Heterogeneous patterns of cognitive decline in Alzheimer’s disease across three domains of cognition

Abstract

Background:

Others have examined heterogeneity in Alzheimer’s disease; however, few have used longitudinal data while accounting for variation in disease stage. We used latent classes to model heterogeneity in the trajectories of three cognitive domains (memory, language, and executive functioning) starting at Alzheimer’s dementia diagnosis.

Objective:

Our aim was to describe the patterns of heterogeneity in cognitive decline across cognitive domains during the course of Alzheimer’s disease and to contextualize our findings by assessing associations with demographic factors and neuropathological measures.

Methods:

We used cognitive data from the Religious Orders Study, the Rush Memory and Aging Project, and the Minority Aging Research Study in a multi-dimensional joint latent class mixed model, which allowed us to estimate cognitive trajectories that varied across cognitive domains and latent classes. We accounted for the uncertainty in latent class assignment and corrected for multiple hypotheses when assessing the association of the latent classes with demographic and neuropathological variables.

Results:

We identified five latent classes differentiated by level of impairment (high to low) and rate of decline (slow to fast). Within each latent class, the pattern of decline did not differ substantially across cognitive domains. Classes were associated with APOE genotype, sex, race, education, and neuritic plaque and neurofibrillary tangle burden.

Conclusions:

Our results highlight global differences in the level of cognitive impairment at diagnosis and the rate of decline rather than differences between domains of cognition. Examination of patterns in the global rate of cognitive decline may improve understanding of heterogeneity in Alzheimer’s disease.

Introduction

Alzheimer’s dementia has been established as a heterogeneous disease in symptom and biomarker presentation.^1–6 It is unclear if heterogeneity occurs along a spectrum or if distinct subtypes of the disease exist. In either situation, subtyping is a useful approach to identify homogeneous groups or to simply condense complex information from multiple data modalities. Much of the literature has considered heterogeneity in neuropathological and imaging data and examined associations with cognition.^2–12 Only a few have used subtyping schemes based on cognitive data.^{1, 13–19} Cognition is the least invasive and most affordable Alzheimer’s dementia biomarker to collect. It may be useful for understanding heterogeneity in the natural history of the disease, and it is plausible that different subgroups would present with distinct patterns of cognitive decline. We sought to identify heterogeneous patterns of decline across three cognitive domains during the course of Alzheimer’s dementia.

Studies that incorporate measures of cognition generally use data from neuropsychological tests. It is common to use total scores, where neuropsychological items are kept in their original scale and summed, or to use sums or averages of z-scores. The weighting of items inherent in these methods does not reflect differences in measurement qualities, such as item difficulty, nor does it account for measurement error in the item responses.²⁰ Additionally, many studies consider cognition globally or as a single domain, which may obscure important differences between individual cognitive domains. To address these concerns, we followed others^{1, 14–17} and estimated scores for memory, language, and executive functioning using confirmatory factor analysis approaches. This allowed us to incorporate multiple indicators of each domain and to estimate item parameters that inform how each item reflects the respective latent trait.

A common characteristic of the subtyping literature is the use of cross-sectional data to determine subtypes. This runs the risk of finding distinct groupings that are an artifact of a particular time point. In a meaningful subgrouping, we expect individuals within a specific subgroup to be similar to one other at multiple time points. Some papers have applied a longitudinal subtyping scheme,^{5, 13} while others have used a quasi-longitudinal approach by performing cross-sectional analysis at multiple time points.^{6, 9, 18, 21} However, longitudinal subtyping approaches are still uncommon in the Alzheimer’s dementia literature, and more studies are needed.^{5, 10}

Another common feature in this literature is the use of the baseline study visit for subtyping in cross-sectional studies^{7, 9, 11} and the anchoring of time at the baseline visit in longitudinal studies.^{13, 18} Using a time point that is unrelated to disease stage, such as study visit or age, does not allow for a comparison of individuals at similar disease stages and gives the appearance of more extensive heterogeneity.^{5, 22} We followed others in anchoring time at Alzheimer’s dementia diagnosis.^{1, 5, 14–17} This may be a reasonable choice given that diagnosis is homogenously defined and subject to much scrutiny, often by consensus panels, in cohort studies of aging.

