Physical fitness and dementia risk in the very old: a study of the Lothian Birth Cohort 1921

Participants were members of the Lothian Birth Cohort 1921 (LBC1921), which is described more fully elsewhere [21]. Most had taken part in general intelligence testing at age 11, as part of the Scottish Mental Survey 1932 (SMS1932) [22]. Childhood cognitive test scores were available for 89.6% of those LBC1921 participants who attended baseline testing. Missing childhood cognitive data was likely to have resulted from school absence on the day of testing (due to sickness, for example), or be because the test results for some schools in Fife (a region neighbouring Lothian) were not found. From 1999, SMS1932 participants were traced and recruited for follow-up in later life, with the aim of investigating normal cognitive ageing [23]. Five-hundred and fifty participants, mostly from the Lothian area of Scotland, enrolled in the study and attended the first wave of testing. Participants were aged approximately 79 years at baseline. Surviving participants who continued to take part in the study were re-tested at four subsequent test waves; at ~ 83, 87, 90 and 92 years of age (approximately 3-yearly intervals). This interval was planned as the period over which the authors expected to be able to detect significant changes in some cognitive test scores, while balancing this against attrition. At each wave, data were collected by questionnaire and in-person testing. Participant deaths were ascertained prospectively, at regular intervals, with details supplied by the General Registrar’s Office, Scotland. The Lothian Research Ethics Committee (test waves 1–3) and the Scotland A Research Ethics Committee (test waves 4–5) provided ethical approval for the study. From wave 4, attending participants provided consent for data linkage and access to health records.

The three fitness measures included in this study were 6-m walk time, grip strength and forced expiratory volume in 1 s (FEV₁) [1]. Walking speed, grip strength and FEV₁ have all been shown to be valuable in evaluating physical and functional capacity [24‐27]. Six-metre walk time was taken as the time in seconds for a participant to walk a measured length of 6 m, at a normal walking pace. Subjects were permitted to utilise any habitually used walking aid, while those who were unable to walk six metres would not participate in the test. A Jamar Hydraulic Hand Dynamometer was used to measure grip strength in kilograms, in the dominant hand; the best of three trials was recorded. FEV₁ was measured using a microspirometer, in units of litres per second; the best of three attempts was recorded. Taking three measurements is in line with guidance for clinical practice [28].

The analyses would also include variables deemed to be of potential importance with regard to their association with either physical fitness or dementia. In a previous study of the risk factors for dementia in LBC1921, positive APOE ɛ4 carrier status was found to increase risk, whereas a history of hypertension was found to reduce risk [29]. These variables were therefore included, along with the following important control variables: age, sex, height, age 11 IQ, smoking status, history of cardiovascular or cerebrovascular disease and history of diabetes. Age, sex and height are all key control variables given the direct impact on physical strength and fitness. Similarly, a history of cardiovascular or cerebrovascular disease and diabetes history were included given the potential association with physical fitness and dementia. When such conditions impact a person’s ability to partake in exercise, fitness is likely to be adversely affected. The inclusion of age 11 IQ as a control variable is important given the known association with dementia [30] and other fitness-related factors [1, 31]. We also considered the potential impact of smoking status given the likely association with physical fitness, and lung function in particular. Whilst other factors were potentially associated with fitness, we could not include all possible risk factors, as this would have led to multiple hypotheses testing. We therefore adopted the hypothesis-driven approach described above, based on previous findings.

APOE ɛ4 status was determined using genomic DNA isolated from participants’ venous blood. Any history of hypertension, diabetes, cerebrovascular or cardiovascular disease, smoking status (ex-, current-, or never-) and sex were self-reported by participants at the first wave of testing. A positive history of cerebrovascular or cardiovascular disease included those reporting previous stroke, transient ischaemic attack, myocardial infarction, angina, coronary artery bypass graft and angioplasty procedures and those who reported a positive history but did not specify the specific nature. A positive history of diabetes included those reporting any history of diabetes (any type or type unspecified). Age at baseline was taken from the number of days between date of birth and date of attendance at wave 1 testing. Age 11 IQ was calculated based on the results of the Moray House Test (MHT) no.12, the general intelligence test taken by participants at age 11 as part of SMS1932 [21]. MHT scores were corrected for age at testing, then converted to IQ-type scores with a standardised mean score of 100 (SD 15). Height in centimetres was measured using a SECA stadiometer.

