Abstract
Evidence regarding the relationship between performance on specific cognitive domains and cause of death is scarce. We assessed whether specific cognitive domains predicted mortality and the presence of any association with specific causes of death in a population-dwelling sample of non-demented older adults. In this population-based, prospective study (NEDICES), 2,390 non-demented subjects ≥65 years completed a brief neuropsychological battery. Cox’s proportional hazards models, adjusted by sociodemographic and comorbidity factors, global cognitive performance, educational level, and premorbid intelligence were used to assess the risk of death. Participants were followed for a median of 9.2 years (range 0.01–10.7), after which the death certificates of those who died were examined. 880 (36.8%) of 2,390 participants died over a median follow-up of 5.5 years (range 0.01–10.5). Using adjusted Cox regression models, we found that hazard ratios for mortality in participants within the lowest tertiles (worse performance) were 1.31 (speed of cognitive processing, p = 0.03); 1.22 (semantic fluency, p = 0.04), 1.32 (delayed free recall, p = 0.003), and 1.23 (delayed logical memory, p = 0.03). Poor performance on delayed recall and speed of cognitive processing tests were associated with dementia and cerebrovascular disease mortality, respectively. Further, poor performance on semantic fluency was associated with decreased cancer mortality. In this study of community dwelling non-demented older adults, worse neuropsychological performance was associated with increased risk of mortality. Performance on specific cognitive domains were related to different causes of death. Of particular note there appears to be an inverse association between poor semantic fluency and cancer mortality.
Keywords
INTRODUCTION
Several population-based prospective studies have demonstrated a marked decrease in survival of people with both dementia and mild cognitive impairment [1, 2]. Specifically, dementia is the main cause of mortality in persons with mild cognitive impairment [2]. In this context, it is not surprising that much research is focused on how to delay cognitive decline in older people and therefore increase the survival of persons with cognitive impairment.
There is a large body of evidence which demonstrates that global cognitive impairment predicts all-causes mortality in older adults [3, 4]. In addition, several studies have concluded that worse performance on memory and other cognitive domains, such as perceptual speed, are associated with an increased mortality rate [5–7]. However, scientific evidence about the relationship between performance in specific cognitive domains and cause of death is scarce [8–15]. In this sense, it would be interesting to know whether worse performance on a specific cognitive domain (e.g., recall or language) may help identify particular groups of individuals who may be at risk of mortality from specific diseases. This strategy offers the opportunity to evaluate its prognostic value in terms of mortality and explore the underlying mechanisms that link cognition and mortality.
In this study, we evaluated the usefulness of several neuropsychological tests, namely speed of cognitive processing, naming, recall, and semantic verbal fluency to predict an increased risk of mortality, after controlling for different covariates in a population-based sample of older people without dementia. Moreover, we aimed to investigate if there were specific associations between the cognitive profile of individuals and cause of death.
MATERIAL AND METHODS
Study population
The data for this research was derived from the Neurological Diseases in Central Spain (NEDICES), a longitudinal, population-based survey of the prevalence, incidence, and determinants of major conditions of the elderly. These included dementia, cerebrovascular disease, Parkinson’s disease (PD), and essential tremor (ET) [1, 16–28]. Detailed accounts of the study population and sampling methods have been published [16–18]. The survey area consisted of three communities: (1) Margaritas (approximately 14,800 inhabitants), a working-class neighborhood in Getafe (Greater Madrid) (urban blue collar area); (2) Lista (approximately 150,000 inhabitants), a professional-class neighborhood in the Salamanca district (Central Madrid) (urban white collar area), and (3) Arévalo (Ávila) (approximately 9,000 inhabitants), an agricultural (rural) zone located 125 km northwest of Madrid. In each community, eligibility was restricted to residents who were aged 65 years or older and those who were present on December 31, 1993, or during six or more months of 1993. Eligible persons who had moved away from the survey area were not traced. In Margaritas and Arévalo, every eligible subject was selected for screening. However, because of the large number of elderly residents in Lista, proportionate stratified random sampling was used to select a sub-sample of 2,113 subjects for screening. The selected study population was 6,395 people, but 481 people were ineligible (census issues, address errors or death), leaving 5,914 eligible subjects, of whom 5,278 were enrolled.
