Empirically Defining Trajectories of Late-Life Cognitive and Functional Decline

INTRODUCTION

Alzheimer’s disease (AD) is a progressive neurodegenerative disease typically characterized by slow cognitive and functional decline; however, considerable variability in rate of progression makes it hard to predict an individual’s future course. AD is a continuum beginning pathologically decades before symptoms appear (preclinical stage) [1]. Symptoms progress through an early to late phase of mild cognitive impairment (MCI) and then worsen to include a more pronounced functional impairment referred to as dementia. Clinicians would like to provide patients with a likelihood of progression and how fast it may occur. Early detection and intervention might delay cognitive deterioration and/or reduce costs, especially in someone who could progress quickly. It is also important to understand disease progression for future clinical trials.

While historically AD progression has been modeled as a linear process, we now know that progression is typically slower in early disease stages and then increases more rapidly as the patient progresses into dementia [2, 3]. Sona and colleagues reviewed 82 studies assessing the role that various factors may have on rapid cognitive decline in AD [4]. In the studies reviewed, disease progression was modeled using individuals’ mean cognitive scores that were regressed on covariates such as diagnosis, age, and education, to determine their impact on disease progression [4]. These approaches attempt to describe the relationships among variables with the goal of identifying significant predictors of outcome. The limitation of these types of analyses is that they do not allow for subpopulations with variable rates of progression and they assume that one trajectory is adequate to describe the whole study population.

Growth Mixture Modeling (GMM) is a modeling approach that looks for relationships among individuals’ trajectories with the goal of classifying groups based on similar growth patterns [5, 6]. This approach enables identification of unobserved subpopulations based on similarities in their growth patterns over time rather than assuming that one trajectory approximates the whole population. Additionally, GMM allows for non-linear (e.g., quadratic) progression curves. Several studies have used GMM to follow trajectories of change in cognition and/or functioning in subjects at various stages of AD. Small and Backman utilized GMM to follow longitudinal trajectories of cognitive change as assessed by decline in Mini-Mental State Examination (MMSE) scores in preclinical AD and reported that a two-group quadratic model (and not a linear model) of cognitive decline provided the best statistical fit to the observed data [7]. Others have used GMM to follow trajectories of cognitive decline in cognitively healthy older adults followed for up to 54 months [8], and in Caucasian subjects with possible or probable AD for up to 13.5 years [9]. Leoutsakos et al. used parallel-process GMM to model class membership jointly as a function of both cognitive and functional decline by assessing change in MMSE and Clinical Dementia Rating Sum of Boxes (CDR-SB) scores and reported four distinct latent classes [10].

The objectives of this exploratory study were to define unobserved subgroups (latent classes) from participants in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database who had similar growth patterns of both cognitive and functional change using GMM, and to identify characteristics associated with those trajectories. We then applied Classification and Regression Tree (CART) analysis [11] to develop a decision tree that uses clinical predictors that could be applied to individuals to determine which trajectory, as determined by GMM, they will most likely follow.

METHODS

RESULTS

The total dataset included 325 HC, 279 EMCI, 372 LMCI, and 216 AD dementia subjects. The average length of follow-up for the FAQ was 16.3 months (SD ± 7.1) and for the ADAS-Cog₁₃ was 16.2 (± 7.1). Overall, 60% subjects were considered amyloid positive with amyloid positivity increasing with advanced disease stage (HC 35.1% , EMCI 51.3% , LMCI 71.8% , AD 91.7%). There were 31 cases that had discrepant amyloid results between CSF and FBP-PET (20 of the 31 discrepant cases were CSF positive but FBP-PET negative, and the opposite was true for the remaining 11 cases).

Table 2 shows a number of GMM models from which the final model was selected. The final model identified three quadratic trajectories across all subjects for cognitive/functional change as measured by the ADAS-Cog₁₃ and FAQ bivariate analysis. While the graphs in Fig. 1 illustrate trajectories using mean values of individuals assigned to each class for each scale separately, GMM took both cognition and function into consideration when determining the classes. Additionally, Fig. 2 illustrates spaghetti plots of individuals’ trajectories in each class, overlayed with the fitted mean trajectory of that class.