Follow-up for the purpose of dementia ascertainment has been described previously and involved the retrospective collection of evidence, from enrolment to age 95 years [29]. Data were collected from death certificates, electronic hospital records, and clinical reviews. Each death record available by 30th June 2016 was examined for evidence of dementia or cognitive impairment. Electronic hospital records and psychiatric records were accessed for participants who were willing and able to consent to data linkage. Any evidence for dementia or cognitive impairment since enrolment in the study – whether a confirmed diagnosis or evidence for a diagnosis – was recorded. Data were collected from electronic medical records to the 16th May 2016. Additional information was available for 26 participants who underwent clinical review by the authors (TCR, JMS), either in the NHS or research setting. Information from such reviews was collected up to 15th December 2016, when all of the evidence from each source was considered at a final dementia diagnosis consensus meeting (RAS, TCR, JMS). The evidence was examined against a previously described list of criteria for ‘probable’ or ‘possible’ dementia diagnosis [29]. Using these criteria, the meeting agreed upon the presence of a diagnosis and the subtype. Any disagreement was resolved through discussion.

The events considered in these analyses would be dementia and death. Determination of these outcomes was as described above. Time to death was taken as the number of days between the date of attendance at wave 1 testing and date of death. For those who did not die, the censoring time was taken as the number of days between wave 1 testing and a date beyond that last date of data collection for any participant. Time to dementia was taken as the number of days between the date of attendance at wave 1 testing and the first date that a dementia diagnosis was noted in any of the available sources. If dementia was recorded on a death certificate only, and no duration was noted, the diagnosis was presumed to predate death by 6 months. If the duration was not given on the death certificate, but a diagnosis was recorded in another source, the earliest such date was used to determine the time to dementia. If sources had recorded cognitive decline, mild cognitive impairment and dementia, the date was taken as the earliest recording of a dementia diagnosis. If dementia was ascertained based on evidence within the records that did not include a formal diagnosis of dementia, the earliest mention of cognitive impairment was used to date dementia (as long as this did not specifically note the absence of a dementia syndrome). For those who did not develop dementia, the time to dementia variable was recorded as either the time to date of death or to a date beyond that last date of data collection for any surviving participant.

Any participant reporting a history of dementia or scoring 23 or less on the Folstein Mini Mental State Examination [32] (MMSE) at baseline was excluded from our analyses. Those without a valid MMSE score at baseline were also excluded, as were those without any follow-up data available. To minimise the potential for classification error, possible dementia cases were excluded from the analyses and probable dementia was used as the primary outcome in this study.

We first performed simple comparison analyses between the group who developed dementia and the group who did not, for each of the included variables. Univariate analyses were completed using either a Pearson chi-square (for categorical variables) or independent samples t-test (for continuous variables) (IBM SPSS, Version 21). A p value of < 0.05 was used to demonstrate a statistically significant difference. All subsequent steps in the analyses were completed using the R statistical software (version 3.3.3) [33].

The second stage of the analyses used binary logistic regression to examine the potential risks for dementia associated with the fitness measures. In a first model (logistic regression model 1) we included the fitness variables along with those factors known to be associated with dementia in our cohort and the other important control variables (FEV₁, grip strength, 6-m walk time, APOE ɛ4 carrier status, height, age, sex, history of hypertension, smoking status and age 11 IQ). The development of probable dementia was the outcome. In the second model (logistic regression model 2), the same variables were included, with the addition of a history of cardiovascular or cerebrovascular disease and a history of diabetes. A history of hypertension was considered separately to these other health variables as a statistically significant association with dementia had previously been observed within this cohort.

The main analyses used Cox regression models with death being included as a competing risk for dementia; in doing so, the influence of fitness on earlier death in the analysis of dementia development due to fitness was considered. The models were completed using standard software for Cox regression with right-censored data and treating death as censored, (R package survival). The first and second Cox regression models (Cox regression models 1 and 2) included the same covariates as those described for logistic regression models 1 and 2. Both models fitted the data acceptably with concordance of 64%. Analysis of the scaled Schoenfeld residuals (R function cox.zph), showed all covariates complied with the proportional hazards assumption. As is recommended practice, we supported the results of Cox regression with graphs demonstrating cumulative incidence for each competing event; illustrating the time-varying risk of dementia, between covariate levels. The unbiased estimate of cumulative incidence was calculated using the Aalen-Johansen estimator [34], (R packages prodlim and mstate).

A total of N = 550 participants recruited to LBC1921 attended baseline testing at age 79 years. We excluded the following participants from the analyses: participants with an MMSE score of less than 24 at baseline (n = 9), participants without a valid MMSE score at baseline (n = 2), participants reporting a history of dementia at baseline (n = 2), and participants with no follow-up data available for the purpose of dementia ascertainment (n = 41). One additional participant was excluded as the calculated time to dementia suggested that dementia predated attendance at wave 1 testing. From the remaining participants (n = 495), a consensus diagnosis of probable dementia was agreed for n = 109. Those participants with a possible diagnosis of dementia were excluded from the analyses (n = 7).

The resulting eligible cohort (n = 488) included 280 females (57.4%) and 419 participants (85.9%) were known to be deceased [29]. This included 331 who died with no diagnosis of dementia. Descriptive statistics for those eligible for inclusion (both with and without dementia), and those excluded are shown in Table 1.

