Antemortem Prediction of Braak Stage

Abstract

We examined the extent to which tauopathy distribution, as determined by Braak staging, might be predicted by various risk factors in older individuals. The Swedish Twin Registry provided extensive information on neuropsychological function, lifestyle and cardiovascular risk factors of 128 patients for whom autopsy data including Braak staging were available. Logistic regression was used to develop a prognostic model that targeted discrimination between Braak stages 0-II vs. III-VI. The analysis showed that Braak stage III-VI was significantly predicted by having one or more APOE ε4 alleles, older age, high total cholesterol, absence of diabetes and cardiovascular disease, and poorer scores on the Wechsler Adult Intelligence Score Information test, verbal fluency, and recognition memory but better verbal recall. The algorithm predicted Braak stage III-VI well (receiver-operating characteristic area under curve: 0.897; 95% CI: 0.842-0.951). Using a cut-off of 50% risk or more, the sensitivity was 85%, the specificity was 70%, and the negative predictive value was 69%. This study demonstrates that tauopathy distribution can be accurately predicted using a combination of antemortem patient data. These results provide further insight into tauopathy development and AD-related disease mechanisms and suggest a prognostic model that predicts the spread of neurofibrillary tangles above the transentorhinal stage.

INTRODUCTION

Cortical tauopathy is frequently observed in the brains of healthy and non-healthy aged people (1-3). Tau deposition in the form of neurofibrillary tangles (NFT) is often described as part of the spectrum of Alzheimer disease (AD), together with β-amyloid (Aβ) and neuritic plaques. As there is still controversy as to the disease-specificity of tauopathy in AD, the recent NIA recommendations on neuropathologic criteria remark that “there is a consensus to disentangle the clinicopathological (diagnostic) term AD from AD neuropathologic changes” (4). Emerging evidence in recent neuropathological studies also demonstrate “NFT+/Aβ-brains,” in which tauopathy is predominantly observed without amyloidosis (3, 5). Suspected non-Alzheimer pathophysiology or primary age-related tauopathy have been proposed as new concepts describing subjects within the low range of Braak staging. There is no consensus on this emerging nosological nomenclature and there are no formal diagnostic criteria for non-AD tauopathies but there is, “broad agreement in numerous clinicopathologic studies that the extent of NFT accumulation correlates with severity of dementia”, better than the amyloid burden (4). Moreover, recent data suggest that Braak stage is the best predictor of age of onset and disease duration in AD patients (6). Therefore, given the prevalence of tauopathy, 2% to 10% (7), and its implication in both AD and non-AD conditions, it is becoming more desirable to predict the distribution of cortical tauopathy antemortem (3), regardless of etiology.

Many disease-influencing factors have already been identified for AD and dementia (8-13), whereas risk-factors for non-AD tauopathy are less well studied (14). Risk factors and their correlation to clinical diagnoses or brain atrophy have been extensively studied. Prognostic risk models for clinical AD have been developed using, among other outcomes, cerebrospinal fluid biomarkers (15-17), complement system proteins (18), various types of neuroimaging techniques (19-22), and neuropsychological assessments (23, 24). A more general prognostic risk score for dementia based on data from various domains of assessment has been difficult to construct because few studies have simultaneous access to these factors (8). The risk score proposed by Kivipelto et al estimates the probability of developing dementia by simultaneously incorporating lifestyle, genotype and cardiovascular factors (25). The discriminability of such models is usually high but they share the shortcoming of using clinical or neuropathologic diagnostic labels as the outcome variable rather than considering the actual neuropathology. To date, while there are risk scores developed for clinical dementia, there seem to be no prognostic studies particularly aimed at predicting NFT distribution using any of these types of data.

A number of ongoing prospective clinicopathological cohort studies are evaluating neuropathological correlates of clinical diagnoses such as dementia or AD (18, 22, 26-31). These studies incorporate a variable range of antemortem data but few have analyzed the predictability of autopsy-confirmed neuropathologic changes such as tauopathy, as categorized by Braak staging (32, 33). Moreover, only a minority of these studies include data on neuropsychological status, which is central to the symptomatology of dementia. Indeed, measurements of cognitive function, in particular the decline thereof, have been shown to correlate well with Braak staging (28, 34-37). Braak staging is a validated neuropathologic method to assess the spread of cortical NFT (38), and meta-analyses of autopsy studies also reveal that it currently is the most significant neuropathologic correlate of dementia (6, 32, 33).