In this study, we used longitudinal measures of memory, language, and executive functioning in a joint multi-dimensional latent class mixed model²³ with time aligned at Alzheimer’s dementia diagnosis. This work addresses calls for more longitudinal studies of subtyping⁵ while anchoring at the time of diagnosis to limit heterogeneity in disease stage. Additionally, we were able to separately describe the trajectories of decline for each domain of cognition, using harmonized cognitive measures developed using modern psychometrics.^{20, 24} We examined associations of the identified latent classes with demographic characteristics and neuropathological measures. This paper, to our knowledge, is the first to longitudinally model heterogeneity in cognition during the course of Alzheimer’s dementia while aligning time at a point relevant to the disease stage.

Methods

The Religious Orders Study, Memory and Aging Project, and Minority Aging Research Study

Data are from the Religious Orders Study (ROS), the Rush University Memory and Aging Project (MAP), and the Minority Aging Research Study (MARS). All three are prospective cohort studies that enroll older adults without known dementia at enrollment. ROS enrolls nuns, priests, and brothers within the United States, MAP enrolls individuals living in northeastern Illinois, and MARS includes older African American individuals. These studies are described in detail elsewhere^{25, 26}. All three studies were approved by an Institutional Review Board of Rush University Medical Center, and all participants signed an informed consent, Anatomical Gift Act, and a repository consent to share data and biospecimens.

The studies administer nearly identical comprehensive neuropsychological assessments, including 21 tests at annual visits. Using the same set of items for each cohort, we chose to estimate scores for four domains of cognition: memory, language, executive functioning, and visuospatial abilities. These domains differ from the six domains constructed in the ROS, MAP, and MARS studies. The score for visuospatial abilities was based on only two items (the intersecting pentagons item from the Mini-Mental State Examination and judgment of line orientation). We excluded this domain from analyses as we did not feel it was estimated precisely enough to include in our longitudinal models.

Alzheimer’s disease (AD) dementia diagnosis was based on the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Disease and Related Disorders Association (NINCDS/ADRDA) criteria.²⁷ The diagnosis protocol has been described in detail elsewhere.^{28, 29} Briefly, a clinical diagnosis is made at each annual visit and is based on computer-scored cognitive testing that is reviewed, along with additional clinical information, by a neuropsychologist.

We included individuals where AD was determined to be the single (probable AD) or one of multiple (possible AD) causes of dementia. Those with an Alzheimer’s dementia diagnosis on their first visit were excluded, as the dementia conversion date is unknown. We required that individuals have cognitive scores for all three cognitive domains at the first Alzheimer’s dementia diagnosis visit. Only visits at and after the Alzheimer’s dementia diagnosis visit were included, and individuals with no visits after diagnosis were excluded (i.e., we required at least two visits). For a complete description of the sample selection, see Supplemental Figure 1. The final sample includes 785 individuals with 3,330 visits.

Neuropathological measures are from autopsy, and detailed methods have been described previously.^30–32 Briefly, after brain removal, each hemisphere was cut coronally into 1 cm slabs using an acrylic plastic jig. Tissues from one hemisphere were fixed in paraformaldehyde for at least 3 days after which they were dissected into diagnostic blocks. Neuritic plaques, diffuse plaques, and neurofibrillary tangles (NFT) were visualized for five regions (entorhinal cortex, hippocampus, midtemporal cortex, inferior parietal cortex, and midfrontal cortex) using a modified Bielschowsky silver stain. Counts were done by a board-certified neuropathologist or trained technician who was blinded to all clinical data. Brain donation is a requirement for enrollment in ROS and MAP and is optional in MARS. Here we used aggregate measures of NFT, neuritic plaques, and diffuse plaques calculated by dividing each of the five regional measures by their standard deviation and averaging. To make comparisons to one of the predominant subtyping schemes in the literature,^{2, 4} we also examined regional NFT measures from the five available regions.

