants co-enrolled in the Advancing Research and Treatment for Frontotemporal Lobar Degeneration Project 1 Study (one MCI, two AlzCS, nine FTD).

bAmyloid PET abnormality is defined as ≥1.48 SUVR.

cTau PET abnormality is defined as ≥1.33 SUVR.

Primary analysis: clinical diagnostic groups within the bivariate amyloid and tau PET distribution

Figure 1 shows the bivariate amyloid and tau PET distribution with individual points colour-coded to represent the five clinical diagnostic groups. Despite a positive rank correlation between amyloid and tau PET overall (rho = 0.57, P < 0.001), rather than showing a prototypical ellipsoid bivariate relationship, few individuals were present in the upper left region. The marginal histograms for amyloid and tau in Fig. 1 illustrate that the distribution of amyloid PET values in the ADRC were bimodal, while in the MCSA they were unimodal with a long right tail even when plotted on log scale. The distribution of tau PET values in the ADRC was unimodal with a long right tail but was unimodal and roughly Gaussian in the MCSA.

Figure 1

Amyloid and tau PET distributions by clinical group overall and within study. Scatter plots of tau PET SUVR versus amyloid PET SUVR among all individuals combined (A) and separately among individuals in the MCSA (B) and ADRC (C). Tau PET and amyloid PET values are in SUVR units but the data is plotted on log scale, which accounts for the uneven spacing. Points are coloured by clinical diagnosis. Histograms in the margins show the distributions of tau PET SUVR (right) and amyloid PET SUVR (top). Axis labels on the top represent amyloid PET values on a centiloid scale.

In Fig. 2A we illustrate results of our clustering procedure based on a three-cluster model with data points coloured to indicate cluster. Black data points were identified as not consistent with any of the three clusters. Ellipses represent the centre 50% of data for each cluster distribution. Three clusters were chosen as optimal based on Bayesian information criterion of −969, −1338, −1405, −1401, and −1367 for one, two, three, four, and five clusters, respectively. We also superimpose previously established amyloid and tau PET cut points that segregate the bivariate distribution into four quadrants: normal amyloid and normal tau (A−T−), abnormal amyloid and normal tau (A+T−), abnormal amyloid and abnormal tau (A+T+), and normal amyloid and abnormal tau (A−T+). The clustering also suggested a quadrant-based interpretation that is similar to the cut point based quadrants. For both methods few individuals were in the upper left quadrant.

Figure 2

Amyloid and tau PET clusters. Scatterplot of tau PET SUVR versus amyloid PET SUVR with points coloured according to the three-cluster classification from a bivariate mixture model (A). Points shown in black represent individuals who were inconsistent with one of the three clusters. The vertical and horizontal lines represent the cut points of 1.48 SUVR for amyloid PET and 1.33 SUVR for tau PET. The ellipses show the centre 50% of the data for the three cluster distributions with a black star indicating the bivariate mean from the clustering. In B, these ellipses are shown along with a square for each clinical diagnosis group representing the bivariate median (centroid) of the tau and amyloid distributions. Tau PET and amyloid PET values are in SUVR units but the data is plotted on log scale, which accounts for the uneven spacing. Axis labels on the top represent amyloid PET values on a centiloid scale.

The cluster ellipses and standard cut points are shown again in Fig. 2B with the bivariate medians (centroids) of the clinical groups added. The centroids of the cognitively unimpaired and FTD groups were located in the A−T− quadrant and within what we call the low-low cluster ellipse. The AlzCS centroid was in the A+T+ quadrant and what we call the high-high cluster ellipse. The probable DLB centroid fell in the A+T− quadrant and within what we call the high-low cluster ellipse. The MCI centroid was on the border between the A−T− and A+T− quadrants and situated between the ellipses for the low-low and high-low clusters.

Figure 3 shows the proportion of individuals in each of the five clinical groups that fall into the cut point based and cluster based quadrants. Using established cut points (Jack ) most cognitively unimpaired individuals fell into either the A−T− (70%) or A+T− (25%) quadrants, but a few were in the A+T+ (4%) and A−T+ (1%) quadrants. Among MCI, 42% were in the A−T− quadrant, 28% in the A+T− quadrant, 27% in the A+T+ quadrant, and only 3% in the A−T+ quadrant. The large majority of AlzCS individuals were in the A+T+ quadrant (89%) but a few were in the A−T− (3%) and A+T− (8%) quadrants. Probable DLB individuals were generally in either the A−T− (38%) or A+T− (44%) quadrants, while the majority of the FTD individuals were in the A−T− quadrant (74%) (Supplementary Table 1). These patterns were similar for the cluster-based quadrants.

