Validating the role of the Australian National University Alzheimer’s Disease Risk Index (ANU-ADRI) and a genetic risk score in progression to cognitive impairment in a population-based cohort of older adults followed for 12 years

Participants were community-dwelling adults residing in the City of Canberra, Australia, or in the neighbouring town of Queanbeyan who had been recruited into the Personality and Total Health (PATH) Through Life project, a longitudinal, population-based study of health and well-being in adults. Cohorts aged 20–24 (20+), 40–44 (40+) and 60–64 (60+) years at baseline were assessed at 4-year intervals for a total of 12 years. The background and procedures for the PATH study are described elsewhere [16]. Written informed consent was obtained from all participants. This study was approved by the human research ethics committee of The Australian National University.

In this study, we used data from the 60+ cohort with interviews conducted in 2001–2002 (n = 2551), 2005–2006 (n = 2222), 2009–2010 (n = 1973) and 2014–2015 (n = 1645). Individuals were excluded if their ethnicity was not Caucasian (n = 107) or if they had a self-reported history of stroke, transient ischemic attack, epilepsy, brain tumours or brain infection (n = 381).

The development of the ANU-ADRI and the methodology underlying its computation have been described previously [15]. The ANU-ADRI can be computed on the basis of up to 15 predictive variables, 11 of which are available in PATH, including age (self-report), sex (self-report), alcohol consumption (calculated according to National Health and Medical Research Council 2001 guidelines [17] using number of drinks per week, with light to moderate intake in males being 0.25–20.5 drinks per week and in females being 0.25–13.5 drinks per week), education (self-reported number of years of education), diabetes (self-reported history of diabetes), depression (assessed using the Patient Health Questionnaire [PHQ-9] [18] following the coding algorithm provided in the PHQ-9 instruction manual, with a score >10 used as a cut-off), traumatic brain injury (self-reported history of traumatic brain injury with loss of consciousness), smoking (self-reported smoking status as current smoker, past smoker or never smoker), social engagement (constructed from four domains for marital status, size of social network, quality of social network, level of social activities; a fifth domain for living arrangements was not available in PATH and thus was computed pro rata as the average of the above-mentioned social engagement variables), physical activity (combined self-reported number of hours performing mild, moderate and vigorous activities, weighted by multiples of 1, 2 and 3, respectively [19]), cognitively stimulating activities (assessed as the number of cognitive activities undertaken in the last 6 months, comprising reading, writing, playing games or attending cultural events), and BMI (weight divided by height squared, expressed in kilograms per square meter). No data were available for the remaining three predictive variables: cholesterol, fish intake and pesticide exposure. The ANU-ADRI is still predictive of the development of dementia, even when a subset of variables is used [15]. Values for predictive variables included in the ANU-ADRI for PATH were selected from baseline measurements or the first occasion on which the variables were measured. To facilitate interpretation, a constant of +13 was added to the ANU-ADRI to change the range to from −13 to +19 to 0–32.

The most significant LOAD risk single-nucleotide polymorphisms (SNPs) identified via genome-wide association studies (GWASs) [11, 20‐24] from 23 loci (ABCA7, BIN1, CD2AP, CD33, CLU, CR1, EPHA1, MS4A4A, MS4A4E, MS4A6A, PICALM, HLA-DRB5, PTK2B, SORL1, SLC24A4-RIN3, DSG2, INPP5D, MEF2C, NME8, ZCWPW1, CELF1, FERMT2 and CASS4) were genotyped using TaqMan OpenArray assays (Life Technologies, Carlsbad, CA, USA) as previously described [25, 26], in addition to the two SNPs defining the APOE alleles, which were genotyped using TaqMan assays as previously described [27]. Using these LOAD risk SNPs, an explained variance-weighted genetic risk score (EV-GRS) [28] was constructed, which is the sum of all the risk alleles across the individual, weighted by minor allele frequency (MAF) and the OR associated with LOAD. The EV-GRS is calculated according to the following formula:for the ith patient, where \( log\left({OR}_i\right) \) = the OR for the jth SNP, \( {MAF}_{ij} \) = the MAF for the jth SNP, and \( {G}_{ij} \) = the number of risk alleles for jth SNP. Individuals with missing genetic data were excluded (n = 240). We weighted the LOAD SNPs using the previously reported OR for LOAD and by the MAF for the CEU reference population (Utah residents with Northern and Western European ancestry; see Additional file 1: Table S1). The EV-GRS was transformed into a z-score.