In the present study, we evaluated the predictability of neuropathologic Braak staging from life-style, genetic and neuropsychological data. We also provide an algorithm for Braak stage estimation that discriminates individuals based on tauopathy distribution in terms of likelihood of Braak stage 0-II vs. III-VI. This dichotomization is based on the increased degree of clinical dementia for Braak stages III-VI and on the neuroanatomical difference between the presence of tau only in the entorhinal cortex (Braak I-II) vs. its presence more widely in limbic areas or neocortex (Braak III-VI).

MATERIALS AND METHODS

Participants and Data

Participants were drawn from 4 related twin-studies of aging based on the population-based Swedish Twin Registry (STR) (39), recently described in detail (40). Data on lifestyle, cognitive functioning and health outcomes were derived from the Swedish Adoption/Twin Study of Aging (41), the Origins of Variance in the Old: Octogenarian Twins study (42), the Aging in Women and Men study (43), and the Study of Dementia in Swedish Twins (44). Inclusion and exclusion criteria are described in each study. Informed consent for postmortem brain autopsy was collected during the studies and included an invitation to participate not premised on cognitive status. There were 134 intact donated brains available from these studies. After brains lacking neuropathologic assessment or with insufficient antemortem data were excluded, the sample comprised 128 brains from patients with sufficient antemortem data that also had neuropathological examination including Braak staging (38). In this cohort, the proportion of subjects with a clinical diagnosis of dementia was 85%; of those, 73% fulfilled criteria for clinical AD.

Data on medical history obtained from enrolment in the studies or from national disease registries and included educational attainment (more than 7 years of schooling), diabetes diagnosis (ever/never), and cardiovascular disease (CVD) diagnoses (ever/never). CVD refers to a diagnosis either of angina pectoris, myocardial infarction, coronary heart disease or stroke or related subdiagnoses. The ages of onset of diabetes and CVD were not recorded, only whether there was ever a diagnosis. Cognitive assessments and medical workup from the 4 Swedish Twin Registry studies followed similar protocols including neuropsychological test batteries (40, 44). The data came from the in-person testing for the last-recorded assessment of each individual. Mean time between the last examination and death was 3.1 years, (SD = 2.53 years). Only those neuropsychological tests on which there were data on more than 60% of the participants were included: the Mini-Mental State Exam, the Consortium to Establish a Registry for Alzheimer's Disease (CERAD) word list test, verbal fluency, the Information subtest (Swedish version of the Wechsler Adult Intelligence Scale, WAIS), the verbal Symbol-Digit test of perceptual speed and Object Naming (Memory in Reality Test). The medical workup and lifestyle data from the last recorded examinations included total fasting cholesterol levels, history of smoking (never, former or current), current alcohol consumption (yes/no), body mass index, systolic and diastolic blood pressure and APOE-genotyping.

The general protocol for clinical assessment in the Swedish Twin studies is described in Gatz et al (44). It broadly followed the CERAD proceedings (45), and included physical, cognitive and neurological examination. Consensus clinical diagnosis of dementia followed the DSMIV criteria (46); differential diagnosis for AD was based on the NINCDS/ADRDA criteria for Alzheimer's disease (47).

Neuropathology

Bielschowsky silver preparations were performed on 6-μm-thick sections from 16 brain regions and the presence of neurofibrillary tangles was assessed and staged according to Braak stages 0-VI (48). Amyloid plaques were assessed semiquantitatively according to CERAD protocol for all brains. Immunohistochemistry for amyloid typing was only available for a few brains. The Braak staging data were split into 2 categories: Braak stage 0-II vs. III-VI. Braak 0 indicates little or no tauopathy. Statistical power was deemed too low to analyze the brackets III-IV and V-VI separately. All neuropathologic assessments of AD were carried out according to the CERAD criteria (49) and the old NIA guidelines (50). This study was approved by the Ethics Committee of Karolinska Institutet (registration number: 2007/151-31/4).

Statistical Analyses

Missing data rarely occur at random and a complete case analysis (deletion of all participants with one or more missing values) leads to loss of statistical power and to biased results. Therefore multiple imputation (10 data sets) was used to address the missing values (51), using the package MI in the statistical program R (version 3.20.0). Data were analyzed by pooling the 10 imputed data sets.