Cognitive Scores

Confirmatory Factor Analysis (CFA) was used to estimate harmonized scores for memory, language, and executive functioning. The methods have been described in detail elsewhere.²⁴ Briefly, granular data on neuropsychological test items were obtained from four studies: the Alzheimer’s Disease Neuroimaging Initiative (ADNI), the Adult Changes in Thought (ACT) study, ROS, and MAP. A panel of experts, including two neuropsychologists and a behavioral neurologist, considered each item and categorized it as predominantly measuring memory, executive functioning, language, visuospatial abilities, or none of these. A separate cross-sectional CFA model was run for each domain. Estimated item parameters were used to extract scores for all individuals and visits. Additionally, item parameters estimated in this ‘legacy’ model were used to estimate domain scores for the MARS study that are on the same metric as the other studies. We prefer these scores to a sum or average of z-scores because they reduce measurement error and take measurement properties of the items, such as item difficulty, into consideration.²⁰

It should be noted that our data included floor scores. These are scores where the lowest possible points were attained on each administered item. In these cases, we know that the respondent performed worse than was measurable by the administered items, but we do not know precisely what their score should be. We were concerned that, if we included all visits with floor scores, we would exacerbate the underestimation of the slope since it would seem like cognition had stopped declining. We, therefore, chose to set all scores within a domain that occurred after the first-floor score to missing. There were 121 individuals with at least one memory score after a floor score resulting in 244 scores treated as missing. There were 170 visits (across 85 individuals) in language and 308 visits (across 152 individuals) in executive functioning with floor scores that we treated as missing.

Joint Longitudinal Latent Class Mixture Model

We used a joint longitudinal latent class mixed model from the lcmm package version 2.1.0 in R.²³ The model is laid out in more detail in Supplemental Section 2. The model can be thought of as a combination of three components: a multinomial logistic model for latent class membership, a separate linear mixed model in each latent class and for each domain of cognition, and a time-to-event model for the mortality in each latent class.³³

Each of the linear mixed models took the same general form, using time since Alzheimer’s dementia diagnosis as the timescale. The outcome variables were the composite scores for memory, language, or executive functioning. An intercept and fixed effects for time since diagnosis and age at diagnosis were included. Individual random intercepts and slopes with respect to time were incorporated to account for the serial correlation of repeated observations. We chose to model time linearly as this best fit our data (see Supplemental Section 3 and Supplemental Figures 2 and 3 for additional details on modeling decisions). We allowed the fixed intercepts and slopes, as well as the variances of the random intercepts and slopes, to vary by latent class. The fixed intercept and slope explained the class-specific average trajectory, while the random effects described the individual variation around those trajectories separately within each class.

Our model, described thus far, assumes that data are missing at random. In our case, this would mean that the probability of missing a particular score for an individual is explained by time, the trajectories of all the scores, and age at Alzheimer’s dementia diagnosis. It is plausible that individuals die for reasons not explained by these variables. Additionally, different subtypes of Alzheimer’s dementia may have different survival times. We modeled time to death as a survival process with time since Alzheimer’s dementia diagnosis as the time scale. We used a distinct proportional hazards model for each latent class with a Weibull baseline hazard and age at the time of Alzheimer’s dementia diagnosis as a covariate.

As we were interested in determining the presence of latent classes based solely on the trajectory of cognition and the survival process, we did not model latent class membership according to covariates. This means that the multinomial logistic portion of the model reduces to a simple probability of class membership.

For a given number of latent classes, the model was estimated via maximum likelihood using a modified Marquardt algorithm. We used ‘gridsearch’ from the lcmm package with 100 random departures and a maximum of 50 iterations to ensure that model solutions were not local maxima. One challenge of any latent class model is determining the optimal number of latent classes. We estimated models with 1 through 8 classes. Following recommendations for latent trajectory studies,³⁴ the optimal number of latent classes was determined according to statistical criteria assessing the goodness-of-fit with the Bayesian Information Criterion (BIC), the separation between classes with the entropy, and a balance of both with the Integrated Likelihood Criterion (ICL). After model estimation, the posterior probabilities of class membership for each individual and each class were extracted. Class assignment was determined based on the class with the highest posterior probability for each individual, referred to as the modal class assignment.

Assessing the Association of Latent Classes with Predictors and Distal Outcomes

It is important to assess the association of the latent classes with external variables to better understand the findings. However, directly assessing the association with modal class assignment is not recommended as this does not account for uncertainty in classification and can bias parameter estimates.³⁵ Methodologically and conceptually, external variables of interest are split into two groups: those seen as predictors of the latent classes and those seen as outcomes (often called distal outcomes) of the latent classes. Many options for estimating associations with both sets of external variables have been proposed.^36–38 We follow the approach taken by Proust-Lima et al.,³³ which is available in the lcmm package.