All procedures were approved by the ethical standards committees on human experimentation at the University Hospitals “12 de Octubre” (Madrid) and “La Princesa” (Madrid). Written (signed) informed consent was obtained from all enrollees.
Study evaluation
Face-to-face evaluations were performed at baseline (1994 to 1995) and then at follow-up (1997 to 1998). Briefly, at the time of their baseline assessment (1994 to 1995) and follow-up assessment (1997–1998) older adults were interviewed face-to-face using a screening questionnaire to collect data on demographics, medications, current medical conditions, lifetime smoking (ever versus never), and lifetime drinking (ever/at least once per week versus never). The questionnaire included items for neurological disorders (cerebrovascular disease, dementia, PD, and ET), and the presence of subjective memory complaints and depressive symptoms. We also assessed the use of antidepressant medications, a marker which may be less prone to biases than a simple screening question. A short form of the questionnaire was mailed to subjects who refused or were unavailable for face-to-face or telephone screening. This form covered demographic characteristics, several neurologic disorders (dementia, cerebrovascular disease, PD, and ET), current medications, and the name of theirfamily doctor.
Participants who screened positive for any neurological disease were examined by one of eight senior neurologists who established standardized methods to perform and interpret the examination at the inception of the study (see http://www.ciberned.es/estudio-nedices for details). For participants who could not be examined, medical records were obtained from their general practitioners, in-patient hospitalizations, and neurological specialists (if they had any visit). The diagnosis of stroke and dementia was based on clinical data and medical record review [19–22]. The World Health Organization clinical definition of a stroke was applied [21, 22]. For the diagnosis of dementia, we applied Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV criteria [29]. Parkinsonism was diagnosed when at least two cardinal signs (resting tremor, rigidity, bradykinesia, and impaired gait/postural reflexes) were present, and PD was diagnosed in these patients when there were no secondary causes or atypical features [23–25]. Diagnostic criteria for ET were similar to those used in the Sicilian Study [30].
The current study was based on data from the second evaluation, which was carried out between 1997 and 1998, since it included a brief standard neuropsychological test battery (see below).
Neuropsychological battery
Many tools have been recommended for screening cognitive deficits in the older adults. Performance on these tests is commonly influenced by age, educational level, ethnicity, and functional limitations [31, 32]. Most screening tools have been developed in countries where secondary education is common. When these tests are used in populations with lower levels of education, there may be a high proportion of false positives [31, 32]. In the second evaluation of the NEDICES study, 54.4% of the 3,816 screened subjects were either illiterate or could only read and write. We therefore designed a neuropsychological battery for a lower educated cohort; it was similar to that used in the “Aging in Leganés” study, a population-based survey carried out in Leganés city (central Spain) [31, 32]. Our selection of tests was also influenced by their ease of use in an epidemiological setting.
Premorbid intelligence was evaluated with the Word Accentuation Test (WAT), which assessed the ability to read with accent thirty infrequent Spanish words (range from 0 [low performance] to 30) [33].
Global cognitive performance
The expanded 37-item version of the Mini-Mental State Examination (MMSE) was administered [34]. This version was adapted from the standard MMSE, which adds three additional items: (1) an attention task (i.e., “say 1, 3, 5, 7, 9 backwards”), (2) a visual order (i.e., a man raising his arms), and (3) a simple construction task (i.e., copying two overlapping circles) [34].
Speed of cognitive processing
The Trail Making Test-A was used and subjects linked, in successive order, the numbers 1 to 25, which were randomly distributed on a sheet of paper [35]. Scoring was based on the time required to complete the task (>5 min versus ≤5 min) and number of errors (higher scores indicate poorer cognitive performance) [35]. In the current research, the cut-off point (>5 min versus ≤5 min) was pre-specified in the protocol. Data on Trail Making Test-B was not analyzed because many older adults could not complete the test.