Figure 3

Amyloid and tau PET groups within clinical diagnosis. Per cent of individuals in each quadrant (A) or cluster (B) within each clinical diagnostic group. (A) Percentages according to amyloid and tau PET groupings based on the established cut points of 1.48 SUVR for amyloid PET and 1.33 SUVR for tau PET. (B) Percentages according to the bivariate mixture model clusters. These are labelled according to amyloid (low or high) and tau (low or high). Those individuals whose values were inconsistent with one of the three clusters were labelled as other.

Secondary analysis: bivariate amyloid and tau PET distribution by age among unimpaired, MCI, and AlzCS individuals

In Fig. 4 we illustrate the bivariate amyloid and tau PET distribution by age restricted to cognitively unimpaired, MCI and AlzCS participants. In the MCSA, nearly all (97%) individuals under 60 years of age were A−T− while those above 60 fell in each of the A−T− (55%), A+T− (35%), and A+T+ (9%) quadrants. In contrast, younger individuals were more equally present in both the A−T− (46%) and A+T+ (49%) quadrants in the ADRC.

Figure 4

Scatter plots of tau PET SUVR versus amyloid PET SUVR by age groups among cognitively unimpaired (CU), MCI, and AlzCS individuals. MCSA individuals are shown in the left column and ADRC individuals in the right column. The vertical and horizontal lines represent the cut points of 1.48 SUVR for amyloid PET and 1.33 SUVR for tau PET. Points are coloured by clinical diagnosis. Tau PET and amyloid PET values are in SUVR units but the data is plotted on a log scale, which accounts for the uneven spacing. Axis labels on the top of the columns represent amyloid PET values on a centiloid scale.

Post hoc sensitivity analyses

In Supplementary Fig. 2 we illustrate correlations between different tau PET reporter regions of interest. The temporal meta-region of interest used in the primary analysis was highly correlated with the entorhinal cortex (rho = 0.86), inferior temporal (rho = 0.99), and lateral parietal (rho = 0.91) regions of interest. The entorhinal cortex had a ceiling effect relative to the other regions of interest. Figure 5 illustrates the bivariate amyloid and tau PET to clinical diagnostic group relationships for these three alternative tau PET regions of interest. Except for the ceiling effect in the entorhinal cortex, the bivariate distributions were similar to that of the temporal meta-region of interest.

Figure 5

Scatter plots of tau PET SUVR versus amyloid PET SUVR among all individuals combined and separately among individuals in the MCSA and ADRC for three alternative tau PET regions of interest. Points are coloured by clinical diagnosis. Histograms in the margins show the distributions of tau PET SUVR (right) and amyloid PET SUVR (top). Axis labels on the top of each plot represent amyloid PET values on a centiloid scale. Tau PET and amyloid PET values are in SUVR units but the data is plotted on a log scale, which accounts for the uneven spacing.

Discussion

The clinical diagnoses in this sample were based on well-defined syndromic presentations. Each syndrome in turn is linked with an expected aetiology. For example, the expected aetiology in probable DLB is α-synuclein pathology, and in AlzCS it is Alzheimer’s disease. Patient management decisions are driven by the aetiologies that physicians suspect underlie specific syndromic presentations (Pontecorvo ; de Wilde ; Rabinovici ). However, a clinical syndrome is often an imperfect indicator of underlying aetiology (Beach ; Jagust ). Our data illustrate that the combination of amyloid and tau PET provides useful information about the aetiology underlying various common neurocognitive syndromes at the individual patient level.

The A+T+ biomarker profile identifies Alzheimer’s disease in vivo in the NIA-AA research framework (Jack ) and about 9 in 10 AlzCS participants in this study were A+T+. The 8% (Supplementary Table 1) of AlzCS participants who were A+T− would be labelled Alzheimer pathological change in the NIA-AA research framework (Jack ). Low levels of Alzheimer’s disease tauopathy that lie below the tau PET detection threshold could account for a portion of the symptoms in A+T− individuals (Knopman ). However, the dementia is likely due, at least in part, to one or more ageing-related non-Alzheimer disorders, e.g. TDP-43 (recently labelled LATE disease; Nelson ) with or without hippocampal sclerosis, α-synuclein deposits, argyrophilic grain disease, or cerebrovascular disease (particularly microscopic cerebral infarctions) (Schneider ; Sonnen ; Nelson ; Kawas ; Botha , ; Petersen, 2018). In A−T− individuals diagnosed clinically as AlzCS, symptoms are almost certainly due to one or more of the aforementioned non-Alzheimer disorders.