Univariate Analysis

All 22 predictor variables were tested individually for association with categorized Braak staging ≥III by means of univariate regression. Those that were associated with the outcome were selected for subsequent multivariable analysis. Research on variable selection in prognostic modeling, as described by the Cochrane Prognostic Methods Group, has shown that small alpha levels are not recommended in small datasets as this leads to poor model performance in independent data. Therefore the selection cut-off criterion of the variables before the multiple regression analysis was set to p < 0.20 (52, 53).

Multivariate Analysis

To examine which combination of prognostic factors gave the best prediction of severity of Braak stage, multivariate logistic regression was conducted with backward selection, using the R package BootstepAIC. Goodness of fit of the models was examined by comparing the Akaike Information Criterion (AIC). Bootstrapping procedures were applied both to select variables and to determine coefficients. Bootstrapping was performed using the package Boot (Bootstrap Functions) originally developed by Angelo Canty for S (2013).

Finally, a risk score was developed based on the best model. Regression coefficients of the model were divided by the lowest coefficient and rounded off to the nearest decimal for the scoring algorithm. The sum of individual scores for the various predictors in the model corresponds to a risk of tauopathy above the transentorhinal level (defined as Braak stages > III). These latter analyses were performed in SPSS for Windows, version 22, released 2013 IBM Corp., Armonk, NY).

Testing of Prognostic Model

The discriminative ability of the risk score output was assessed by using receiver-operating characteristic curves, and data on observed vs. predicted cases were used to calculate model sensitivity, specificity and prediction values.

RESULTS

The sample for analysis included 128 donated brains from individuals born between 1900 and 1934, and 22 different predictors. Baseline descriptive statistics are reported in Table 1A and Table 1B.

Table 1A.

Descriptive Characteristics for Individuals with Braak 0-II Scores (n = 41)

Variable	Subcategory	Mean (SD)	Range	Frequency (%)
Lifestyle Variables
Sex	Male			18 (43.9)
	Female			23 (56.1)
More than basic schooling (7 years)				10 (24.4)
One or more APOE ε4 allele				11 (26.8)
Current alcohol use				10 (24.4)
Smoking status	Never			17 (41.5)
	Former			18 (43.9)
	Currently			6 (14.6)
Neuropathological Variables
Dementia diagnosisc				29 (70.7)
	Clinical ADc			12 (29.3)c
Post-mortem Neuropathologic AD				10 (24.4)
Health Variables
Age in years at last examination		82.7 (6.7)	66 – 95
Body Mass Index		24.6 (3.7)	17.6 – 35.2
Cholesterol		5.4 (1.1)	2.9 – 8.3
Systolic blood pressure		145.5 (20.5)	102 – 190
Diastolic blood pressure		80.7 (8.6)	61 – 100
Time between last assessment and death (yrs)		3.0 (1.7)	0.0 – 6.46
Diabetes diagnosis				16 (39.0)
Age of onset of diabetes (yrs)d		80.5 (7.3)d	66 – 87d
Cardiovascular Diagnosis (CVD)				31 (51.2)
Age of onset of CVD (yrs)		75.4 (9.8)	50 – 88
Cognitive Variables
CERADb immediate recall		8.5 (4.1)	0.0 – 19.0
CEARDb delayed recall		1.5 (1.9)	0.0 – 8.0
CERADb recognition of words not on list		8.4 (2.6)	0.0 – 12.6
CERADb recognition of words on list		6.7 (2.8)	0.0 – 10.6
Verbal Fluency		9.0 (5.4)	0.0 – 25.0
Object Naming (Memory in Reality Test)		8.9 (2.0)	0.0 – 13.7
Mini Mental State Examination		19.8 (6.8)	2.0 – 30.0
WAISa information		10.6 (4.4)	2.0 – 22.0
Symbol-Digit Test		12.8 (10.6)	0.0 – 38.0

^aWechsler Adult Intelligence Score – Information subtest score.

^bCERAD Word list.

^c12 clinical AD cases of 29 with a dementia diagnosis, thus 41% of clinical AD for those with dementia.

^dIncomplete data.

AD, Alzheimer disease; CERAD, Consortium to Establish a Registry for Alzheimer's Disease; yrs, years.

Table 1B.