The screening and clinical assessment methods at waves 1–3 are described elsewhere [29, 30] and are briefly summarised here. At each wave, the same predetermined cut-off derived from a battery of cognitive tests was used for inclusion of participants in a sub-study on mild cognitive disorders and dementia. Participants from the full cohort were selected for clinical assessment if they had any of the following: (1) a Mini Mental State Examination (MMSE) [31] score <25; (2) a score below the fifth percentile score on immediate or delayed recall of the first list of the California Verbal Learning Test [32]; or (3) a score below the fifth percentile on two or more of the Symbol Digit Modalities Test (SDMT) [33], Purdue Pegboard with both hands [34] or Simple Reaction Time [35]. At wave 4, participants were selected for review if they met any of the following criteria: (1) MMSE score <25 or ≤2.5^th percentile on one or more cognitive test, (2) previous diagnosis at waves 1–3, (3) subjective decline ≥25 on the Memory and Cognition Questionnaire (MACQ) or (4) decline in MMSE score ≥3 points.

The criteria for the clinical assessment for cognitive impairment at waves 1–3 has been published by our group elsewhere [30]. They involved a structured clinical assessment for dementia conducted by one of two physicians, a neuropsychological assessment, and the Clinical Dementia Rating [36], which were used together to formulate a consensus diagnosis.

Owing to the large number of participants screened for review at wave 4, diagnosis was based on neurologist review of interview data as outlined below and in Fig. 1. For each of the 1644 participants with interview data at wave 4, assessment data were screened for signs of decline on the basis of the following criteria (screen 1): a previous diagnosis of a cognitive disorder at waves 1, 2 or 3 or either evidence of cognitive impairment on the MMSE (≤24) or performance on one or more cognitive tests ≤6.7th percentile at wave 4 (immediate recall task, delayed recall task, SDMT, F words, A words, Boston Naming Test, Simple Response Time task, Choice Response Time task, Purdue Pegboard dominant, Purdue Pegboard non-dominant, Purdue Pegboard both, Digit Span Backward, Trail Making Test B, Stroop words, Stroop colour-word test). Additionally, participants had to show evidence of either subjective decline (score ≥25 on the MACQ [32]) or evidence of decline (>3-point decline in MMSE score since wave 3) or evidence of consistent cognitive impairment over time (MMSE ≤24 at waves 3 and 4).

All data derived from the health survey and cognitive testing as well as informant interview were collated into a spreadsheet case file for each participant. This case file (screen 2) automatically screened each participant for meeting criteria for any one of the following diagnoses: Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), major neurocognitive disorder (NCD); DSM-IV dementia; DSM-5 mild NCD; MCI; age-associated cognitive decline; age-associated memory impairment; DSM-IV amnestic disorder not otherwise specified; DSM-IV mild NCD; and DSM-IV other cognitive disorder. Major criteria for meeting most of these diagnoses were operationalised as any of the following: (1) concern of self or informant of significant cognitive decline (MACQ ≥25 or Informant Questionnaire on Cognitive Decline in the Elderly >3.31 or history of dementia diagnosis); (2) substantial impairment on at least one cognitive domain relative to wave 4 normative data (cut-offs less than −2 SD for dementias, less than −1.5 SD for mild cognitive disorders); (3) interference with independence and instrumental activities of daily living (IADL; self-reported IADL impairment or Bayer IADL scale score >3.11 or informant-reported everyday cognitive difficulties); (4) not exclusively during delirium (cognitive changes of >6 months’ duration, onset of cognitive changes preceding informant report of onset of delirium-like symptoms); and (5) not due to another co-existing disorder (PHQ9 < 9 and no reported history of schizophrenia or other psychosis). Those meeting criteria for one or more diagnoses (n = 368) were screened for case file review by a research neurologist. Diagnoses were made for 301 of these cases, of which 60 complex cases were selected for diagnostic consensus based on the following criteria: (1) comorbid depression, (2) other comorbid psychiatric conditions, (3) stroke and (4) DSM-5 major NCD without memory impairment. Following consensus diagnosis with a clinician specialising in psychiatry, the final diagnoses included 85 dementia/major NCD, 196 mild cognitive disorders (MCI/mild NCD), and 34 other mild or medically related cognitive disorders.

