Age-related deficit accumulation and the risk of late-life dementia

A growing number of factors are associated with dementia risk. Reports just from 2014 reify vascular risk factors (in the degree of intracranial artery stenosis) [1], implicate impaired sleep [2], and raise questions about gynecologic surgery [3], and pesticides [4]. Even recognizing that seemingly disparate factors might share common mechanisms, it remains unclear what to make of the number and diversity of risk factors. Their very disparity might hold a clue. Inasmuch as the risk factors appear to have so little in common, their collective role might simply be to induce aberrant repair processes [5].

Despite these many new exposures associated with dementia, age remains the single biggest risk factor. Likewise, it is also the biggest risk factor for death, and this may help us understand how to interpret the growing list of dementia risk factors. The risk of death increases exponentially with age, but not everyone of the same age has the same risk of dying. People at an increased risk of death compared with others of the same age are said to be frail [6]; this greater risk typically obtains for other adverse outcomes, too, including, institutionalization, health service use, and worse health.

How best to make operative the concept of frailty is disputed, but as detailed later, both well-established methods (that is, the frailty syndrome/phenotype and a frailty index (FI) based on the accumulation of health deficits) have been linked to late-life cognitive impairment. The FI counts how many deficits a person has (broadly defined by biological and clinical characteristics); it can be calculated simply as the ratio of the deficits present in a person to the total number of deficits considered in a given study setting [7]. A possibility linking age-related accumulation of health deficits and dementia is that the latter represents the failure of a high order, integrative function (for example, cognition) in a system that is close to failure (for example, in people who are frail) [8]. By that line of reasoning, late-life dementia can be understood with less attention to specific risk factors, and more to how these risk factors operate in the aggregate [9]. In other words, just as the higher risk of death in older adults reflects their frailty more than it does their age, so too might the increased risk of dementia arise in relation to their general health status, which can be represented by the number of health deficits that they have accumulated.

Some of this is not new, in that, for example, dementia risk is known to be related to the combined effect of many factors [10]-[14]. Likewise, several reviews suggest associations between frailty and cognitive decline [15],[16], especially when death is modeled as a competing risk [9]. As Barnes and Lee [9] pointed out, the most widely used approach to dementia prediction uses weighted scales that combine a small number of selected risk factors, each of which is significantly associated with dementia [9]. Even so, only moderate accuracy results [9],[10],[17]. Notably, many such scales involve age and genetic risk factors (for example, ApoE4) that are hardly modifiable [9],[17]. Approaches to pick the best possible predictors, including the best number, and their optimal weights are difficult: empirically, multiple combinations of different variables can have equivalent value in predicting outcomes.

Such considerations also inform our approach, but with a different take that, in our view, has consequences for both dementia epidemiology and potentially for management. Instead of picking out factors that only individually are statistically significant, and then seeing whether they survive multivariable modeling, we were struck by the fact that different subsets of variables can give comparable predictions. This suggests both considerable diversity in individual health and multiple dependencies among health measures [8]. In other words, seemingly insignificant factors, as judged by statistical considerations alone, can still impact the system and modify risk.

When these seemingly insignificant factors are combined, they may help reveal the state of the system [7],[8]. In this regard, we reported that poor general health, as manifest by the presence of multiple health deficits, increases dementia risk [18]. Note that this was shown to be the case for health deficits that otherwise were not known to be directly linked to dementia. Note too that these factors worked in combination; indeed, in contrast to the convention of including in the multivariable risk model only items that were individually significantly associated with dementia, we proposed an index made up solely of nontraditional risk factors (that is, the frailty index of nontraditional risk factors), most of which were not individually associated with dementia risk [18]. That approach (of including all risks, without regard to their individual statistical significance) borrows from signal-detection methods long used in information theory [19]. Recently, again by using this approach, our group has replicated the finding that nontraditional risk factors combine to predict cognitive decline [20].

