Predicting Risk of Cognitive Decline in Very Old Adults Using Three Models: The Framingham Stroke Risk Profile; the Cardiovascular Risk Factors, Aging, and Dementia Model; and Oxi-Inflammatory Biomarkers

Stephanie L. Harrison, PhD; Anton J. M. de Craen, PhD; Ngaire Kerse, PhD; Ruth Teh, PhD; Antoneta Granic, PhD; Karen Davies, PhD; Keith A. Wesnes, PhD; Wendy P. J. den Elzen, PhD; Jacobijn Gussekloo, PhD; Thomas B. L. Kirkwood, PhD; Louise Robinson, MD; Carol Jagger, PhD; Mario Siervo, PhD; Blossom C. M. Stephan, PhD


J Am Geriatr Soc. 2017;65(2):381-389. 

In This Article


All of the data sets used in this study were from longitudinal population-based studies of health and aging in very old adults.

Newcastle 85+ Study

All adults born in 1921 who were permanently registered with a participating general practice in Newcastle upon Tyne and North Tyneside (northeast England) were invited to participate.[19,20] A trained research nurse administered a multidimensional health assessment in the participant's usual place of residence. Of 1,453 eligible people invited to participate, 845 had baseline data for the detailed multidimensional health assessment and general practice record review. Those with a history of stroke or dementia were excluded from all analyses, in line with previous studies examining associations between the Framingham Stroke Risk Profile and cognitive function.[21–23] Six hundred sixteen participants had complete clinical and laboratory data and no previous history of stroke or dementia at baseline. Follow-up assessments took place 18, 36, and 60 months from baseline.[24]

Leiden 85-plus Study

Of the 705 residents of Leiden who turned 85 between September 1, 1997, and September 1, 1999, and were eligible to participate, 599 responded and participated in the study.[25] At baseline, 444 participants had complete clinical and laboratory data and no previous history of stroke or dementia and were included in this analysis. Participants were visited at their usual place of residence for a detailed health assessment, and their medical records from their primary care physician were also reviewed. There were five annual follow-up assessments.


The cohort was derived from two separate populations comprising Māori (indigenous people in New Zealand) and non-Māori.[26] The current study included only the non-Māori cohort aged 85 years old at baseline to allow comparison with the other included cohorts. Individuals born between January 1 and December 31, 1925, who resided within the Lakes or Bay of Plenty District Health Board areas when study enrollment was completed in 2010 were recruited. Of the 870 eligible non-Māori individuals, 516 enrolled in the study.[27] At baseline, 396 who had complete clinical and laboratory data and no previous history of stroke or dementia were included in this analysis. Participants were given the choice to meet at their usual place of residence or at another site to complete a structured face-to-face standardized questionnaire and a detailed health assessment; general practice medical records were reviewed. This analysis included data from three annual follow-up assessments.

Assessment of the Risk Prediction Models

The FSRP and the CAIDE models were determined in each study using baseline data. Table S1 presents the variables included in each model and their measurements. All variables, with the exception of education and physical activity, were measured similarly across the cohorts. Information on apolipoprotein E4 was not available for the LiLACS NZ data set. For analysis, FSRP and CAIDE scores were divided into study-specific tertiles to create low-, intermediate-, and high-risk groups, with the low-risk group used as the reference category.[11,28]

Biomarkers and Oxi-inflammatory Load

Three biomarkers previously shown to be associated with CVD were selected from the blood results, including two inflammatory biomarkers (CRP, IL-6) and one biomarker for oxidative stress (homocysteine).[4] Details of the biomarker assays in each study can be found in the Appendix S1. The distributions of the biomarkers were skewed, so the data were log transformed. A fixed value of 0.1 was added to each data point, including 0 values, to calculate the logarithmic value. To determine the combined effect of all three biomarkers, a cumulative score of the standardized z-scores for each log-transformed biomarker value was created and is referred to as a participant's oxi-inflammatory load. Homocysteine was not available for the LiLACS NZ data set, so the cumulative score could not be calculated in this cohort.

As in previous analyses, biomarkers and oxi-inflammatory load scores were grouped into deciles, and participants were allocated to one of three groups (<10th percentile, 10th–90th percentile, >90th percentile), with the middle category being used as the reference.[29] A sensitivity analysis grouping oxi-inflammatory load scores into three tertiles (with the middle category used as the reference) was also conducted.

To examine whether oxi-inflammatory load scores could improve prediction of cognitive impairment using the FSRP or the CAIDE model, analyses were repeated after adding points to the CAIDE and FSRP scores based on oxi-inflammatory load values of participants; specifically, six points were added for the highest tertile of oxi-inflammatory load, three points for the middle tertile, and no points for the lowest tertile.

Cognitive Assessment

Global cognitive function was assessed in all three cohorts using the standard Mini-Mental State Examination (MMSE; range 0–30).[30] The MMSE was conducted at baseline, 36 months, and 60 months in the Newcastle 85+ Study; at baseline and annually for 5 years in the Leiden 85-plus Study; and at baseline and annually for 3 years in the LiLACS NZ Study. Cognitive impairment at baseline and incident cognitive impairment at each follow-up was defined as a score of 25 points or less.

Domain-specific cognitive functions including attention, information processing, and episodic verbal recognition memory were assessed in the Newcastle 85+ Study using the Cognitive Drug Research (CDR) System. Details of the cognitive assessments in the Newcastle 85+ Study have been published elsewhere.[31] The CDR System was conducted at baseline, 18, and 36 months.

In the Leiden 85-plus Study, speed of information processing was measured using the Letter Digit Coding Test. Attention was measured using the Stroop Test Part 3. Memory was measured using the 12-word learning test. These tests were administered at baseline and annually for 5 years. There were no domain-specific tests administered in the LiLACS NZ Study.

Statistical Analysis

Cox proportional hazards models were run to determine whether FSRP and CAIDE scores and oxi-inflammatory load were associated with impairments in global or domain-specific cognitive function (dichotomized variables) over the follow-up period in each study. Tests of the proportional hazards assumption were run for each model and were not violated. Biomarker models were adjusted for potential confounding factors including sex, years of education, current alcohol consumption, and smoking status.

Following this, a Meta-analysis of the highest HR category was conducted to estimate the pooled effects of the FSRP, CAIDE, and oxi-inflammatory load scores on prospective risk of impaired global cognitive function (MMSE scores). HRs and 95% confidence intervals (CIs) for each study were entered into the models. Statistical heterogeneity was assessed using the I[2] and the Q tests, and P < .10 was chosen as a cut-off for heterogeneity.[32] A fixed-effect model was applied because of the lack of heterogeneity (Q test, P > .10), and forest plots were generated for the FSRP, CAIDE, and oxi-inflammatory load scores. Comprehensive Meta-Analysis 2 software (Biostat, Engelwood, NJ) was used to conduct the analysis.

Linear mixed models were used to examine change in continuous test scores for each cognitive measure. Cognitive test scores that were positively skewed (power of attention, simple reaction time, Stroop Test Part 3) were logarithmically transformed, and MMSE scores (which were negatively skewed) were corrected using the formula where K is the maximum score, and x is the participant's score. Each model included the risk model or biomarker score (cross-sectional effect), time (change in cognitive scores over time), and an interaction term between the risk model or biomarker score and time (additional effect of the risk model or biomarker score). All data were analyzed using Stata version 13.0 (Stata Corp., College Station, TX).