Clinically diagnosed MCI was based on the Petersen criteria at waves 1 and 2 [37], whereas the Winblad criteria [38] were used at waves 3 and 4. Clinically diagnosed dementia was based on the DSM-IV criteria [39] at all waves. At wave 4, there were 14 participants who were not interviewed but were known to have dementia on the basis of informant reports and medical records. Owing to the small number of individuals classified with dementia, participants with either MCI or dementia were grouped into a single MCI/dementia category.

To complement the clinical diagnosis of MCI, a broader MCI-TB classification was applied to the entire PATH sample [40] at each wave on the basis of education-adjusted cognitive performance (Table 1). The PATH sample was first stratified by education (0–12 or 13+ years). Within each of these strata, individuals were classified as MCI-TB if they scored 1.5 SD below the mean on one or more of the psychometric tests used to assess the following cognitive domains: perceptual speed measured using the SDMT [33], episodic memory assessed using the immediate recall of the first trial of the California Verbal Learning Test (recall-immediate) [32], working memory measured using the Digit Span Backward from the Wechsler Memory Scale [41] and vocabulary assessed by the Spot-the-Word test [42].

All statistical analyses were performed using R version 3.1.2 software [43]. Because missing values can reduce power and introduce bias in the resulting estimates, missing values that were not attributable to attrition for the predictive variables used in the construction of the ANU-ADRI and the MCI-TB (see above) were imputed using an implementation of the random forests algorithm available in the ‘missForest’ package in R [44, 45]. This left 2078 individuals available for analysis. Additional file 2: Table S2 shows the proportion of missing variables for each variable.

We first evaluated the risk of progression from CN to MCI/dementia using Cox proportional hazards models with age as the time scale and the ANU-ADRI and EV-GRS included as predictor variables in the same model. The outcome of interest in these models was the time to first diagnosis of MCI/dementia, with those subjects who did not develop MCI/dementia at their last assessment right-censored. HRs and 95% CIs were given for the time to MCI/dementia analysis. Concordance index (c-index) for the prediction of conversion from NC to MCI/dementia was calculated. Cox proportional hazards models were estimated using the ‘survival’ package in R.

To evaluate a more complex model of disease progression, MSMs were used to examine the association between the ANU-ADRI and EV-GRS and transitions between cognitive states. MSMs allow the modelling of competing risks and backward transitions between states (i.e., recovery) [46]. Hidden Markov models can be used to estimate misclassification error, and the effects of covariates can be allowed to vary by transition [46]. The MSMs used in this analysis modelled cognitive deterioration and cognitive recovery by allowing transitions and backward transitions between CN, MCI or MCI-TB states. Backward transitions from dementia were not allowed, whereas death was used as a fourth absorbing state (Fig. 2). Individuals with only a single observation (i.e., no recorded transitions) were excluded from the analysis (n = 204). Individuals lost to attrition were considered right-censored. The ANU-ADRI and the EV-GRS were included as covariates in the same model. Maximum likelihood estimates of parameters in the MSMs were obtained with the Broyden-Fletcher-Goldfarb-Shanno optimisation method. Normalisation was applied to the likelihood function to improve numerical stability. Because the likelihood is maximised using numerical methods, an input of initial values is required to start the search for a maximum. MSMs were fitted using ‘msm’ [46] in R, and multiple models were run using different sets of initial values to ensure the robustness of the parametric estimates. See Additional file 3 for more details on the structure of MSMs.

