Study population
We conducted a case-cohort study based on the ESTHER study (Epidemiologische Studie zu Chancen der Verhütung, Früherkennung und optimierten Therapie chronischer Erkrankungen in der älteren Bevölkerung [German]). In this prospective cohort study implemented in Saarland, Germany, 9940 women and men aged 50 to 75 years at baseline were recruited during a general health checkup by their general practitioners between 2000 and 2002. Besides an age of 50–75 years, the inclusion criteria for the ESTHER study were physical and mental ability to participate in the study as well as knowledge of the German language. Participants were followed up concerning the incidence of major diseases and mortality 2, 5, 8, 11, 14, and 17 years after baseline. For details, see Löw et al. [17]. The sociodemographic baseline characteristics and common prevalent chronic diseases were similarly distributed in the respective age categories as in the German National Health Survey, a representative sample of the German population [17].
Dementia ascertainment
Information about a dementia diagnosis was collected during the 14- and 17-year follow-ups of the ESTHER study. The median follow-up time was 16.3 years (interquartile range: 13.5–17.0 years), and the maximum was 19.4 years due to the 2-year period of baseline recruitment. In brief, the collection of dementia diagnoses included sending standardized questionnaires to the study participants’ general practitioners (GPs). Participants who had dropped out during previous follow-ups due to ill health or had died were included in the dementia ascertainment through the GPs as well. If the GPs were aware of a dementia diagnosis for their patients, they were asked to provide a date of diagnosis and all available medical records documenting a dementia diagnosis. The latter included records from neurologists, psychiatrists, memory clinics, or other specialized providers. If the GP provided a mixed dementia diagnosis, available medical records were screened for an underlying Alzheimer’s disease or vascular dementia background and considered as Alzheimer’s disease, vascular dementia, or both. The Alzheimer’s disease diagnosis guideline used in Germany during the follow-up period of the ESTHER study was the one of the National Institute on Aging and the Alzheimer’s Association [18].
Biomarker measurements
Inflammation-related, blood-based proteins were measured from serum samples collected during the health checkup at baseline (2000–2002). Blood samples were sent to the study centre and stored at −80°C until biomarker measurements took place in three waves in March 2018, December 2018, and September 2020 (referred to as time points t1, t2, and t3 in the following). At the time of the measurements, 10–25 μl of serum was extracted from different aliquots that had been thawed twice and sent with dry ice to the laboratories, which analysed the samples with the Olink Target 96 Inflammation panel, Olink Proteomics, Uppsala, Sweden. At t1 and t2, samples were analysed in the laboratory of Olink Proteomics, Uppsala Science Park, SE-75183 Uppsala, Sweden. At t3, the measurements were performed in the Research Unit Protein Science, German Research Center for Environmental Health, Helmholtz Center Munich, Heidemannstraße 1, 80939 München, Germany.
The Olink panels are based on a proximity extension assay technology (PEA) [19, 20]. Details on the reliability and stability of the technology are described elsewhere [21]. In brief, oligonucleotide-labelled antibody probe pairs are allowed to bind to their respective target proteins in the samples. Only if two antibodies are in close proximity, a polymerase chain reaction (PCR) reporter sequence is formed by DNA polymerization. This sequence is detected and quantified using high-throughput real-time quantitative PCR (qPCR) (Fluidigm® BiomarkTM HD system). The Olink Target 96 Inflammation panel allows the measurement of 92 biomarkers per sample. A list of all biomarkers of this panel is displayed in Supplemental Table 1.
At t1, t2, and t3, 22, 15, and 5 plates were used, respectively. To avoid batch effects, cases and controls were randomly distributed across plates and adjusted according to included interpolate controls. The average intra-assay coefficient of variance among all 92 measured biomarkers was 7%, 4%, and 3% at t1, t2, and t3, respectively. The average inter-assay coefficient of variance was 12%, 10%, and 10% at t1, t2, and t3, respectively. Furthermore, the quality of each serum sample was assessed by Olink technology [22]. All samples were measured successfully, and the number of quality control warnings was below 4% in all three timepoints. Of the 1435 randomly selected controls and 393 incident dementia cases, 46 serum samples of participants were excluded due to a quality control warning by Olink.