The predictors investigated here were selected to align with common demographic factors discussed in the subtyping literature, as well as measures aimed at understanding potential comorbidities that could affect cognition. We included self-reported sex and race, presence of an APOE ε2 or ε4 allele, years of education, self-reported history of hypertension up to the time of Alzheimer’s dementia diagnosis, history of diabetes from self-report or medications up to the time of Alzheimer’s dementia diagnosis, and an index for history of vascular disease burden up to Alzheimer’s dementia diagnosis, which incorporates self-reported measures of claudication, stroke, and heart conditions. All pairwise latent class comparisons were conducted for each measure and the Benjamini-Hochberg false discovery rate (FDR) was used to correct for multiple hypothesis testing.³⁹

We assessed the association of latent classes with aggregate NFT, diffuse plaque, and neuritic plaque burden. Together, these can be considered a measure of the global AD pathology burden.²⁹ To compare our results with previously identified subtypes, we also modeled regional NFT burden from the inferior parietal cortex, entorhinal cortex, midfrontal cortex, midtemporal cortex, and hippocampus. All pairwise comparisons were corrected for the FDR. These analyses were conducted in the subsample of participants with autopsy data.

Results

Model Selection and Class Descriptions

The final sample included 785 individuals across 3,330 visits. The sample included more women than men (25% men), and a majority of participants self-reported as White (82%). The average number of visits per person was 4.2, with an average follow-up time of 3.6 years. A description of the sample is provided in Table 1.

Table 1.

Sample Summary Statistics. Values are mean (standard deviation) or n (%). All statistics summarize individual-level measures except for ‘Visits with floor scores’, which gives the total number of visits (and percent out of all visits) that were at the floor.

	Full Model Sample	ROS	MAP	MARS
Age at Diagnosis	86.5 (6.6)	85.9 (6.3)	87.8 (6.4)	83.5 (6.7)
Male	195 (24.8%)	94 (26.4%)	81 (23.5%)	20 (23.8%)
Any APOE ε2 allele	82 (11%)	40 (11.7%)	32 (9.9%)	10 (12.5%)
Any APOE ε4 allele	263 (35.3%)	116 (33.9%)	114 (35.2%)	33 (41.2%)
Years of education	16.2 (3.7)	17.9 (3.4)	14.9 (3.2)	14.5 (3.4)
Self-reported race:
White	645 (82.2%)	319 (89.6%)	326 (94.5%)	0 (0%)
Black / African American	137 (17.5%)	35 (9.8%)	18 (5.2%)	84 (100%)
Other and unknown	3 (0.4%)	2 (0.6%)	1 (0.3%)	0 (0%)
Died	681 (86.9%)	334 (94.1%)	289 (83.8%)	58 (69%)
Age at death	91.6 (5.8)	90.9 (5.8)	92.9 (5.7)	89.1 (5.1)
Visits per person	4.2 (2.5)	4.6 (2.8)	4.1 (2.2)	3.3 (1.7)
Years of follow-up	3.6 (2.7)	3.9 (3)	3.4 (2.4)	2.8 (2.2)
Score at Diagnosis:
Memory	−0.9 (0.5)	−0.8 (0.5)	−0.9 (0.4)	−1 (0.5)
Language	−0.7 (0.5)	−0.7 (0.5)	−0.7 (0.5)	−0.9 (0.6)
Executive Function	−0.5 (0.5)	−0.4 (0.5)	−0.5 (0.5)	−0.7 (0.6)
Visits with floor scores:
Memory	204 (6.1%)	103 (6.3%)	87 (6.2%)	14 (5%)
Language	142 (4.3%)	82 (5%)	50 (3.5%)	10 (3.6%)
Executive Function	275 (8.3%)	146 (8.9%)	109 (7.7%)	20 (7.1%)
N visits	3330	1641	1409	280
N individuals	785	356	345	84

The joint longitudinal latent class mixed model was run for 1 through 8 classes. As shown in Figure 1, the statistical criteria were quite similar for the 5, 6, and 7, class models. The ICL shows a distinct leveling off at 5-classes, with the 7-class model being marginally better (lower) compared to the 5- and 6-class models. Of these three models, entropy was best (highest) for the 5-class model, while the BIC was best (lowest) for the 7-class model. To avoid overfitting the data, we decided to interpret the simplest of the three models in our main analysis and moved forward with the 5-class model. However, we checked that the relative patterns of decline across classes for the 7-class model were similar to those of the 5-class model (Supplemental Figure 4). Supplemental Figure 5 shows the evolution of the individual, modal class assignment across the 8 models.