Descriptive Characteristics for Individuals with Braak III-VI Scores (n = 87)

Variable	Subcategory	Mean (SD)	Range	Frequency (%)
Lifestyle Variables
Sex	Male			27 (31.0)
	Female			60 (69.0)
More than basic schooling (7 years)				12 (13.8)
One or more APOE ε4 allele				47 (54.0)
Current alcohol use				19 (21.8)
Smoking status	Never			42 (48.3)
	Former			36 (41.4)
	Currently			9 (10.3)
Neuropathological Variables
Dementia diagnosisc				80 (92)
	Clinical ADc			56 (64.0)c
Post-mortem Neuropathologic AD				84 (96)
Health Variables
Age in years at last examination		84.4 (6.7)	67.2 – 100
Body Mass Index		23.5 (3.6)	14.1 – 33.3
Cholesterol		6.1 (1.3)	3.0 – 8.9
Systolic blood pressure		148.2 (23.8)	92.0 – 200.0
Diastolic blood pressure		81.2 (12.3)	52.0 – 116.0
Time between last assessment and death (yrs)		3.2 (2.9)	0.0 – 15.0
Diabetes diagnosis				6 (6.9)
Age of onset of diabetes (yrs)d		75.5 (11.5)d	55 – 87d
Cardiovascular Diagnosis (CVD)				44 (50.6)
Age of onset of CVD		79.0 (8.7)	57 – 98
Cognitive Variables
CERADb immediate recall		5.9 (4.9)	0.0 – 18.0
CEARDb delayed recall		0.9 (1.1)	0.0 – 5.0
CERADb recognition of words not on list		5.4 (3.3)	0.0 – 12.0
CERADb recognition of words on list		4.9 (3.0)	0.0 – 10.7
Verbal Fluency		7.1 (4.1)	0.0 – 19.0
Object Naming (Memory in Reality Test)		8.0 (2.4)	0.0 – 11.7
Mini Mental State Examination		17.9 (9.0)	0.0 – 43.6
WAISa information		7.2 (3.8)	0.0 – 22.0
Symbol-Digit Test		10.1 (9.8)	0.0 – 40.0

^aWechsler Adult Intelligence Score – Information subtest score.

^bCERAD Word list.

^c64 clinical AD cases of 80 with a dementia diagnosis, thus 70% of clinical AD for those with dementia.

^dIncomplete data.

AD, Alzheimer disease; CERAD, Consortium to Establish a Registry for Alzheimer's Disease; yrs, years.

Clinical and Neuropathological Assessment

Within the sample, 109 out of 128 individuals had a clinical diagnosis of dementia; 80 of those fulfilled criteria for clinical AD. Neuropathologic assessment showed that 93 of the 128 brains demonstrated signs of AD change using the diagnostic nomenclature of either “possible, probable and definite AD” (Table 2; Fig. 1). Of the 93 brains with AD changes, 36 had a second neuropathologic diagnosis (Fig. 2).

Table 2.

Distribution of Neuropathologic Diagnoses (n = 128)

Neuropathology Diagnosis	Subcategory	Frequency (%)	Average Braak score
Definite ADa/ High		29 (22.6%)	5.3
Probable ADa / Intermediate		35 (27.3%)	4.2
Possible ADa /low		29 (22.6%)	3.2
Not AD or otherb		35 (27.3%)	1.1
	Infarcts	6	-
	Vascular dementia/disease or similar	5	-
	Lewy body dementia	4	-
	Frontotemporal lobe degeneration	4	-
	Argyrophilic grain disease	3	-
	Parkinson disease	2	-
	Hippocampal sclerosis	2	-
	Asymptomatic (no pathology)	2	-
	Otherb	7	-

^aWith or without concomitant diagnosis, see diagram 2C.

^bOther diagnoses included: Progressive supranuclear palsy (PSP), Dementia Not Otherwise Specified, amyloid plaques, meningioma/metastases, mixed diagnoses but not AD, unclassifiable/damaged.

AD, Alzheimer disease.

Figure 1.

Distribution of neuropathologic Alzheimer disease (AD) diagnosis vs. Braak stages (n = 128).

Figure 2.

Distribution of concomitant diagnoses among the brains with Alzheimer disease (AD) changes (n = 93 out of 128).

Prognostic Model Estimation

The results from the univariate logistic regression with the categorized outcome variable, Braak stages 0-II vs. III-VI, are presented in Table 3. In total, 16 variables were significantly associated with the outcome at a significance level of p < 0.20. Table 4 shows the model obtained after variable selection by bootstrapping. The initial model, including all selected variables, had an AIC of 117 whereas the final model after eliminating 5 variables had an AIC of 108; a lower AIC indicates higher goodness of fit.