Protein levels are reported as Normalized Protein eXpression (NPX) values, a relative quantification unit logarithmically related to protein concentration. The number of samples with values below the lower limit of detection (LOD) varied strongly by biomarker and is shown in Supplemental Table 1. In total, 20 biomarkers with > 25% of the values below LOD were excluded from all analyses (grey-shaded biomarkers in Supplemental Table 1). Thereby, 72 out of the 92 biomarkers were considered evaluable markers. Biomarker values below the LOD were replaced by LOD/√2. The normalization of raw data was conducted with the R (R Core Team, 2020, version 3.6.3) package “OlinkAnalyze”, developed and maintained by the Olink Proteomics Data Science Team [23]. To normalize data from three different measurement points (t1, t2, and t3), reference sample normalization was used based on 17 and 16 bridging samples between each two measurement points (t1–t3 and t2–t3, pair-wise bridging). For details of the procedure, we refer to the white paper of OLINK Proteomics [22].
Covariate assessment
Data on sex, age, education, body mass index (BMI), physical activity, and life-time history of depression were collected during baseline assessment through a standardized self-administered questionnaire. The history of coronary heart disease (CHD) and diabetes mellitus were obtained from physician diagnoses. Furthermore, anti-diabetic drugs reported by the GP were used to complement diabetes mellitus diagnoses. Participants were considered to have cardiovascular disease (CVD) based upon CHD diagnoses from GPs or self-reported history of myocardial infarction, stroke, pulmonary embolism, or revascularization of coronary arteries. TaqMan single-nucleotide polymorphism (SNP) genotyping assays were used to determine apolipoprotein E (APOE) genotypes. More precisely, genotypes were analysed in an endpoint allelic discrimination using a PRISM 7000 Sequence detection system (Applied Biosystems) [24].
Inclusion and exclusion criteria
The selection of study participants from the ESTHER cohort for this case-cohort analysis is shown in Fig. 1. ESTHER participants were eligible for selection as cases or random controls. Participants were excluded if they withdrew consent to contact the GP (n=248), if the GP withdrew consent to be contacted (n=304), or if the GP could not be reached anymore (n=1035). Thereupon, dementia diagnosis information was requested for 8353 participants and received for 6940 participants (response rate: 83.1%). Moreover, participants were excluded if dementia incidence status could not be ascertained by GP questionnaires (either diagnosis was not available (n=412) or not confirmed (n=171)) or if blood samples were not available (n=73). Thus, information from 6284 participants was available for analyses. Except for their age, participants with available dementia information and blood samples and those who could not be used for analyses had comparable baseline characteristics (Supplemental Table 2). Fewer included than excluded participants were in the oldest age group of 70 to 75 years (13.8% compared to 19.0%) reflecting the challenges to obtain dementia information about deceased study participants from their GPs.
Olink inflammation panel measurements were performed in a case-cohort design among 1435 randomly selected study participants and all incident dementia cases of the rest of the cohort (n=393). To check if the random selection was successful, we compared the baseline characteristics of selected and non-selected controls and observed no substantial differences between the groups (Supplemental Table 3). After excluding participants with quality control warning, 389 incident dementia cases and 1393 randomly selected participants were available. As the randomly selected participants included 115 incident dementia cases, the study population comprised 504 participants with incident dementia and 1278 randomly selected controls.
Statistical analyses
First, to describe factors associated with dementia risk, categorized baseline characteristics of all-cause dementia cases and controls were compared using the χ2-test. Second, odds ratios (ORs) for all-cause dementia were estimated with a multivariate logistic regression model, including all baseline characteristics shown in Table 1.