To understand better how frailty and age influence dementia risk, we now consider (1) how traditional (including vascular risk factors and cognition-related measures) and nontraditional risk factors (including measures of health that are not considered as dementia risks) compare in their predictive ability, and (2) how the predictive value of dementia risk might vary by the number of deficits considered. We addressed these questions by reevaluating the Canadian Study of Health and Aging (CSHA), an established cohort with 10-year dementia follow-up. Multiple indices were constructed by using various numbers of randomly selected deficits that either have or have not been associated with dementia risks. To isolate dementia from death, these outcomes were assessed separately. Given the likely gender differences in dementia and mortality in relation to deficit accumulation [8],[21],[22], we evaluated the risks separately for men and women.

Overall, cognitively healthy older adults (men or women) who survived 10 years were at baseline relatively younger, better educated, and had lower levels of frailty, compared with those who became cognitively impaired or demented, or who died (Table 2). Cognitively healthy survivors were also more likely to have been married. Although those who died were the least healthy at baseline, the baseline profiles of people who died or developed dementias were otherwise similar (Table 2). In each group, women were older on average and less likely to have a living spouse compared with men (Table 2). Each deficit (except language inability in women) was individually associated with an increased risk of death, whereas only some deficits (27 of 42 in men; 25 of 42 in women) were individually associated with dementia (Table 1).

When the deficits were combined, both the frailty (all-factor) index and the nontraditional risk-factor index increased with age; this was not the case for the vascular risk-factor index (Figure 2). Adjusted for age, the odds ratios per additional deficit increased for both death and dementia in both women and men, albeit a greater odds for each adverse outcome per FI increment in men (Table 3). Regarding objective 1 (the nature of the risk factors), in each risk-factor index, subjects in the tertile with the highest index scores had the highest mortality and dementia rates, whereas those with the lowest index scores had the lowest, showing that seldom did a person with an FI <0.2 (that is, with fewer than nine of 42 deficits) develop dementia over 10 years (Figure 3).

Regarding objective 2, in a multivariable logistic regression model adjusted for age, when the number of deficits considered in the index increased, the ORs for death and dementia each increased for the index. Likewise, the impact of age decreased. For death, when the number of deficits included in the index increased from 5 to 42, the Wald statistics decreased by 16% for age (from 341to 286) and increased by 71% for the index in men (from 37 to 126). For women, the corresponding changes were a 22% decrease for age and a 73% increase for the FI. For dementia, when the number of deficits included in the index increased from 5 to 42, the values of the Wald statistic were a 15% decrease for age and 64% increase for the index in men; 11% decrease for age and 63% increase for the index in women. By contrast, the change in the beta estimates for the constants were negligible: -13.63 versus -13.72.

The FI, constructed by using all 42 variables, best predicted subjects who died and those who developed dementia over 10 years (Figure 4: diamond symbols; AUC = 0.71 ± 0.02 for men, 0.72 ± 0.02 for women for death; AUC = 0.67 ± 0.03 for men, 0.66 ± 0.03 for women for dementia). The predictive power was associated with the total number of variables considered: by using any given variable, the AUC for dementia was usually <0.55, and seldom exceeded 0.60. The AUC increased with the number of randomly selected variables under consideration, and gradually stabilized when n > 25, at which the values converged (Figure 4).

In this secondary analysis of a well-established population-based study of cognitive health, we examined the predictive power of risk indices constructed by using differing types and numbers of deficit variables for 10-year dementia and mortality. The data showed that the death and dementia outcomes were both closely associated with baseline health status, as represented by the level of deficit accumulation in an FI. More interesting, the predictive value of the risk indices increased as a function of the number of deficits included in the index. Here, the deficits covered a wide range of self-reported health problems, including diseases, symptoms, and disabilities, some of which were known dementia risk factors, and some were not. As such, our study provides evidence suggesting that the risk of dementia and death may be determined more by the overall state of health, and notably more by how many health problems an individual has, and less by exactly what these problems are. This general result held true for both men and women.

