This retrospective cohort study was based on datasets from two separate research projects. These were the WPHC (19) and the Dementia Prevalence Survey (11).

The WPHC was a community based screening program of 3,033 Aboriginal and Torres Strait Islander individuals aged 13 years and over living in 26 rural and remote communities in northern Queensland, Australia, between March 1998 and December 2000 (19). Non-Indigenous residents and children aged ≤ 12 years were excluded from the study. An additional WPHC was undertaken between 2005 and 2007 in a subset of communities, which involved the follow up of some participants and recruitment of new participants (20, 21).

The Dementia Prevalence Survey was a cross-sectional assessment of the prevalence of dementia among 276 Torres Strait Islander and Aboriginal peoples aged 45 years and over living in all the populated island and mainland communities in the Torres Strait and Northern Peninsula Area of Far North Queensland between 2015 and 2018 (11). Participants were screened with an adapted version of the Kimberley Indigenous Cognitive Assessment (KICA) tool and received a clinical assessment for dementia or CIND, as part of a comprehensive geriatric assessment.

The WPHC and the Dementia Prevalence Survey overlapped in seven communities in the Torres Strait and five communities of the Northern Peninsula Area. In the years between the studies, some participants had moved to a community that was included in the Dementia Prevalence Survey and not the original WPHC. As a result, the current study comprised participants from eight communities in the Torres Strait and five communities of the Northern Peninsula Area. Ethics approval to link the datasets from these two studies was granted on 07/12/2018 by the Far North Queensland Human Research Ethics Committee (HREC/18/QCH92-1262).

A sample size calculation using the “Kelsey formula” for an independent cohort study (22) indicated 273 participants would be required [i.e., α = 0.05, β = 0.80, Prevalence not Exposed (P0) = 0.28, Prevalence Exposed (P1) = 0.44, Relative Risk (RR) = 1.6, Ratio of exposed to unexposed (r) = 1.00]. The parameters for this formula were based on a smaller pilot linkage of the data where albuminuria showed a 60% increase in later life CIND/Dementia risk (i.e., RR = 1.6).

Probabilistic data linkage was undertaken over four stages by the primary author (FT) in Stata 15 (College Station, TX: StataCorp LLC) using the package “dtalink.” The weights and thresholds that defined “Exact Matches” and matches requiring review are described in Supplementary Figure 1. The Stata “calcweights” command determined the weighting for identifiers at each stage. At the end of each stage, successfully linked records were removed from each dataset, so the number of unlinked records in both datasets was reduced for the next stage. After four rounds of probabilistic linkage, a final review of both datasets was also undertaken. A final linked dataset of identifiers from both data sources was prepared and provided to a co-author for review (ST). A Unique Linkage Key was generated and used to combine the full WPHC and Dementia Prevalence Survey datasets into a final linked study dataset for analyses (Figure 1).

A detailed description of the WPHC methods has been published elsewhere (19). The WPHC dataset contained baseline information on waist circumference (centimeters, cm), weight (kilograms, kg), Body Mass Index (BMI), current alcohol consumption (Yes/No), and cigarette smoking (i.e., self-reported “currently smoked tobacco,” Yes/No). Food intake was measured via a 24-h recall method, which included the number of serves of fruit and vegetables. Physical activity was measured using a 7-day recall method. Adequate levels were ≥5 days in the previous 7 days where a participant undertook moderate physical activity for ≥20 min. Hypertension (HBP) was defined as systolic ≥140 mmHg and/or diastolic ≥90 mmHg, and/or current anti-hypertensive medication. Fasting venous blood samples were used to collect measures for triglycerides, total cholesterol, high density lipoprotein (HDL), low density lipoprotein (LDL), gamma-glutamyl transferase (GGT), and fasting glucose. Very Low-Density Lipoprotein (VLDL, mmol/l) was derived post-hoc from triglycerides. Diabetes was identified through self-report (confirmed through medical record check) or a baseline fasting glucose >7.8 mmol/l. Hyperglycemia (HBG) was defined as fasting glucose ≥5.5 mmol/l. First catch urine samples were self-collected and analyzed for albumin creatinine ratio (UACR). Albuminuria, which is an excess of the protein albumin in the urine and an early indicator of kidney damage and a surrogate marker systematic microvascular dysfunction (23), was defined as UACR ≥2.5 mg/mmol for males and ≥3.5 for females (24).