Table 3.

Univariate Prognostic Factor Associations with Braak Stages ≥III (n = 128)

Predictors	B	SEM	T	P value
Cardiovascular Disease	−1.977	0.436	−4.535	<0.001
Diabetes diagnosis	−3.152	0.776	−4.065	<0.001
CERADb recognition of words not on list	−0.253	0.068	−3.693	<0.001
One or more APOE ε4 alleles	1.401	0.403	3.478	<0.001
WAISa information	−0.17	0.051	−3.328	<0.001
Cholesterol level mmol/L	0.444	0.161	2.760	0.006
CERADb immediate recall	−0.121	0.047	−2.548	0.01
Dementia diagnosis	1.246	0.517	2.408	0.02
CERADb recognition of words on list	−0.154	0.064	−2.395	0.02
Body Mass Index	−0.112	0.053	−2.131	0.03
Object Naming (Memory in Reality Test)	−0.187	0.095	−1.966	0.05
Verbal Fluency	−0.072	0.041	−1.757	0.08
Mini Mental Status Exam Score	−0.034	0.023	−1.483	0.14
CERADb delayed recall	−0.17	0.126	−1.344	0.18
Symbol-Digit Test	−0.025	0.018	−1.295	0.2
Age in years at last examination	0.035	0.027	1.269	0.2c
Sex	0.452	0.379	1.192	0.23
7 years of schooling (yes/no)	−0.399	0.474	−0.844	0.4
Systolic blood pressure	−0.005	0.008	−0.671	0.5
Diastolic blood pressure	−0.009	0.016	−0.598	0.6
Time between last study and death	−0.019	0.071	−0.276	0.8
Smoking status	−0.036	0.269	−0.135	0.9

^aWechsler Adult Intelligence Score – Information subtest score.

^bWord list.

^cThe line above shows the cut-off for p < 0.2.

CERAD, Consortium to Establish a Registry for Alzheimer's Disease.

Table 4.

Multivariate Model with Predictors of Categorized Braak Stage After Bootstrapping and Risk Score Points

Predictors	B (original)	Bootstrapped B	95% CI	Risk score pointsc
Constant	−10.59	−12.93	−19.3 – 0.93	−88
One or more APOE ε4 alleles	1.211	1.35	−0.40 – 2.65	9,2
Total cholesterol mmol/L	0.857	0.955	0.10 – 1.54	6,5 × value
Diabetes diagnosis	−4.59	−5.47	−6.77 – −2.17	−37
Cardiovascular Disease	−1.70	−2.08	−2.84 – −0.03	−14
Age at last examination	0.116	0.146	−0.033 – 0.202	Age in years
WAIS informationa	−0.148	−0.173	−0.440 – 0.170	−1,2 × test score
Verbal Fluency	−0.216	−0.244	−0.554 – 0.150	−1,67 × test score
CERADb immediate recall	0.212	0.257	−0.145 – 0.456	1,76 × test score
CERADb recognition of words not on list	−0.141	−0.181	−0.325 – 0.174	−1,24 × test score

^aWechsler Adult Intelligence Score – Information subtest score.

^bCERAD Word list.

^cThe sum of risk score points corresponds to the probability of developing a Braak stage of or above III and is determined by the following equation: $P(Braak\ge III)=1∕[1+{\mathrm{e}}^{(-\text{intercept}+\sum (\beta \_1\times \text{Predictor}\_1+\dots \beta \_\mathrm{n}\times \text{Predictor}\_\mathrm{n}))}]$

CERAD, Consortium to Establish a Registry for Alzheimer's Disease; CI, confidence interval.

The risk score algorithm for the prognostic model is given in Table 4 and corresponding risk probabilities are presented in Table 5. A high total score corresponds to an elevated risk of Braak stage III-VI. In the model, APOE ε4 genotype, higher total cholesterol, older age and better CERAD immediate recall score (correct answers) were risk factors for Braak ≥III (positive βs), whereas higher scores on WAIS Information, on verbal fluency, as well as on CERAD recognition of words not on list (correct answers), presence of diabetes, and presence of cardiovascular disease were protective factors (negative βs). In the univariate analysis, the CERAD immediate recall score predicts Braak staging in the expected direction (negative β), but this switched in the multivariate analysis (positive β), presumably reflecting covariation among cognitive measures.

Table 5.