Semantic fluency
Subjects were asked to name as many different animals and fruits as they could in 60 s [36]. Lower scores indicate worse cognitive performance.
Naming and recall
The six objects test was used [31]. Initially, the subject had to name six common visually presented objects (range 0 [lower cognitive level] to 6); after the coding phase, both immediate and delayed free recall (5 min later) tasks of the six objects (range 0 [lower performance] to 6) were applied. Moreover, participants completed a story recall task, which required listening to a brief narrative with six informational bits (range scores: 0–6) [31]. The participants recalled the information after hearing the story (immediate logical memory) and after a 20-min delay (delayed logical memory). Lower scores meant worse performance in this task.
Functional activity and psychological wellbeing
A Spanish adaptation of the Pfeffer Functional Activities Questionnaire (FAQ) was also administered [37]. The FAQ assesses ten common activities that require complex cognitive and social functioning [38]. The performance on these tasks is rated from 0 (normal) to 3 (dependent) and the total score ranges from 0–30 [38]. The Spanish version removes two items (assembling tax records, business affairs, or paper, and playing a game of skill, working on a hobby) and adds three additional items (handles own medications, can go outside alone, greets appropriately), and the total score for the eleven items ranges from 0 (completely independent)–33 (completely dependent) [37]. We also assessed the psychological well-being using the Philadelphia Geriatric Center Morale Scale (PGCMS) [39].
Exclusions and final sample for the analyses
Out of 5,278 participants screened at baseline (1994–1995), 672 were lost to follow-up either because they declined (n = 112) or were unreachable (n = 560), and 790 died before they were contacted the second time (see Fig. 1). Three years later, out of the 3,816 who participated at the second wave (1997–1998), we excluded 673 participants because they had an incomplete neuropsychological assessment and 753 due to them having dementia or other neurological conditions, such as cerebrovascular disease, PD, or ET. We decided to exclude these neurological conditions because the main aim of this study was to assess what cognitive domains may predict mortality and its association with specific causes of death in a population-dwelling sample of elders without dementia. These factors are known to increase both the risk of cognitive impairment or dementia, [40–43] and mortality [25, 44]. Therefore, the eligible sample for the study consisted of 2,390 subjects.
The group of 1,426 subjects who were excluded because they had an incomplete neuropsychological assessment or any neurological disease were older than those who were eligible (mean = 77.4±7.0 versus 75.1±5.8 years, Mann-Whitney test, p < 0.001), less educated (267 [18.9% ] versus 222 [9.3% ] illiterate, chi-square = 82.66, p < 0.001) and a larger proportion were women (862 [60.4%] versus 1,369 [57.3%], chi-square = 3.69, p = 0.05).
Follow-up data on death were collected until May 1, 2007. The date of death was obtained from the National Population Register of Spain (Instituto Nacional de Estadística). In the Spanish context, a doctor completes a death certificate on all individuals at the time of death in accordance with the recommendations of the World Health Organization. The certificate is then sent to the local authority in the municipality where the person had been living, and the information is collected in a National Register. The assignment of any death cause is based on the major illness or injury which started the chain of pathological events that directly led to death (http://www.who.int/topics/mortality/en/). Using the International Classification of Diseases –ICD (9th Revision for deaths that occurred prior to 1999, [http://www.cdc.gov/nchs/icd/icd9.htm] and 10th Revision [http://www.cdc.gov/nchs/icd/icd10.htm] for deaths occurring from 1999 to 2007, the NEDICES investigators classified the cause of death into 6 main categories: dementia, cerebrovascular disorders, cardiovascular disorders (pulmonary embolism, congestive heart failure, myocardial infarction, heart or aortic rupture, and asystole), respiratory diseases, cancer, and other causes (infections, trauma, genitourinary or gastrointestinal disorders).