Class 1 (n = 162, 13.6%) had the highest baseline FAQ and ADAS-Cog₁₃ scores and steepest pattern of cognitive and functional worsening of all classes. In contrast, Class 3 (n = 819, 68.7%) had the lowest baseline FAQ and ADAS-Cog₁₃ scores and minimal change on both over time. Class 2 (n = 211, 17.7%) had an intermediate pattern with baseline FAQ and ADAS-Cog₁₃ scores that were more impaired than Class 3 but less than Class 1, and worsening on both scales that was slower than for Class 1.

Table 1 shows baseline demographics by latent class. Significant differences were observed among the three classes across most variables including age (p = 0.0021), APOE4 status (p <  0.0001), amyloid status (p <  0.0001), diagnosis (p <  0.0001), alcohol abuse (p = 0.0474), and all clinical and cognitive measures (p <  0.0001 except GDS p = 0.0007). No significant differences were observed for gender, education, or comorbidities of smoking, cardiovascular disease, or endocrine disease. On average, Class 1 subjects were older and more likely amyloid (92%) or APOE4 (69.1%) positive. The majority of Class 1 was diagnosed with AD dementia (79%) or LMCI (17.9%). This group showed the greatest amount of impairment across measures of cognition and function and the greatest neuropsychiatric symptom severity.

In contrast, Class 3 subjects were younger and less likely to be amyloid (48.2%) or APOE4 (38.3%) positive. Class 3 included HCs (39%) or were diagnosed with EMCI (29.9%) or LMCI (29.1%), and showed the least amount of impairment across all measures of cognition and function, and the lowest neuropsychiatric symptom severity.

Class 2 subjects were HCs (0.9%) or diagnosed with EMCI (13.7%), LMCI (49.8%), or AD dementia (35.5%). Subjects were close in age to Class 1, and a higher percentage than Class 3 was amyloid (84.4%) or APOE4 (64%) positive but slightly less than Class 1. Baseline scores on cognitive, functional, and neuropsychiatric scales were intermediate between Classes 1 and 3.

Figure 3 shows FAQ or ADAS-Cog₁₃ mean scores by diagnostic group graphed adjacent to each class’s trajectory line. As expected, the arrays for mean score lines were ordered according to level of impairment of each diagnostic group and either above or below the latent class line.

CART analysis identified the FAQ as the input variable most predictive of latent classes found in GMM with two decision nodes that separated subjects into their most likely class at an estimated prediction accuracy of 82.3% (SD 1.11%) (Fig. 4). The first node partitioned Class 3 if the FAQ score was <5. The remaining subjects were split at the second node into Class 2 if the FAQ was <14, and into Class 1 if the FAQ was ≥14.

DISCUSSION

We approached the topic of clinical trajectories by using a person-centered modeling approach (GMM) in contrast to most other reports that use variable-centered approaches [4]. This enabled us to discern unobserved subpopulations based on similarities in their growth patterns over time rather than to assume that one trajectory approximates the whole population or to specify subgroups based on one or more variables. We found that the combination of cognitive and functional change over a 24-month period followed three distinct trajectories, represented by one class with the lowest baseline cognitive and functional impairment and virtually no change (Class 3) and two classes that both progressed but were separated by their degree of baseline impairment. Additionally, there were some clinical characteristics that differed across the classes, such as diagnosis, APOE4, and amyloid status. Our study included cases ranging from HC through AD dementia as diagnosed in the ADNI cohort, while other studies may not have addressed this whole spectrum. Further, we only selected ADNI cases where amyloid status was known, whereas most prior reports do not have the benefit of knowing amyloid status. Finally, GMM allowed quadratic in addition to lineartrajectories.

Our study found three trajectories of cognitive and functional change. Class 1 had the highest baseline FAQ and ADAS-Cog₁₃ scores and the steepest pattern of cognitive and functional worsening. In contrast, Class 3, the largest group, had the lowest baseline FAQ and ADAS-Cog₁₃ scores and minimal change on both scales. Not surprisingly, its constituents had diagnoses suggesting earlier stages of disease (% HCs>EMCI = LMCI). It is notable that almost half the individuals in this class of non-progressors were amyloid positive and we may have seen this class progress with a longer follow-up period. Class 2 had an intermediate pattern with baseline FAQ and ADAS-Cog₁₃ scores that were more impaired than Class 3 but less than Class 1, and worsening on both scales, but less steep than for Class 1. Consistent with this pattern, Class 2 constituents, by diagnosis (% LMCI>AD dementia>EMCI), reflected less overall disease severity than in Class 1 (% AD dementia>LMCI>EMCI). A longer follow-up period could confirm whether the trajectory for Class 2 would eventually look like that for Class 1 as patients become more impaired. It is important to note that we did not look at the relationship between cognitive and functional decline because GMM was a bivariate analysis, nor do we know how the scales perform relative to each other in a psychometric fashion.