Risk-Score Interpretation

Total score	Risk of Braak ≥ III	Sensitivity	Specificity
<−25	0%	92.8 %a	75.4%a
−25 < score < −10	0 – 30%	98.8%b	31.7%b
−10 < score < 5	30 – 50%	94.2%b	56.0%b
5 < score < 15	50 – 80%	80.4%b	82.9%b
15 < score	80 – 100%	51.7%b	97.5%b

^aNote: For cut-offs below −25 points, the sensitivity and specificity refer to the case of zero risk rather than risk of Braak ≥ III.

^bNote: calculated using the lowest score point in the bracket as cut-off and with no upper limit.

Model Evaluation

The receiver-operating characteristic curve for the final model, displaying the overall discriminability between a Braak stage score of 0-II vs. III-VI, is shown in Figure 3. The area under the curve of the model was 0.897 (95% CI: 0.842-0.951), which indicates high-level discriminability. At the cut-off level of 50% probability, the risk score has a sensitivity of 85% and a specificity of 70%. For the same cut-off the negative predictive value is 69%, the positive predictive value 86% and the accuracy is 80%. Table 5 shows sensitivity and specificity for the other risk groups based on the various cut-off risk scores.

Figure 3.

Receiver-operating characteristic curves (ROC) for the prediction model. The area under the ROC curve indicates the level of discrimination between patients with predicted Braak stage 0-II and predicted Braak stage of III-VI, that is, the discrimination between tauopathy below or above the transentorhinal stage. Area under the curve (AUC) = 0.897; 95%, confidence interval (CI): 0.842-0.951.

DISCUSSION

This study shows that it is possible to predict large-scale tauopathy distribution from genotype, health outcomes and neuropsychological variables for older individuals, most of whom have a clinical dementia diagnosis. The model predicts Braak Stages III-VI well, that is, the abundant presence of NFT in hippocampus, amygdala and in the neocortex as opposed to only in the entorhinal cortex, i.e. subclinical Braak stages 0-II. A risk-score algorithm was created that discriminates between these 2 groups.

The widely used Mini-Mental State Exam score, education status and blood pressure measured the last time that the patient was seen are poor predictors of high Braak staging. APOE ε4 has been significantly associated with AD and it has also been shown that APOE status influences regional pathology in AD (54-56). The present model corroborates these results, showing that APOE ε4 was a significant predictor of widespread tauopathy. One imaging study showed that AD tauopathy was not predicted by APOE genotype but those results were not verified neuropathologically and they concerned cognitively normal aging (14).

Both diabetes and CVD were significantly protective in the univariate and multivariate analyses. Diabetes is a risk factor for increased general mortality (57), cardiovascular disease (58), and dementia (59-61) but pathological studies reveal inconsistent results regarding the risk for AD-type pathology: some results show no enhanced risk of AD-type pathology (62), whereas others demonstrate smaller hippocampal volumes (63, 64), or pancreatic amyloid and tau deposition (65). CVD is also a risk factor for dementia, more so for midlife CVD than later (66-68), and this is partly related to diabetes (58) and to obesity or high body mass index (10, 11). One possible reason that diabetes and CVD were protective predictors for widespread tauopathy may be the unexpected proportion of participants with pathologies other than tauopathy (e.g. vascular, infarcts or Lewy body dementia) and, therefore, associated with lower Braak stages. Diabetes and CVD exert their risk through vascular lesions and not through tauopathy; therefore, having these diseases would predict pathology other than NFTs, leading to the present result. It should also be noted that those with uncontrolled diabetes, hypertension, or CVD are at greater risk of dementia (and vascular dementia in particular) compared to diagnosed but well-controlled patients (60). Diabetic and cardiovascular patients in Sweden receive annual or biannual medical screening including adequate prophylaxis, making uncontrolled disease unlikely. Recent studies also suggest that anti-diabetic treatments, both oral and insulin, can contribute to slower cognitive decline in AD patients or even potentially lower AD incidence (69, 70). Similar studies on treatments for vascular risk factors suggest similar effects (71). Furthermore, survival bias (i.e. the survival of the healthiest participants despite illness) could explain why the individuals who survive after diabetes or CVD also have from lower Braak stages. There is no significant difference in age of death between the participants with or without these diagnoses. As age of onset of CVD and diabetes influences whether these variables are considered protective or not, the participants were stratified by age of onset of CVD or diabetes in a post-hoc analysis, but no significant difference in Braak stage could be defined in this sample.