As a sensitivity analysis for the MCI-TB analysis, more stringent criteria were investigated with MCI-TB based on a score of 1.5 SD below the mean on two or more of the above-mentioned psychometric tests. Additionally, we performed a complete case analysis to ensure that our imputation method was not biasing the observed results.

Baseline distributions of education, depression, sex, ANU-ADRI, raw cognitive tests scores and cognitive states at each wave for the PATH cohort are described in Table 1. Participants who completed all four waves of interviews had a higher level of education than participants who completed only the wave 1 interview (t = −6.8, df = 331.3, p < 0.001). Participants were followed for an average of 9.6 years (after accounting for loss due to attrition) and a total of 13.9 years. Group differences in the sub-indices of the ANU-ADRI between CN and either MCI/dementia or MCI-TB can be found in Additional file 4: Table S3. The distribution of the ANU-ADRI and EV-GRS scores is shown in Fig. 3. As expected, the proportion of individuals classified as MCI/dementia increased over the course of the study, whereas the proportion of individuals classified as MCI-TB remained stable (Table 1). By wave 4, 36% of the cohort had been lost to follow-up, with 57, 54 and 94 individuals deceased by waves 2, 3 and 4, respectively, and an additional 280, 267 and 329 individuals being lost to follow-up for other reasons (e.g., refusal, left catchment area) at waves 2, 3 and 4, respectively.

Between any two waves, a greater proportion of people transitioned from CN to MCI-TB (10.5%) than from unimpaired to MCI (2.6%), indicating that MCI-TB is a broader categorisation of cognitive impairment. A smaller proportion of individuals transitioned in the opposite direction—from MCI-TB to CN (31.3%)—than from MCI to CN (44%), indicating that MCI-TB is also a more stable category (Table 2).

A higher ANU-ADRI (indicating greater risk) score was associated with an increased risk of progression to both MCI/Dementia and MCI-TB (Table 3). The EV-GRS was not associated with progression to either MCI/Dementia or MCI-TB. The interaction between the ANU-ADRI and the EV-GRS was non-significant for the MCI/Dementia (HR 0.99 [95% CI 0.96–1.01], p = 0.33) and MCI-TB (HR 0.99 [95% CI 0.98–1.01], p = 0.11).

In the sensitivity analysis, using a more stringent MCI-TB criterion (scoring 1.5 SD below the mean on two or more tests) confirmed that the ANU-ADRI was associated an increased risk of progression from CN to MCI-TB (HR 1.08 [95% CI 1.05–1.10], p = < 0.0001). In the complete case analysis, the ANU-ADRI remained significant for both the MCI/dementia (HR 1.06 [95% CI 1.02–1.09], p = 0.001) and MCI-TB (HR 1.036 [95% CI 1.01–1.04], p = 0.007) models.

A higher ANU-ADRI score was associated with an increased risk of transitioning from CN to MCI or MCI-TB (Fig. 1, Table 4). The probability of transitioning from CN to either MCI or MCI-TB after 12 years for individuals scoring 1 SD below the mean on the ANU-ADRI was 10%, and for individuals scoring 1 SD above the mean, the probability of transitioning was 20%. A higher ANU-ADRI score was not associated with transitions from CN, MCI, dementia or MCI-TB to death; with transition from cognitive impairment to dementia; or with cognitive recovery from MCI or MCI-TB to CN. The EV-GRS was associated with an increased risk of transitioning from CN to dementia, with the probability of transitioning from CN to dementia for individuals scoring 1 SD above the mean being 1.3%. The interaction between the ANU-ADRI and the EV-GRS was not significant for any of the transitions for either the MCI or MCI-TB models.

In the sensitivity analysis, using a more stringent MCI-TB criterion (Additional file 5: Table S4) confirmed that the ANU-ADRI was associated an increased risk of progression from CN to MCI-TB (HR 1.12 [95% CI 1.07–1.17]). For the complete case analysis (Additional file 6: Table S5), the ANU-ADRI remained statistically significant for both the models for transition from CN to MCI (1.06 [1.02–1.09]) and from CN to MCI-TB (HR 1.05 [95% CI 1.01–1.08]).