Statistical analyses
All statistical analyses were performed using SPSS, version 21.0 (IBM Corp., NY, USA). Age, number of medications, PGCMS, and neuropsychological test scores were not normally distributed (all Kolmogorov-Smirnov test, p < 0.05), even after log-transformation. Therefore, we used Mann-Whitney tests to analyze these continuous variables. The chi-square test was used to analyze categorical variables. Moreover, we used Cox proportional hazards models to estimate the relative risk of mortality based on cognitive tests scores. Cognitive test scores of individuals were divided into tertiles with the aim of comparing hazard ratios (HR) with 95% confidence intervals (CI) of individuals in the lowest tertile versus all other tertiles [reference group]. This non-conservative segmentation strategy was selected in order to identify individuals with subtle cognitive deficits on each domain. The time variable was the years from the date of the follow-up evaluation (1997–1998) to either (i) May 1, 2007 in living participants or (ii) the date of death in participants who had died prior to May 1, 2007. Different baseline variables were considered as potential confounders in Cox proportional hazards analyses. These factors were age in years, gender, geographical area (Lista, Arévalo, and Las Margaritas), ever smoked (ex-smokers and current smokers), ever drank [ex-drinkers and current drinkers], arterial hypertension, diabetes mellitus, heart disease, cancer (including active and inactive tumors), depressive symptoms or antidepressant use, subjective memory complaints, number of medications, regular physical exercise, and PGCMS (total score). In the adjusted models, we considered those variables that were associated with the vital status in univariate analyses (p < 0.10). A value of p < 0.10 rather than p < 0.05 was conservatively chosen to select different sources of confusion from the univariate analyses. In the models, we also adjusted for global cognitive performance (37-MMSE total score), educational level, and premorbid intelligence.
We tested the proportional hazards assumption for each model generating time dependent covariates, interactions of the predictors and function of survivaltime. Time-dependent variables were included in the adjusted models [45], but none of them were significant, indicating that the proportional hazards assumption was not violated [45].
Finally, in order to examine the association between cognitive test scores (lowest tertile versus all other tertiles) and the dichotomous outcome of multiple causes of death, logistic regression analyses (odds ratios [OR] and 95% CI were performed adjusting for variables that were associated with the vital status in univariate analyses (p < 0.10).
RESULTS
The mean duration of follow-up for 2,390 participants was 8.0 years (median 9.2 years, range 0.01–10.7). Eight hundred and eighty (36.8%) of 2,390 participants died after a median follow-up of 5.5 years (range 0.01–10.5). Figure 1 shows the flow diagram of the study.
Baseline characteristics of the participants who died versus those who were still alive are shown in Table 1. As expected, those who died were older, predominantlymen, took more medications, were more likely to have ever smoked or drunk and did less regular physical exercise. In addition, they were more likely to have hypertension, diabetes mellitus and heart disease. Moreover, there was a higher proportion of the deceased participants from Arévalo (rural area) and the PGCMS (total score) was lower in participants who had died versus those who were still alive.
Neuropsychological performance of deceased and alive participants is shown in Table 2. Groups significantly differed in all the cognitive domains, including global cognitive performance, speed of cognitive processing, semantic fluency, naming, and recall. Deceased participants also had a higher FAQ score (i.e., poorer function) at baseline.
In unadjusted Cox models, the risk of mortality was associated with the lowest tertiles of each neuropsychological test and the Pfeffer’s FAQ (Table 3). We performed Cox proportional hazards analyses that were adjusted for several covariates such as age,gender, geographical area, educational level, premorbid intelligence, 37-MMSE total score, lifetime smoking, lifetime drinking, diabetes mellitus, hypertension, heart disease, cancer, number of medications, regular physical exercise, and PGCMS. After the adjustments, we found that the risk of mortality was associated with the lowest tertiles for speed of cognitive processing (time to complete the Trail Making Test-A), semantic fluency (fruits and animals), delayed recall (pictures and prose), and the Pfeffer’s FAQ(Table 3).