The dementia prediction model of Kivipelto et al incorporates sex, education, systolic blood pressure, body-mass index and low physical activity, but shares the common factors of age, APOE ε4 and total cholesterol with the present model. These discrepancies may indicate the divergence between predicting dementia diagnosis and predicting tauopathy distribution. In Kivipelto et al, the area under the curve for the proposed models ranged between 0.7 and 0.8 (25), whereas the discriminability of the present model is high (area under the curve ~0.9), perhaps reflecting the advantage of predicting neuropathology as opposed to dementia diagnosis.

The evidence supporting the link between cognitive function and Braak staging is ample, usually involving sufferers from dementia and variability in global and specific cognitive function correlates highly with the Braak stage (25, 28, 32-35, 37). NFT spread, in particular, is associated with decline in multiple cognitive abilities even among older persons who do not qualify for a dementia diagnosis (72). The meta-analysis of Sonnen et al concludes that Braak staging currently is the best known neuropathological correlate of cognitive decline and of dementia (33). Moreover, the false-positive rate of concurring high-level NFT and intact antemortem cognition, namely individuals without a dementia diagnosis but with high Braak stage, is very low. In a review of 555 brains from individuals without dementia, only 15 (2.7%) had Braak stage V-VI (32); therefore, Braak staging is a reliable indicator of antemortem cognition (32, 36). In sum, clinical severity of cognitive decline and widespread presence of tauopathy are unlikely to diverge in this type of sample although the latter may diverge from clinical diagnosis of dementia or AD as concluded in the most recent diagnostic guidelines (4).

A Braak stage-based risk score is an improvement compared to other studies because it is independent of constantly updating clinical diagnostic criteria and permits better outcome resolution than with categorical clinical diagnoses. As it is speculated that high Braak stages are more directly related to AD-pathology than other outcomes (4, 6), predicting high Braak stage may be more disease-specific to AD than more general methods, such as radiologically verified structural changes in the brain. Predicting tauopathy distribution makes the prognostic model more related to specific brain pathophysiology than to effects that may be secondary to the disease or unrelated to the disease such as vascular lesions. Of the participants in the study, 32% have a Braak stage of 0-II, that is, no tauopathy above the transentorhinal stage, which is less characteristic of AD than Braak III-VI. This type of tauopathy distribution is therefore more likely a result of some other non-AD mechanism, perhaps more in line with the hypotheses of the suspected non-Alzheimer pathophysiology or primary age-related tauopathy concepts or other dementias (3, 5).

A strength of the present study is that it draws upon a wealth of antemortem data across several domains, in particular neuropsychological data (9 of the 22 variables), which are not always available in studies of neuropathology (18, 22, 26-31). The sample size of the cohort in this study was rather small and hence it suffers somewhat from diminished statistical power. Nevertheless, access to donated brains is rare, and in comparison to neuropathological studies the present sample size (n = 128) can be considered large, making the study worthwhile, particularly in combination with its rich antemortem data. Of the 10 studies reviewed in the meta-analysis by Sonnen et al, only the Hisayama Study has a significantly greater autopsy sample (n = ~800), but it is not a prognostic study (27, 33). Another limitation of the present study, despite its extensive data compilation, is some incompleteness of data. However, as argued, in a setting with complete data, a selection bias would occur because only the healthiest individuals are capable of participating in every new wave of data collection. Through the process of multiple data imputation, this effect is decreased (51).

The data were analyzed cross-sectionally and the sample consists only of elderly individuals, most, but not all of whom had a clinical diagnosis of dementia. The generalizability of the risk score is therefore confined to older individuals in the general population with newly discovered cognitive impairments or dementia diagnoses, speculatively within 1.5 standard deviations from the mean age of 84 years, i.e. 74 to 94 years (85% of the sample). The average time between the last examination and death was 3.1 years, and the t values (1.26) and p values (0.205) for age in the univariate analyses clearly reflect the finding that age in that range is not a strong predictor of tauopathy distribution.

Future studies of tauopathy and antemortem prediction of Braak stage could benefit from larger samples of brains for greater statistical power. Specifically, a truer prognostic design could be achieved with a larger clinical sample, including subjects with a full range of healthy aging, cognitive impairment and several subtypes of dementia, such as AD, vascular dementia, Lewy body dementia and frontotemporal dementia. By stratifying according to clinical status, the prognostic value of a predicted Braak stage would provide additional insight.