In other words, each deficit adds information, even those that individually do not significantly increase dementia risk. It was also interesting to note that for both death and dementia, women tolerated deficits better than men did, as evidenced by lower risks per deficit. This result parallels similar investigation by our group and others in understanding age-related deficit accumulation and the risk of death [22],[23].

It is well understood that people are more likely to develop health problems at variable rates as they age. Important for our study, the aging brain subject to late-life cognitive dysfunction appears to be especially deficit prone [16]. For example, just over the past decade, more than 20 dementia risk factors have been identified for sporadic Alzheimer disease, the most common cause of late-life dementia, which are in addition to advanced age and genetic determinants (for example, carrying the ApoE e 4 allele). This long list has included, in addition to the items cited at the outset from 2014 [1]-[4], traumatic head injury, stroke and transient ischemic attack (TIA), hypertension, heart diseases, diabetes mellitus, obesity, high-fat diet, metabolic syndrome, low dietary intake of antioxidants, fish, vegetables, or fruits, reduced high-density lipoprotein cholesterol, deficiency of vitamin B or trace minerals, hypothyroidism, high homocysteine, abnormal serum-hemoglobin level, sleep apnea, anxiety, depression, poor resilience, anemia, smoking, alcoholism, anesthesia, environmental toxic exposure and pollution, low physical, cognitive, or social activities, and low education, income, or social status [10],[23]-[27].

Although it is reasonable to expect that more health problems will join this long list, our study supports aggregating these items to assay whether a new “risk factor” is informative. In short, we support the proposal that this is best done explicitly, in a frailty/deficit accumulation index, rather than implicitly, through age [9]. Here too, we note that age is not part of the FI. Instead, the FI represents an alternative way to tease out information that otherwise might be lost in age-adjusted models. That consideration, too, informed the strategy of evaluating the impact of age in the multivariable logistic regression model as we increased the number of health deficits that we included in the FI (Figure 4): the decline in the effect of age is another way of demonstrating the cumulative increase in information for the FI as more items are included in it.

In this way, our study also contributes to the current literature relating general health to adverse outcomes. An FI composed of a variety of health deficits has been used to predict multidirectional changes in cognitive test scores (for example, getting worse, getting better, or remaining the same), and not just in relation to dementia/no dementia, or a dichotomous outcome that includes the “cognitive impairment, not dementia” category [28],[29]. An index made up of baseline health deficits that are neither cognitive risk factors (for example, hypertension, heart disease), nor cognition-related measures (for example, foot problems) has also been found to predict dementia [18]. More recent work has extended this approach from dementia to other multiply determined illnesses that are both common and age associated.

In a re-analysis of data from the Canadian Heart Health Survey, non-traditional risk factors were combined to predict an elevated risk of coronary heart disease, and remained so in analyses adjusted for age and for traditional risk factors, also combined in a risk-factor index [30]. In short, consider the analogy with the presentation of late-life disease. When many other illnesses are present, it is more likely to result in so-called “atypical” manifestations; for example, myocardial infarction in a frail patient is as likely to be seen with delirium or a fall, whereas older adults with heart disease and not a lot of other health deficits are more likely to have chest pain [8],[31]; so too must we accept that in people with many health deficits, it is unwise not to consider their combined effect.

Can this approach of aggregating risk factors aid our understanding of mechanisms? The FI approach has been criticized for lacking such specificity [16], but it might in fact be more revealing in exactly this way. Consider that many of the risk factors included in the FI would appear to have little in common with each other, except that they are injurious in some way. This is true of a larger set of risk factors that might include other items not measured here, such as paternal age at birth, or specific toxin exposures, or details about head injury. So at a broad level, if what these risk factors have in common is the need for the body to mount responses, then in people with aberrant repair processes, or exhausted ones, various types of damage can accumulate [5]. This can be formulated precisely, by using the mathematics of queuing theory. In brief, deficit accumulation occurs when the number of insults to which an organism is exposed (that is, the damage rate) surpasses the ability for the damage to be either removed or repaired, and so deficits occur and accumulate [32]. The length of a queue (here, the number of deficits) is a function of the number of people arriving at the queue (for us, the damage rate), and how quickly the people can be processed in the queue (here, the repair/removal response). Indeed, a recent study of the neuropathology of Alzheimer disease in relation to frailty showed the best explanatory power only when Alzheimer disease pathology, macroinfarcts, and nigral neuronal loss all were considered in a single model [33].