Primary cause of death registered on the death certificates differed significantly depending on the cognitive performance (Table 4). Cerebrovascular disorders and dementia were more often reported in those who required more than 5 minutes to complete the Trail Making Test-A. In addition, cardiovascular diseases were significantly more frequently reported in those in the lowest tertiles of the semantic fluency tasks for animals. Dementia as cause of death was increased with poor performance in semantic fluency (animals and fruits). On the contrary, cancer was less often reported by those in the lowest tertiles of semantic fluency (animals and fruits) vs. those in the remaining tertiles. Finally, as expected, dementia was reported significantly more by those individuals in the lowest tertiles of delayed recall tests (recall of pictures and prose) than those located in the upper tertiles. Other causes of death and cerebrovascular disorders were also related to poor performance on delayed freerecall tasks.
Types of cancers listed on the death certificates did not differ significantly by tertiles of semantic fluency (animals and fruits), except for malignant neoplasms of respiratory and intrathoracic organs, which were less reported in those individuals in the lowest tertiles of semantic fluency (Table 5).
In logistic regression analyses that adjusted for age, gender, geographical area, lifetime smoking, lifetime drinking, diabetes mellitus, hypertension, heart disease, cancer, number of medications, regular physical exercise, and PGCMS total score (i.e., p < 0.10 level in univariate analyses), we found that (1) dementia mortality was associated with the lowest tertiles of delayed free recall for pictures (OR [odds ratio] = 4.05, 95% CI 1.91–8.58 p < 0.0001) and prose (OR = 2.74, 95% CI 1.34–5.60 p = 0.01); (2) cerebrovascular disease mortality was associated with the Trail Making Test-A (>5 minutes) (OR = 2.23, 95% CI 1.11–4.49 p = 0.02); and (3) cancer mortality was associated with the lowest tertiles of semantic fluency for animals (OR = 0.69, 95% CI 0.49–0.97 p = 0.03), but not with fruits (OR = 0.84, 95% CI 0.60–1.19 p = 0.30). Other relationships between death causes and cognitive performance did not reach statistical significance after adjustment.
In secondary analyses, we excluded the individuals with cancer at the entry of the study (N = 93). In regression analyses that adjusted for age, gender, geographical area, lifetime smoking, lifetime drinking, diabetes mellitus, hypertension, heart disease, number of medications, regular physical exercise, and PGCMS total score (i.e., p < 0.10 level in univariate analyses), we found that the lowest tertiles of semantic fluency for animals was associated with cancer mortality (OR = 0.68, 95% CI 0.46–0.99 p = 0.04), but not with fruits (OR = 0.83, 95% CI 0.57–1.19p = 0.31).
The association of cognitive scores at baseline and mortality was of similar magnitude using baseline age as the time scale in alternative analyses (data not shown).
DISCUSSION
In this prospective population-based study, we found that poor performance on delayed recall and speed of cognitive processing tests were associated with dementia and cerebrovascular disease mortality, respectively. Meanwhile poor performance on semantic fluency was associated with decreased cancer mortality after controlling the effect of multiple covariates. Interestingly, our results suggest that, even using non-conservative cut-off points to classify individuals with poor cognitive performance (the lowest tertile include people in the normal range of scoring), there were specific and significant associations between worse performance in cognitive domains and cause specificmortality.
The underlying mechanisms that modulate the association between cognitive ability and mortality are poorly understood. In this respect, it is known that cognitive reserve proxies (e.g., education) delay the cognitive decline and are associated with decreased risk of dementia, [46, 47] which is considered a major cause of death [1]. Moreover, healthy habits (e.g., physical exercise and diet) and social engagement are potential mediators of the relationship between cognitive performance and risk of death [48]. In addition, childhood intelligence quotient has been significantly associated with survival up to 76 years, indicating that deficient brain development is correlated with later adulthood illness [49]. Finally, there are genetic and environmental factors associated with the risk of cognitive impairment, [50, 51] which is in turn a risk factor of dementia and mortality [2].
It is intriguing that specific neuropsychological domains might relate to causes of mortality. Worse performance on delayed recall (pictures and prose) was associated with dementia mortality. This finding is consistent with previous data which showed that the impairment on memory domain was associated with increased dementia mortality [2].