There was variation in screening attendance across the two WPHC time points. For example, some participants had first WPHC data only, others had second WPHC only, and some had data from both health checks (see Figure 1). There was also variation in the completeness of the baseline measures during each screening (see Supplementary Table 1). To ensure maximum completeness of baseline information and the greatest number of participants, this study used the “earliest” and “latest” WPHC measures available for each participant. The “earliest” available measure was data from the first WPHC screen, supplemented with data from the second screen for participants who had not attended the first one. The “latest” available measure was the reverse, that is, data from the second WPHC screen for participants who attended, supplemented with information from the first screen for participants who had not attended the second one. The main two analyses for this study were the relationship between 1) each participant's earliest WPHC measures and later dementia assessment and 2) their latest WPHC measures and later dementia assessment. Although self-reported highest level of education was captured at follow up, this variable related to early life and was considered as predictive. Education was analyzed as a categorical variable (i.e., primary school level only, some or all secondary school, or post school) and as a dichotomous variable in modeling (i.e., primary vs. secondary/post school).

Continuous outliers for the earliest and latest WPHC measures, by CIND/Dementia status, were identified using Quartiles 1 and 3 (i.e., Q1 and Q3) and the Interquartile Range (i.e., IQR). Values that were lower than Q1-(IQR^*1.5) or greater than Q3+(IQR^*1.5) were considered outliers and winsorized by replacing these values with the next lowest or highest non-outlying value (25, 26). Most continuous variables had five or fewer outliers, with the exception of GGT, UACR, and glucose (see Supplementary Table 2).

The main outcome measure was a diagnosis of “normal,” “CIND” or dementia from the Dementia Prevalence Survey. This variable represented a consensus diagnosis, made by a panel of Geriatricians and an Older Person's Psychiatrist, who blind reviewed results from the comprehensive geriatrician assessments to obtain consensus diagnoses based on criteria from the Diagnostic and Statistical Manual for Mental Disorders, 4th Edition (DSM IV-TR) (18). Participants were classified as normal cognition, dementia, or CIND. The latter group comprised people who met DSM IV-TR criteria of Cognitive Disorder-Not Otherwise Specified (NOS) or Amnestic Disorder-NOS, such as cognitive decline without significance impact of activities of daily living. The full details of this method have been published elsewhere (11). While dementia and CIND are different, both states are characterized by an objective decrease in cognitive functioning relative to age expectations (18), and share similar risk factors (3).

All analyses were undertaken using the Stata 15 software package. Categorical variables were tested against the outcome measure using Pearson chi-square tests for independence or Fisher's Exact tests for expected cell counts <5. Cramer's V was used for effect sizes for categorical analyses (Table 1; Supplementary Table 3). Continuous variables were assessed for normality and examined using mean and Standard Deviation (i.e., SD or ±) when normally distributed or otherwise as median (Interquartile Range—IQR), with appropriate tests of significance (e.g., Independent samples t-test or Kruskal-Wallis rank sum tests). Effect sizes were calculated using Cohen's D and Rosenthal's Z score conversion, respectively (Table 1; Supplementary Table 3). Univariate generalized linear modeling was undertaken to create risk ratios (RRs), with CIND/Dementia as a dichotomized outcome and WPHC measures as predictor variables. Age was significantly associated with many WPHC risk factors and was also the most prominent risk factor for CIND/Dementia. To address this confounding, all modeling analyses were also adjusted for age (i.e., aRR) at time of the dementia assessment (Table 2; Figures 2, 3; Supplementary Table 4). P-values <0.05 were considered statistically significant.