The mean time to death for the n = 419 participants who were known to be deceased was 3144.5 days (SD: 1517.5). For those who survived, the time to death was taken as an arbitrary point beyond the last date of data collection for any participant; 6500 days. The mean time to dementia for the n = 109 who developed dementia was 3535.7 days (SD: 1283.3). A total of n = 379 participants remained dementia free; for these participants, time to dementia was taken either as the time to death for those who died (n = 331, mean = 2863.0 days, SD: 1469.4), or an arbitrary point beyond the last date of data collection for any participant (n = 48, 6500 days).

Also included in Table 1 are the results of the group comparison analyses. Univariate analyses demonstrated little difference between those participants who developed dementia and those who did not. Only smoking status (p < 0.001) and APOE ɛ4 (p < 0.001) carrier status demonstrated statistically significant difference (p < 0.05).

Univariate analyses were also performed to compare those who died with those who survived. Lower FEV₁ and greater 6-m walk time were both associated with an increased risk of death (p = 0.02 and p = 0.01, respectively). The results for these univariate analyses are available in Additional file 1: Table S1.

The results of both logistic regression analyses demonstrated that APOE ɛ4 remains an important risk factor for dementia after age 79 years (logistic regression model 2: Odds Ratio (OR) 2.52 (95% confidence interval: 1.50, 4.22), p < 0.001). Both initial models also suggested that increased FEV₁ at age 79 increased the risk for subsequent dementia (logistic regression model 2: OR 1.93 (1.07, 3.57), p = 0.03). The only other variable that reached statistical significance was height, and only in the second model; increased height was shown to decrease risk for subsequent dementia (OR 0.95 (0.91, 1.00), p = 0.04).

In both Cox regression models, APOE ɛ4 continued to be an important predictor for dementia (Cox model 2: Hazard Ratio (HR) 2.85 (95% confidence interval: 1.85, 4.41), p < 0.001). The results did however demonstrate that once death was considered within the analyses, the association between FEV₁ and dementia did not reach statistical significance in either model (Cox model 2: HR 1.30 (0.74, 2.30), p = 0.37). Neither grip strength (Cox model 2: HR 0.98 (0.94, 1.02), p = 0.35) nor walking speed (Cox model 2: HR 0.99 (0.87, 1.13), p = 0.90) was associated with dementia in either model. No other variable was demonstrated to be associated with dementia.

For the purposes of calculating cumulative incidence, FEV₁ results were divided into two groups, at a value of 1.8 l per second. This value was close to both the sample mean and median. The stacked cumulative incidence plot shown in Fig. 1 demonstrated that increased FEV₁ was associated with decreased risk of death but not dementia (Fig. 1). A second cumulative incidence plot confirmed that the presence of an APOE ɛ4 allele was associated with an increased risk for dementia (Fig. 2).

A significant strength of this study lies in the suitability of the cohort for these analyses. The minimal cultural, ethnic and geographical variability within the narrow-age cohort minimises confounding errors and means that the results are specific to the oldest old. We recognise, though, that this limits the generalisability of the results, and so additional studies are required in other groups. The availability of a childhood IQ score is a rare strength of the study, as is the detailed follow-up completed since enrolment in the study. The high mean baseline MMSE score (28.3, SD: 1.5) for those included in the analyses increases the likelihood that we identified incident cases of dementia, rather than prevalent cases. By including multiple measures of fitness it was possible to consider several components of fitness. Whilst this was a strength of our study, it could be argued that additional tests could have provided a more comprehensive assessment of fitness. We selected the additional variables for inclusion in the analyses based on the available evidence, but it is possible that there was some residual confounding due to a variable not considered in this study. It is not however possible to exclude all such possibilities without negatively affecting the analyses through multiple hypotheses testing. The dementia ascertainment method in this study has previously been shown to be effective and comparable with expected rates [29]. Having determined dementia cases retrospectively using existing data, it was not possible to accurately date the onset of dementia. For this reason, we cannot be entirely confident in the accuracy of the time to dementia variable. The method used to determine date of onset in this study, is however conservative and it is likely that dementia onset preceded our assigned date by some time. As such, the true competing risk of death is almost certainly less than we are adjusting for in the model. While we recognise this limitation of our study, it is not unique to our method; it is notoriously difficult to pinpoint the exact date of dementia onset, even in studies using prospective clinical follow-up. This is particularly true given that the most common form of dementia – Alzheimer’s disease – has a gradual onset. A further limitation of our study is the possibility of missed cases. Our results could also have been affected by using an outcome of probable dementia of any subtype. The number of participants with each subtype of dementia was too few to perform individual analyses. It is also possible that the overall size of our study cohort had an impact on our findings, and that a larger sample size would have produced differing results.