Figure 1.

Fit Statistics for the 1 through 8 class models. Bayesian Information Criteria (BIC); Integrated Completed Likelihood Criterion (ICL). The BIC and ICL scales are model-dependent. A lower value is considered better fit. BIC measures fit based on the value of the likelihood function with penalties for the number of parameters estimated and sample size. Entropy is based on the posterior probabilities and ranges from 0 to 1 with a value closer to 1 representing more well separated classes (by definition a one class model has an entropy of 1). The ICL is a function of both the BIC and the posterior probabilities and aims to strike a compromise between model fit and class separation.

The trajectories of the three longitudinal processes for the 5-class model are shown in Figure 2 and parameter estimates are shown in Supplemental Table 1. The top panel of Figure 2 shows the individual raw score trajectories grouped by modal class and the middle panel shows the individual-specific estimated trajectories. The final panel shows the average class trajectories for someone with an age at diagnosis of 85, which is near the median of the distribution. Since the effect of age at diagnosis was allowed to vary by latent class, the relation between the latent classes could have differed across diagnosis ages, however, we did not find that to be the case (Supplemental Figure 6). Supplemental Figure 7 visually summarizes the fit of the longitudinal processes to the data. For clarity, we named the classes based on the pattern of decline, which was largely consistent across cognitive domains. We differentiated two groups, which we named ‘Slow decline, high’ and ‘Slow decline, low’, from the other groups based on their slower rate of decline. Those in the ‘Slow decline, high’ group had a higher level of memory, language, and executive functioning at Alzheimer’s dementia diagnosis compared to those in the ‘Slow decline, low’ group. Similarly, the ‘Moderate decline, high’ and ‘Moderate’ groups had a faster rate of decline compared to the two slow decline groups and were differentiated from each other by the level of cognition across the three domains at diagnosis. The final group, which we called ‘Fast decline’, had a low to intermediate intercept compared to the other groups and declined the most rapidly.

Figure 2.

Raw and predicted trajectory of the three longitudinal processes in the 5-class model. Colors represent modal class assignment. The top panel shows raw score trajectories. The middle panel shows subject-specific predicted trajectories. The bottom panel shows the average class predicted trajectories for an age at diagnosis of 85 with shaded regions showing the 2.5 and 97.5 percentiles of the Monte Carlo approximation of the posterior distribution. Proportions of the sample that fall into each class using modal assignment are given in parentheses.

Generally, the pattern of decline within a class was similar across domains of cognition; the relative level of impairment and rate of decline in one cognitive domain was mirrored by that in each of the other domains. There were two exceptions to this general pattern. First, the ‘Moderate decline, high’ class had the highest scores in language and executive functioning at diagnosis compared to the other classes, but it had lower memory scores at diagnosis relative to the ‘Slow decline, high’ class. It is difficult to determine whether these relative differences in intercepts would be statistically or clinically meaningful. The second exception to the general pattern is the trajectory of decline in the ‘Slow decline, low’ class. There is a slow decline in memory and language for this class, but the decline (slope) is not significantly different from zero in executive functioning. Again, it is difficult to know if this divergence from the general pattern is meaningful or a result of measurement imprecision. Note that while the cognitive domain scores have similar ranges, the scale for each domain is arbitrary and direct comparisons of specific values across domains of cognition should not be interpreted.

The latent classes were well separated (Supplemental Table 2). The average posterior probability within each modally assigned class was 0.83 or higher. Over 70 percent of individuals in each modally assigned class had a posterior probability of belonging to that class that was 0.7 or higher. Based on modal class assignment, the ‘Moderate’ class was the largest, making up 42 percent of the sample. The ‘Slow decline, high’ and ‘Moderate decline, high’ classes were the next largest with 27 and 16 percent, respectively. The ‘Fast decline’ class (8 percent) and the ‘Slow decline, low’ classes (8 percent) were the smallest.