Poor performance on semantic fluency was linked with increased mortality from cerebrovascular disease and decreased mortality from cancer. In this context, one study found verbal fluency was significantly associated with greater all-cause and/or cardiovascular mortality [13], but to our knowledge none have established a link with decreased risk of cancer mortality. It should be noted that impairment on semantic fluency (when measured by naming animals) is a predictor of Alzheimer’s disease [52]. Therefore, we would perhaps expect those with low semantic fluency to develop and die with Alzheimer’s disease rather than cancer, which is coherent with previous reports showing an inverse association of cancer and neurodegenerative diseases (e.g., Alzheimer’s disease) or cognitive decline [53, 54]. As yet undiscovered mechanisms may either promote a neurodegenerative process (uncontrolled cellular destruction) or annul other conditions, namely, cancer (uncontrolled cellular proliferation) [54]. Both cancer and neurodegenerative disorders are characterized by a disarrangement of cell-regulation mechanisms, with increased cell survival and proliferation in the former and with increased cell death in the latter process [54]. We note in our study that this observation was only statistically significant for animals after adjusting for different covariates. Previous studies have indicated that category fluency measures are not totally equivalent and may be differentially affected by diverse variables related to theindividual [55].
Poor performance on psychomotor speed was associated with cerebrovascular disease. In line with this latter finding, slower reaction times and poorer memory scores have been related to an increased risk of mortality from cardiovascular disease, stroke, and respiratory disease in the UK Health and Lifestyle Survey [10]. However, other population-based cohorts suggested a less specific association between cognitive domains and death causes. For instance, poorer scores on cumulative verbal learning and vocabulary tests predicted mortality in a sample of 3,572 individuals aged between 49 and 93 years. Only the learning test emerged as the significant predictor of a particular category of deaths, namely ‘other causes’ [9]. Differences in terms of cohort characteristics, covariates taken into account and classification of death causes could explain these findings discrepancies.
Our study has some limitations. First, the NEDICES study only included people aged 65 and older. However, the relationship between cognition-mortality in older people is of special interest, since prevalence of cognitive problems is higher in this population [19, 20]. Second, we used a relatively restricted neuropsychological test battery that did not sample all cognitive domains. However, our battery was designed to meet the constraints of an epidemiological study with almost 4,000 elderly subjects where the level of education is usually low [31]. In this sense, the educational factor may partially explain the low performance of our sample in the Trail Making Test-A compared to other populations. Further, it is possible that a percentage of participants were unable to complete the test within 5 minutes due to sensory, motor, or visual problems. However, the inclusion of these data is an inherent issue with larger population-based studies. Third, while the baseline sample comprised 3,816 participants, the final sample comprised 2,390 due to the exclusions. This fact limits the generalization of results to the whole cohort. Finally, the presence of reverse causality related to subclinical disease cannot be excluded in this study.
This study also has several strengths. First, the study was population-based, allowing us to assess a group of people without dementia who were unselected for treatment considerations. Second, the mortality data analyses were conducted prospectively. Finally, we were able to adjust for the potential confounding effects of multiple important factors, including global cognitive measurement and premorbid intelligence.
Using a prospective, population-based design, we found that increased risk of mortality was associated with poor speed of cognitive processing, poor semantic fluency, and poor delayed free recall. Performance on specific cognitive domains were related to different causes of death, notably the association between impaired semantic fluency and decreased cancer mortality. Further investigations using a more extensive assessment and different follow-up periods are needed to confirm these specific associations.
Footnotes
ACKNOWLEDGMENTS
Additional information about collaborators and detailed funding of the NEDICES study is located at 
. This research was supported by FEDER funds. Dr. Benito-León is supported by the National Institutes of Health, Bethesda, MD, USA (NINDS #R01 NS39422), the Commission of the European Union (grant ICT-2011-287739, NeuroTREMOR), and the Spanish Health Research Agency (grant FIS PI12/01602). Dr. Bermejo-Pareja is supported by NINDS #R01 NS39422 from the National Institutes of Health, Bethesda, MD, USA and from the Commission of the European Union (grant ICT-2011-287739, NeuroTREMOR).