Multivariable modeling was limited by the small sample size. To accommodate this limitation, bootstrapping with the Stata command “swboot” was used to inform variable selection (Table 3). First, variables that were at or below p = 0.200 in univariate generalized linear modeling were selected into a single multivariate model (Model 1). Non-significant variables or those highly correlated with other variables (i.e., Pearson's r ≥ 0.3) were removed, and the remaining variables entered into the “swboot” command with 200 and 100 repetitions for earliest and latest risk factors, respectively. Variables that were present in 25% of the bootstrap models were entered into Model 2, and the bootstrap method was repeated. Non-significant variables, that were highly correlated and had low presentations in the bootstrapping (i.e., <50%) were then removed, and the remaining variables were entered into the final model (Model 3) (Table 3). Multicollinearity was assessed with the STATA command “coldiag2” and presented as a Condition Index, with values 10–30 indicate moderate collinearity (27). Although age lost significance during modeling, this was due to collinearity with other variables and once these variables were removed, age remained significant. Analyses were repeated with subsets of the data limited to participants aged <60 years at the time of baseline assessment, to examine the potential effect of older participants at baseline.

Figure 1 shows that 65 participants in WPHC and Dementia Prevalence Study were linked over four rounds of probabilistic linkage based on multiple combinations of unique identifiers. The weights assigned to each unique identifier and the thresholds used to determine a successful linkage are described in Supplementary Figure 1. One participant was missing dementia diagnosis status at follow up and was excluded from further analyses. The final linked dataset comprised 64 individuals, of which 9.6% (n = 6) were diagnosed with dementia between 2015 and 2018, 23.4% (n = 15) with CIND, and 67.2% (n = 43) were cognitively normal. Due to the small sample size, participants with dementia or CIND were grouped together for the main analyses (n = 21, 32.8%). The participants were living in eight communities in the Torres Strait and five in the Northern Peninsula Area. While most were from geographical locations with larger populations, their overall distribution was not representative of census population estimates for the entire region (data not tabled).

Figure 1 shows the majority (59%) of linked participants took part in both WPHC assessments, although this was lower for participants who later had CIND/Dementia (42.9%) compared with those without (67.4%) (p = 0.019). Figure 1 also provides the average time in years to follow up dementia assessment by participation in a WPHC screen.

The mean age of the sample was 48.9 years (SD = 10.8, range = 25–68) at baseline and was mostly female (70.3%). Table 1 shows the distribution of risk factors at the earliest WPHC assessment, by later CIND/Dementia status. Participants with CIND/Dementia were ~10 years older at the earliest WPHC assessment (mean = 53.9 ± 10.3) and dementia screen (71.0 ± 9.9) compared with cognitively normal participants (46.6, p = 0.010 and 63.3, p = 0.005, respectively). Participants with CIND/Dementia were more likely to only have a primary (50.0%) school education compared with those who were cognitively normal, although significance was only at trend level (25.6%, p = 0.086). There were high rates of hypertension, diabetes, albuminuria, hyperglycemia, and combinations of these risk factors in both groups. While these factors were generally more frequent among those who later developed CIND/Dementia, the effect sizes for differences were modest. Mean body weight (83.9 kg, range = 60.3–106.7) and median VLDL (0.7, range = 0.3–1.1) were significantly lower in the CIND/Dementia group compared with the cognitively normal group (92.9 kg, range = 66.7–128.2, p = 0.040 and 0.9, range = 0.3–1.8, p = 0.044, respectively). Mean systolic blood pressure (145.2 ± 27.2) in the CIND/Dementia group was significantly higher compared with the cognitively normal group (mean = 131.3 ± 16.4, p = 0.034).

Generalized linear modeling (Table 2; Figures 2A,B) showed age at the WPHC was positively associated with dementia risk (aRR 1.05, 95% CI 1.01–1.10, p = 0.011). While secondary or post-school education appeared to be protective against later life CIND/Dementia compared with primary school only, this effect was attenuated after adjusting for age (aRR = 0.58, 95% CI 0.27–1.23, p = 0.155). After adjusting for age, waist circumference and triglycerides were lower among people who later developed CIND/Dementia (aRR = 0.98, 95% CI 0.95–1.00, p = 0.075 and aRR = 0.63, 95% CI 0.41–0.97, p = 0.034, respectively), while UACR was higher (aRR 1.05, 95% CI 1.01–1.09, p = 0.006).