In a univariate, descriptive analysis, the median and interquartile range (IQR) of all inflammation-related protein levels of all-cause dementia, Alzheimer’s disease, and vascular dementia cases were separately compared with those of controls, using the Wilcoxon rank sum test. Additionally, in a multivariate approach, the ORs per one standard deviation (SD) increase of each inflammation-related protein were assessed separately with each outcome (all-cause dementia, Alzheimer’s disease, and vascular dementia incidence) in logistic regression models adjusted for potential confounders. In models for Alzheimer’s disease incidence, study participants with other (e.g. vascular dementia) or unknown dementia forms were excluded. The same was applied for the outcome vascular dementia incidence by excluding Alzheimer’s disease and other non-vascular dementia cases. The models were adjusted for age, sex, education, physical activity, BMI, CVD, diabetes, depression, and APOE genotype. All variables were used as categorical variables, as described in Table 1, except age, which was modelled continuously. The covariates were selected because they were statistically significantly associated with all-cause dementia, Alzheimer’s disease, or vascular dementia in a previous analysis with the ESTHER study participants [25]. Statistical test results were corrected for multiple testing by the Benjamini and Hochberg method for all tests carried out for one outcome [26]. A false discovery rate (FDR) < 0.05 was applied as the threshold for statistical significance. In a sensitivity analysis, the multivariate logistic regression model was repeated using weighted Cox regression, according to Barlow et al. [27].
We further aimed to identify those inflammation-related proteins whose association with a dementia outcome was independent of other inflammatory biomarkers. Therefore, all biomarkers, which were significantly associated with a dementia endpoint after FDR correction, were tested for the independence of the association by forward elimination. Only biomarkers having the strongest, independent, positive association with the outcome entered the regression model with the threshold for statistical significance of p<0.05 in the following logistic regression analysis. Moreover, biomarker clusters were built and named based on the identified independent biomarkers. All other biomarkers of the Olink inflammation panel, which were highly correlated (Spearman’s correlation coefficient r > 0.5) [28] with an independent biomarker, were put in its cluster. One biomarker might be in more than one cluster. We favoured this statistical approach over a principal component analysis because it has a higher transparency and is easier to reproduce by others, its results are easier to interpret, and the associations of the biomarkers with the dementia outcomes are being acknowledged in the decision about the number of clusters.
The associations of the independent biomarkers with dementia endpoints were further analysed in subgroup analyses based on age, sex, obesity, diabetes, history of CVD, and APOE ε4 polymorphism. These factors were selected a priori because they are important dementia risk factors and determinants of inflammation [29]. Apart from this, interaction terms were tested. In addition, the dose-response relationships between the independent biomarkers and dementia endpoints were assessed with restricted cubic spline curves [30].
Several sensitivity analyses were performed. To check for potential reverse causality, the associations between the independent biomarkers and dementia endpoints were analysed stratified by time of diagnosis (in the first 10 years of follow-up vs later years). Competing risk of death was examined by excluding subjects without dementia diagnosis who died before their 80th birthday, the average life expectancy of the cohort’s population. Fractional polynomials with first-order terms were used to determine each biomarker’s best fitting function with the outcomes [31]. The linear function was the best fitting one for almost all biomarkers (67 of 72). Since the low number of biomarkers with deviations in the best fitting function from linearity (6.9%) could have resulted from multiple testing, all were modelled linearly. In a sensitivity analysis, the multivariate logistic regression analysis was repeated with the five non-linear biomarkers modelled with their best fitting function. Finally, to examine the impact of persons with a potential acute infection on the overall results, subjects with C-reactive protein (CRP) levels > 20 mg/L were excluded.
To our knowledge, missing values of covariates were missing at random. The highest proportion of missing values was found for APOE polymorphism (7.5%). Thus, multiple imputation was used to impute missing values. Variables shown in Table 1 were used for the imputation model. Twenty data sets were imputed with the Markov Chain Monte Carlo (MCMC) method separately by sex with the SAS procedure PROC MI. All analyses were performed based on those 20 datasets with the SAS procedure PROC MIANALYZE.
Statistical tests were two-sided using an alpha level of 0.05. All statistical analyses were conducted with the Statistical Analysis System (SAS, version 9.4, Cary, North Carolina, USA).