Figure 3 shows the cumulative incidence of death for each of the latent classes for those with an Alzheimer’s dementia diagnosis at age 85. Supplemental Figure 8 compares the cumulative incidence for five different diagnosis ages. The pattern shown in Figure 3 is apparent for those with a diagnosis age of 80 or older, which includes 90 percent (n=655) of the sample. The ‘Fast decline’ class had the highest incidence of death and consequently the shortest average follow-up time (2.6 visits over 1.7 years). The ‘Moderate’ and the two slow decline classes had similar, intermediate levels of incidence and average follow-up time (‘Moderate’ 3.7 visits over 3 years; ‘Slow decline, high’ 4.5 visits over 3.8 years; ‘Slow decline, low’ 3.9 visits over 3.2 years). The ‘Moderate decline, high’ class had the slowest increasing cumulative incidence and the longest average follow-up time (6.1 visits over 5.8 years).

Figure 3.

Cumulative Incidence of death for the 5-class model. Cumulative incidence of death assuming an Alzheimer’s dementia diagnosis at age 85. Lines show point estimates for the cumulative incidence by modally assigned class. Bands show 2.5 and 97.5 percentiles from the Monte Carlo approximation of the posterior distribution.

Association of latent classes with demographic predictors

We evaluated the association of variables external to the latent class model. Figure 4, Panel A shows the associations with hypothesized predictors of the classes while accounting for the uncertainty in classification. ‘Moderate decline, high’ was arbitrarily chosen as the reference class and is represented by the line at zero. Supplemental Table 3 shows results for all pairwise comparisons along with q-values corrected for the FDR.

Figure 4.

Association of latent classes with external variables. In all panels ‘Moderate decline, high’ is the reference class represented by the line at 0. Error bars are 95% confidence bands for the test of mean difference between each class and ‘Moderate decline, high’. (A) Comparisons are expressed as log odds ratios relative to the reference class. Some standard errors were estimated with considerable error due to small cell sizes. We recommend interpreting only the direction of association rather than the coefficient values. For example, confidence bands below zero suggest evidence for a reduced likelihood (for categorical variables) or quantity (for numeric variables) of the trait/characteristic in individuals categorized into the respective latent class relative to ‘Moderate decline, high’. The log odds ratio for the presence of an APOE ε2 allele in the “Slow decline, low” class (grey) is −8.93 and is thus not shown here. This comparison was particularly sparse as shown by the large standard error and we consider it to be non-interpretable. (B) Comparisons represent average burden measure deviations from the reference class. All burden measures are averages of standardized region measures. (C) Units are NFT counts relative to the reference class. Models in panels B and C control for time since first diagnosis, and points may be interpreted as coefficients in a regression with the specified measure as the outcome and the ‘Moderate decline, high’ class as the excluded category.

Individuals in the ‘Slow decline, high’ class were less likely to have an APOE ε4 allele compared to other classes (‘Moderate decline, high’ and ‘Moderate’ p<0.01, q<=0.01; ‘Slow decline, low’ p<0.01, q=0.02) except the ‘Fast decline’ class where the difference was not significant (p=0.05, q=0.18). The ‘Slow decline, high’ class also included significantly more individuals self-reporting as Black / African American compared to the ‘Moderate decline, high’ (p<0.01, q=0.01) and ‘Moderate’ (p<0.01, q<0.01) classes. Comparisons with ‘Fast decline’ (p=0.01, q=0.08) and ‘Slow decline, low’ (p=0.02, q=0.09) were in the same direction, but not significant.

Individuals in the ‘Slow decline, low’ class had fewer years of education on average compared to ‘Moderate decline, high’, ‘Moderate’, and ‘Slow decline, high’ (p<0.01, q<=0.01). The comparison with ‘Fast decline’ was in the same direction, but not significant (p=0.01, q=0.06).

The ‘Moderate decline, high’ class had a smaller proportion of self-reported males compared to other classes (‘Moderate’ p=<0.01, q=0.03; ‘Fast decline’ p=<0.01, q=0.02; ‘Slow decline, high’ p<0.01, q<0.01), however the comparison to ‘Slow decline, low’ (p=0.02, q=0.09) was not significant.

Association of latent classes with neuropathological distal outcomes

We examined aggregate measures of diffuse plaques, neuritic plaques, and NFT burden at autopsy using methods that accounted for the uncertainty in latent class assignment. The measures were averages of standardized counts from five regions and were available for 578 individuals who had died and gone to autopsy (Supplemental Table 4). All models were adjusted for the time between diagnosis and autopsy. Figure 4, panel B shows the results with the ‘Moderate decline, high’ class as the reference. All pairwise comparisons are shown in Supplemental Table 5.