Previous studies have described trajectories of change in cognition and behavior in subjects with MCI or AD dementia. Xie and colleagues [16] used group-based trajectory analysis to identify distinct cognitive change patterns among a cohort of 187 MCI patients from two geriatric outpatient clinics over a course of up to 3.5 years. Five trajectories were identified and labeled based on their baseline MMSE score. Their two least cognitively impaired groups (MMSE of 27 and 29) had a stable course, which is consistent with our findings for Class 3.

Other studies with longer patient follow-up periods (up to 13.5 years) [9, 10] found that cognitive function, as measured by the MMSE, eventually declined to very low levels in almost all of the trajectories. In contrast, studies that delineate a group with a stable cognitive trajectory [2] or “nonprogression” in cognition and function (like ours) have the shortest follow-up periods (2 to 3.5 years). Wilkosz and colleagues [9] used latent class modeling and found six trajectories of cognitive and behavioral decline in a cohort of 201 Caucasian subjects with possible or probable AD dementia followed for up to 13.5 years. The concomitant variables included in the best latent class trajectory model were initial MMSE and age. Also, two of our co-authors (JSL and CGL) completed an analysis to identify classes of progression trajectories in incident AD dementia cases from the Cache County Dementia Progression Study in up to 8 years of follow-up, and to identify baseline predictors of membership for each group [10]. Using bivariate GMM based on MMSE and CDR-SB scores, those authors identified four latent classes. Class 1 was the largest (72%) and had theslowest progression as compared to the other three classes with more rapid worsening. In a multivariate model, only MMSE was a statistically significant predictor of likely class membership. However, these two studies only included subjects with dementia. Our study also included HC and EMCI subjects who progress more gradually over many years [17]. Overall, our findings are consistent with reports that the course of cognitive impairment in clinically diagnosed AD is variable, starts slow and progressively increases over time [2, 17].

A recent review of 82 studies that reported at least one factor associated with rate of progression for an AD patient population considered an extensive list of potential factors such as age, age of onset, gender, education, family history of dementia, MMSE and CDR scale baseline scores, comorbidities, APOE genotype, and cholinesterase inhibitor use [4]. Study results were often contradictory, and no reliable conclusions regarding factors possibly associated with rapid cognitive decline could be made; however, a trend was found for younger, more educated and/or more impaired patients. The more rapid decline in the higher educated is consistent with the hypothesis that higher cognitive reserve protects from early clinical manifestation of AD, but later patients deteriorate more rapidly as the disease progresses.

We found that class membership was associated with baseline age, APOE4 and amyloid status, diagnosis, and all cognitive, functional, and neuropsychiatric scores. Subjects that progressed were more often APOE4 or amyloid positive, older in age, and later in their disease stage. A number of studies, but not all [18], report that younger subjects have a more rapid decline in cognition compared to older subjects, as reviewed by Sona et al. [4]. However, we found an opposite association for age with class membership where subjects in Class 1 with the steepest progression were older than in Class 3, though we did not control for time from disease onset or disease severity, and AD and HC groups were age-matched, while EMCI and LMCI groups were younger on average. This finding may be driven by the fact that amyloid portends poorer outcome, especially when combined with other pathologies that increase with aging. Our findings are consistent with a number, but not all, of other studies that report greater cognitive decline in subjects with worse baseline cognitive impairment [3, 20]. APOE4 positivity is associated with a greater decline in function [21] and cognition [22, 23], although this relationship has not been observed consistently [24–27]. Possible explanations have focused on linear versus non-linear statistical models [22], or different AD clinical populations [23]. Our study results are consistent with a number of studies showing that amyloid positivity is associated with worse baseline scores and disease progression [1, 28–33]. Additionally, a faster decline in cognition is reportedly associated with increased neuropsychiatric symptom severity including agitation and wandering [34], hallucinations and delusions [35], and depressive symptoms [36]. Our results showed a higher NPI-Q score was associated with faster progression, but the overall severity of neuropsychiatric symptoms was low in our sample.