Variables significant at the p ≤ 0.200 level during univariate analyses were age and highest education dichotomized, and the earliest measures for body weight, waist circumference, systolic and diastolic blood pressure, HDLC, VLDL, triglycerides, HbA1c, and UACR (see Table 2). These were entered into a single multivariate analysis (i.e., Model 1) for bootstrapping assisted variable selection (Table 3). The variables diastolic blood pressure, HbA1c, VLDL, and weight were removed from further modeling due to low significance and/or collinearity with other variables. Bootstrapping removed HDLC (i.e., variable selection <50). The remaining variables were analyzed in a multivariate model (Model 2) and systolic blood pressure and UACR were then removed due to non-significance, low selection in bootstrapping, and high correlation with age. In the final model (Model 3), which had moderate multicollinearity (i.e., Collinearity Index = 28.02) increasing age was significantly associated with CIND/Dementia risk, more education was protective, and waist circumference and triglycerides were again significantly lower among people who later developed CIND/Dementia. These variables were also selected frequently in bootstrapping (i.e., >50% of models).

At the time of the latest WPHC measures, the mean age of participants was 52.9 (SD = 11.4, range = 25–75). Adequate levels of physical activity were protective for later CIND/Dementia risk (aRR 0.41, 95% CI 0.20–0.85, p = 0.016) (Supplementary Tables 3, 4; Figure 3). Bodyweight and VLDL were negatively associated with CIND/Dementia (p = 0.042 and p = 0.019 respectively). After adjusting for age, albuminuria was the only vascular factor with positive risk ratios at both the latest (Figure 3B) and earliest (Figure 2B) WPHC assessments. This trend also held when albuminuria was comorbid with hypertension. However, these ratios were modest and non-significant, reflecting the small cell sizes and wide confidence intervals. Multivariate modeling of latest WPHC measures (Table 3), with bootstrapping and removal of variables for collinearity, produced a final model where age was significantly positively associated with CIND/Dementia risk, while more education, body weight, and triglycerides were significantly negatively associated with CIND/Dementia risk. In this model, sufficient physical activity measured at the latest WPHC was protective for later CIND/Dementia. However, this was inconsistent with the trend from the earliest WPHC data (see Figures 2B, 3B).

The earliest measures of waist circumference and body weight were examined in post-hoc analyses. Waist circumference was not significantly associated with other non-anthropometric baseline study variables. Weight was negatively correlated with age (Pearson's R = −0.285, p = 0.023). For comparison, all the main analyses with the earliest data were repeated on 51 participants aged <60 years at the baseline measures. The main findings for age, education, weight, and triglycerides were replicated (Supplementary Table 5).

The small sample size was below the number required based on sample size calculations. This reduced the generalizability of our findings, which should be interpreted with caution. As there were a small number of participants with CIND or dementia, we were required to combine these two distinct syndromes into a single dichotomous outcome measure, which also reduced generalizability. The small sample size also limited modeling reliability, resulted in moderate collinearity (i.e., Condition Index 10–30) and likely contributed to inconsistent associations between the earliest and latest health check measures and later CIND/dementia status. Further, the baseline data did not include assessments of cognition and included some older participants, who may have had prodromal dementia 10–20 years before follow-up assessment. However, our comparative analyses limited to participants aged <60 years at baseline produced the same main findings, suggesting that CIND/dementia at baseline was unlikely. The population is also culturally unique, which should be considered when generalizing findings to other First Nations populations, including in Australia. The baseline WHPC measures were also limited to “snapshots” of an individual's health, and no long-term information about medication prescribing, treatment, or change over time was available. Despite these limitations, this modest study is the first to describe long-term risk factors in a population with a high prevalence of CIND/Dementia and is the first to leverage data linkage of disparate studies to reduce participant burden. While the baseline measures were not comprehensive, they are the same measures that are collected routinely in primary care in Australia, so have real world applicability. The results may also have some global relevance to other populations that also experience high lifetime exposure to potentially modifiable dementia risks.