In short, if we accept that cognition in older adults is formally complex and dynamic (and this can be demonstrated with simple clinical and imaging tools [29]), then it will be important, in achieving a fuller understanding of how cognition becomes impaired, to have tools that allow dynamic changes to be understood. That this approach is indifferent to exactly which mechanisms are involved is more a strength than a weakness. Second, from a clinical standpoint, this is a cautionary note to take into account the general health of the patient, a focus that can get lost in subspecialty practice. Our data suggest that good management of people with dementia should pay attention to other active illnesses as well as medical management.

Our data must be interpreted with caution. The results underscore the competing outcome effect between death and dementia: subjects might die before they have a chance of developing dementia. Here, subjects who had troubles with kidney, bladder control, chest pain, high blood pressure, heart/circulation at baseline were at such a high risk of death that these factors did not increase the risk of dementia in the survivors. Moreover, whereas 10-year survival status was known for all participants, cognitive reassessments were made at the 5-year follow-ups in CSHA. In consequence, some subjects could have died with undiagnosed dementia. This, and the 14% survivors with an unknown cognitive outcome, made it possible that the dementia rate of the sample was somewhat underestimated. Note too, in the index variables, we treated each deficit equally (that is, without variable weighting), even though it is clear that different health problems could contribute differently to an outcome, and the predictive value may appear higher with use of weighted scales [9],[11],[34]. An obvious advantage of not assigning a specific weight to a variable is that it improves generalizability, as the result may depend less on which exact deficits are available to be used.

Our data contribute to the discussion about biomarkers. Whatever their specific advantage, even individually significant measures such as early amyloid beta retention, CSF, serum, and neuroimaging biomarkers generally result in no better than moderate prediction accuracy [17],[35],[36]. This may simply reflect that too few biomarkers are being deployed to account for the various ways in which dementia can develop. Further, whatever the statistical technique used, it seems likely now that, from a biologic standpoint, the independence of one biomarker from another must be reexamined. In this regard, the parallel case exists in relation to mortality prediction in relation to health deficits, in which different subsets of variables can give comparable results in the prediction of adverse outcomes; many commentators note that this reflects not just diversity in individual health, but multiple dependencies among health measures [8],[16]. In that case too, whether a set of variables represents “the best” predictors varies by target outcomes, even within the same dataset. Here, the predictive value of the index became increasingly greater as the number of variables used increased, until it was sufficiently large (for example, >25). Whether further information offsets noise in the data requires further evaluation in other datasets. Given that a complex system can be reflected by the redundancy of acquired deficits [8],[19], when a sufficient number of deficits are taken into account, each adding a bit of information to the system’s profile, resulting in broad coverage of diverse health states, the selection of a small number of specific variables shows no advantage.

Here, too, we observed a relatively low predictive value for dementia, which was lower than death prediction. One possibility is that dementia represents a state that can involve complex origins related to the dynamics of brain structure and function [29],[37]. In addition, relations of risk factors to dementia outcomes are complicated, varying by follow-up duration and sample profiles [27],[38],[39]. In this regard, evaluations of structural and functional brain health can be useful [27],[40],[41], enlightened partially by studying whole-brain atrophy and lesion changes in aging and dementia [29],[42]. Further research will test how health status combined with brain changes can help with dementia prediction.

The predictive value of 10-year dementia risk varies by the number of deficits considered. Traditional and nontraditional risk factors did not have significantly different predictive ability. The variety of items associated with dementias suggests that some part of the risk might relate more to aberrant repair processes than to specifically toxic insults. The epidemiology of late-life illness must consider overall health status.