Probable DLB participants were distributed between the A−T− and A+T− quadrants (Fig. 3). The A−T− individuals likely represent more pure α-synuclein pathology while the A+T− individuals represent α-synuclein pathology plus Alzheimer pathological change (Walker ; Irwin ; Coughlin ). Because α-synuclein pathology and Alzheimer’s disease often coexist, patients may present with overlapping clinical features (Walker ). The combination of amyloid and tau PET may aid the clinician by indicating the status of a patient within the Alzheimer continuum and thereby informing on aetiologies underlying symptoms. Many probable DLB patients also meet neuropathological criteria for Alzheimer’s disease at autopsy (Walker ; Irwin ) and therefore a higher number of A+T+ probable DLB patients might have been expected in our study. However, flortaucipir retention is greater in AlzCS than in probable DLB in most areas of the brain (Gomperts ; Kantarci ) and when tau is present in pathologically confirmed DLB, the pathological burden is usually lower than it is in Alzheimer’s disease (Walker ; Coughlin ). We used a conservative cut point for tau PET (Jack ) here, and the tau burden in mildly affected probable DLB (median Short Test of Mental Status score 29, Table 1) patients may have fallen below this cut point.

Most FTD participants were A−T−, which was anticipated as these individuals are not considered to lie in the Alzheimer continuum (Fig. 3). Four, however, had high amyloid PET levels. All four were over age 60 and may have had fronto-temporal lobar degeneration plus Alzheimer continuum pathology (Jack ). Discrimination between fronto-temporal lobar degeneration and Alzheimer’s disease on clinical grounds can occasionally be difficult and this distinction may be aided by joint amyloid and tau PET. Flortaucipir was developed to bind to the 3R/4R form of pathological tau that characterizes Alzheimer’s disease (Chien ; Lowe ; Lemoine ; Marquie ). If a tau PET ligand is developed that is sensitive to the tau deposits in primary tauopathies, then many patients in the FTD spectrum would likely exhibit an A−T+ biomarker profile when imaged with such a ligand.

MCI participants were distributed across the A−T−, A+T−, and A+T+ quadrants (Fig. 3). This reflects the aetiological heterogeneity of the MCI syndrome (Markesbery ). A+T+ and A+T− individuals are considered to lie within the Alzheimer continuum (Jack ) and aetiologies underlying cognitive impairment in these participants are likely similar to those discussed above for AlzCS individuals (Forsberg ). A−T− MCI participants may have non-Alzheimer neuropathology or non-degenerative causes for impairment (e.g. sleep disorders, depression, cerebrovascular disease) (Wisse ; Landau ) and the A−T− biomarker profile suggests excluding Alzheimer’s disease as a likely aetiology. The potential value of these biomarkers in patient management is evident. Physicians may be likely to prescribe Alzheimer’s disease medications for an A+T+ MCI patient, but unlikely to prescribe these medications for a patient with the same syndromic presentation but who is A−T− (Pontecorvo ; Rabinovici ).

Among cognitively unimpaired participants, 70% were A−T− and 25% were A+T−. While some cognitively unimpaired individuals had amyloid PET values in the same range as AlzCS, none had tau PET values in the AlzCS range. Further, the few cognitively unimpaired individuals with tau PET values above the tau PET cut point had values that were close to the normal/abnormal cut point (Fig. 4). This suggests that high levels of amyloid are compatible with unimpaired cognitive status while high levels of tau are not, a finding that is consistent with prior evidence that cognitive impairment in Alzheimer’s disease is more closely linked to tau than amyloid-β (Arriagada ; Gomez-Isla ; Bennett ; Nelson ; Ossenkoppele ).

Which biomarker abnormality appears first, amyloid versus tau, has been a source of controversy in the Alzheimer field. Because earlier events can cause later events but not the reverse, temporal ordering of biomarkers has likewise been linked to discussion of cause and effect. Our data are cross-sectional and cannot prove specific claims about temporal evolution of biomarkers or cause and effect; however, the prevalence of various amyloid and tau profiles in this large series of participants can inform on the ‘plausibility’ of different possible temporal evolution patterns. Every individual begins life cognitively unimpaired and A−T− (lower left quadrant) and the large majority of AlzCS individuals are A+T+ (upper right quadrant) (Figs 3 and 4). Therefore every A+T+ AlzCS individual had to follow one of the following transition pathways: (i) A−T− to A+T− to A+T+; (ii) A−T− to A−T+ to A+T+; or (iii) from A−T− directly to A+T+. The near absence of individuals in the A−T+ quadrant, along with the observation that all tau PET values in that quadrant lie close to the cut point, makes the second pathway very unlikely. The clear gap in data points along the middle of the 45° diagonal line makes the last option unlikely as well. The placement of clinical group centroids on the tau versus amyloid plot (Fig. 2B) couples the well-established clinical progression of cognitively unimpaired to MCI to AlzCS with a parallel pathway from A−T− to A+T− to A+T+. Finally, Fig. 4 illustrates the bivariate amyloid and tau PET distribution for different ages among cognitively unimpaired, MCI, and AlzCS participants. The clear trend is progression from A−T− to A+T− with age in many participants, with some individuals reaching A+T+. In aggregate these data lead to the conclusion that the typical pathway within the Alzheimer aetiological continuum from cognitively unimpaired to MCI to AlzCS is likely paralleled by transitions from A−T− to A+T− to A+T+. When combined with the evidence suggesting that high levels of amyloid are compatible with unimpaired cognitive status while high levels of tau are not, our data supports the position that within the Alzheimer continuum it is the transition from A+T− to A+T+ that is associated with onset of severe cognitive symptoms. Our data along with that from other studies (Arriagada ; Ingelsson ; Bateman ; Benzinger ; Xiong ; Leal ; Hanseeuw ) support a dynamic biomarker model (Jack , 2013) in which amyloidosis promotes or is permissive for the spread of tauopathy and it is the latter proteinopathy that is associated with cognitive symptoms.