We did not find class membership to be associated with gender, education, or comorbidities of smoking, cardiovascular disease, or endocrine disease, although there is other evidence that vascular risk factors are associated with greater functional and cognitive decline [26, 38]. Regan and colleagues [39] note though these risk factors may contribute to AD initially, they are not part of the underlying disease process. Alcohol abuse showed a significant difference among the trajectory classes in our analysis, but the percentage of patients reporting alcohol abuse was relatively low among the groups (range: 3.3% to 8%), making this finding less reliable. Those with higher cognitive reserve (more years of education) diagnosed with AD show an initial delay in symptom expression, followed by a more rapid rate of cognitive decline [4]. In the current study, education did not differ among the three classes, though most subjects in the ADNI cohort have a high level of education, limiting the ability to study this factor.

CART analysis can uncover complex interactions among predictors of trajectory class membership which could be difficult to identify using traditional multivariate techniques [11]. The result is a decision tree with a series of if-then scenarios to best allocate individuals to a particular class when applying the variable cutoff values that produced the best fit. This provides a straightforward approach to apply clinical predictors. The CART analysis identified two decision nodes, both based on FAQ results, to predict which subjects most likely belong to the GMM classes. While the ADAS-Cog₁₃ and FAQ results jointly defined the original GMM classes, we chose to exclude the ADAS-Cog₁₃ because it is utilized in research and not available to a clinician. However, we ran a separate CART analysis that included the ADAS-Cog and the FAQ remained the main differentiator. It is noteworthy that baseline cognitive variables or amyloid status were not better predictors of GMM membership. The FAQ is a functional measure that includes instrumental activities of daily living, such as preparing balanced meals and managing personal finances, and can be easily administered in a clinician office. Perhaps the FAQ cutoff values of 5 and 14, as identified through CART analysis, can facilitate identification of who will progress in other patient populations. These findings may extend the initial report suggesting a cut point of ≥9 (“dependent” or a score of 3 in three or more activities) to indicate impaired function consistent with dementia [40]. However, CART analyses are exploratory for generating hypotheses that need further testing.

There are some limitations to this study that should be taken into consideration. When determining amyloid status, we used CSF Aβ₄₂, PiB-PET, or FBP-PET results. However, 31 cases had discrepant amyloid results (20 of 31 discrepant cases were CSF positive but FBP-PET negative). Possible reasons for this include the fact that CSF measures soluble Aβ and PET measures insoluble deposited amyloid plaque. Additionally, there is variability across CSF assays and laboratories, with no standardized cutoff for amyloid positivity [41]. ADNI attempts to control for this variability by using a centralized laboratory for the CSF assays but CSF Aβ specimens must be handled carefully and there could be site variability in this handling. For our analyses, we used the latest ADNI cutoffs for CSF amyloid positivity [14]. Another possible cause for the amyloid test discordance could be false negative scan results, which have been observed with FBP-PET [42]. CART uses most likely class membership, and cannot accommodate the fact that a subject could be in another class with smaller probability. Although our sample size of 1,192 was relatively large, the follow-up period of up to two years (mean about 16 months) may not be long enough to see progression, especially earlier in the disease spectrum. In addition, the relatively low incidence of comorbidities, limited racial representation and high years of education in the ADNI population may limit generalizability of these results. We included HCs in the analyses as some are still at risk for AD and could be transitional toward impairment and indeed, some were amyloid positive (35%).

Our findings make a unique contribution to the literature on longitudinal change in older persons, including those clinically diagnosed with MCI and AD dementia and may help inform clinicians and patients when amyloid biomarker testing is not available. Moreover, patients with MCI were further subdivided into EMCI and LMCI to better capture how those categorical designations populated the three latent classes. Our sample was larger than most prior reports and the amyloidstatus of patients was known. Third, we looked at changes in both cognition and function. Lastly, we used CART analysis to develop a decision tree, and identified two decision nodes based on the FAQ that separated patients along the three latent classes. These decision nodes may further the understanding of the FAQ scale cutoff values in relation to severity of impairment.

In summary, the combination of cognitive and functional change over a 24-month period followed three distinct trajectories where positive APOE4 and amyloid status were more associated with a subpopulation that had greater decline over time (64% – 69% and 84% – 92% , respectively). The FAQ score may help to predict trajectory. This type of information could be shared by the physician to help patients and families plan for the future. However, our results are exploratory and should be confirmed in other populations with longer term follow-up.