The distributions of tau PET values using different reporter regions of interest were highly correlated, particularly the temporal meta-region of interest and the inferior temporal region of interest (Supplementary Fig. 2). Aside from the tau PET ceiling effect in the entorhinal cortex, the bivariate distribution was not greatly different using the entorhinal cortex compared to that seen with other tau reporter regions of interest (Fig. 5), which would imply that the typical progression from A−T− to A+T− to A+T+ is a generalizable conclusion across different choices of tau reporter region of interest.

An important caveat to the conclusion above is that autopsy data show that medial temporal tauopathy likely precedes neocortical amyloidosis in most people (Braak ). Medial/basal temporal lobe tauopathy with no or minimal amyloidosis has been labelled primary age-related tauopathy (Crary ), a neuropathological entity attributed to ageing that is distinct from Alzheimer’s disease (Hyman ). It is unclear how well flortaucipir or any currently available tau PET ligand captures low levels of tau pathology in primary age-related tauopathy.

This study has several strengths. First, the large number of participants (n = 1343) in this sample lends credibility to the conclusion that a fundamental feature of the biology of Alzheimer’s disease is that amyloidosis is required for extensive 3R/4R tauopathy. Second, all participants were well characterized clinically by the same group of physicians, neuropsychologists, and study coordinators and variability was further reduced by using the same set of magnetic resonance and PET scanners and identical image analysis pipelines. Third, we examined both population-based and clinic-based samples. Population-based samples capture relationships between clinical presentation and imaging that exist commonly in community dwelling individuals, but will not capture more uncommon clinical presentations. In contrast, our clinic-based sample captures less common clinical presentations. When combined, we were able to study the full spectrum of the bivariate amyloid and tau distribution.

This study has several limitations. First, in vivo imaging has fundamental limits in detection sensitivity in comparison to direct examination of tissue at autopsy. Sub-threshold levels of both amyloid (Leal ) and tau may be clinically relevant. We cannot know in any individual whether or not subthreshold levels of either proteinopathy are present. Second, the findings are specific for the two PET tracers used and could differ somewhat with different amyloid and tau PET tracers. Third, cut points for amyloid PET have been validated against neuropathology (Murray ; La Joie ). However, because of its more recent development far less autopsy-imaging validation exists for tau PET. Questions about which areas of the brain should be measured and what constitute valid or useful cut points are still under investigation (Cho ; Johnson ; Scholl ; Villemagne ; Maass ; Mishra ; Pontecorvo ; Lowe ). However, the continuous bivariate amyloid and tau PET distributions (Figs 1, 2, 4 and 5) circumvent the limitations of cut points and illustrate that the primary conclusion that very high tau is not seen in the absence of amyloid would not change with different cut points. Fourth, while clinical diagnoses were all made blinded to PET results in the MCSA, this was not always the case for participants in the ADRC (given its clinic-based nature). However, all participants met rigorously defined, established clinical diagnostic criteria that do not require imaging/biomarker information. Fifth, many individuals presenting to the ADRC with AlzCS (a referral clinic sample) are not representative of AlzCS in the community. The former tend to have a younger age of onset and higher levels of neocortical tau uptake (Ossenkoppele ; Lowe ). Sixth, perhaps the most significant limitation, is one shared by all research in this field. Currently available biomarkers that are disease-specific do not fully capture the neuropathological complexity of the ageing brain. Non-Alzheimer pathologies, for which no biomarkers presently exist, are highly prevalent in older individuals (most often co-morbidly), and exert significant effects on cognition (Schneider ; Sonnen ; Nelson ; Kawas ; Botha , ; Petersen, 2018). A major need in the field is the development of either imaging or biofluid biomarkers for these common non-Alzheimer neuropathological entities.

Click here for additional data file.