Neuritic plaque burden was lower in the ‘Slow decline, high’ class compared to other classes (‘Slow decline, low’, ‘Moderate’, ‘Fast decline’ p<0.01, q<=0.01), however, the difference was not significant compared to ‘Moderate decline, high’ (p=0.03, q=0.08). Similarly, ‘Slow decline, high’ also had a lower average NFT burden compared to ‘Slow decline, low’, ‘Moderate’, and ‘Fast decline’ (p<0.01, q<=0.01), but not ‘Moderate decline, high’ (p=0.15, q=0.26). The ‘Moderate decline, high’ class had a lower neuritic plaque burden compared to ‘Moderate’ (p=0.01, q=0.02) and ‘Fast decline’ (p=0.01, q=0.03), and a lower average NFT burden compared to ‘Moderate’ (p<0.01, q<0.01) and ‘Slow decline, low’ (p<0.01, q=0.01).

We were also interested in the regional variation of NFT burden across the latent classes and examined five regional measures of tangle counts. Figure 4, panel C and Supplemental Table 6 show these results.

Compared to the ‘Moderate’, ‘Slow decline, low’, and ‘Fast decline’ classes, the ‘Slow decline, high’ class had on average less NFT burden in the inferior parietal (‘Slow decline, low’ and ‘Moderate’ p<0.01, q<0.01; ‘Fast decline’ p=0.01, q=0.03), entorhinal (‘Slow decline, low’ p=0.01, q=0.03; ‘Moderate’ p=0.01, q=0.02; ‘Fast decline’ p=0.01, q=0.04), midfrontal (p<0.01, q<0.01 for all), and midtemporal (p<0.01, q<=0.01 for all) cortices. The ‘Moderate decline, high’ class had lower NFT burden in the inferior parietal cortex compared to ‘Moderate’ and ‘Slow decline, low’ (p<0.01, q<0.01 for both) and in the midfrontal cortex compared to ‘Moderate’ (p<0.01, q<0.01), ‘Slow decline, low’ (p<0.01, q<0.01), and ‘Fast decline’ (p=0.01, q=0.02).

Sensitivity Analyses

Overall, the ‘Slow decline, high’ class had the lowest AD pathology burden. One potential explanation is that many of these individuals had dementia caused by non-AD pathology or by a mix of AD and other pathologies. To investigate this, we ran two sensitivity analyses. First, we restricted to individuals who underwent an autopsy and had sufficient data to be assessed for pathologically confirmed AD (n=573). Within this sample, we looked for associations between pathologically confirmed AD cases (N=468), defined as those with an intermediate or high likelihood of AD following the NIA-Reagan criteria,⁴⁰ and the latent classes. Supplemental Table 7 shows the number of individuals with no or low likelihood of AD following NIA-Reagan criteria in each modally assigned class without correction for uncertainty in class assignment or the FDR. Supplemental Table 8 incorporates these adjustments and shows that the ‘Slow decline, high’ class included significantly fewer pathologically confirmed AD cases compared to all classes except ‘Moderate decline, high’. Supplemental Table 9 shows that there are still many individuals modally assigned to the ‘Slow decline, high’ class (22% compared to 27% in the full sample). Additionally, the average entropy, which can be thought of as a level of confidence in individual class assignment, was no different for the ‘Slow decline, high’ class in the autopsy sample compared to the full sample. In fact, all classes had the same or higher average entropy in the autopsy sample. This suggests that the patterns observed in the ‘Slow decline, high’ class, as well as in the other classes, were likely not driven by those with no or little AD pathology, even though the ‘Slow decline, high’ class included more individuals without AD as defined by the criteria.

For the second sensitivity analysis, we re-ran all association analyses in the sample of individuals with a high or intermediate likelihood of AD as indicated by the NIA-Reagan criteria. The model with all predictors did not converge. However, the models excluding either the variable for Black/African American, APOE ε4 allele status, hypertension, diabetes, or vascular disease burden did converge. We chose to interpret the results excluding diabetes as no associations were found with this variable in the main results. The interpretation of results would not be different if we instead examined one of the other models that converged. The pattern of associations between the predictors and latent classes were similar (Figure 5, panel A), but with less precision than in the full sample (Supplemental Table 10). The one exception was that there were too few Black/African American individuals in this sample (N=36) to interpret the associations. The patterns of associations with the average and regional neuropathological variables were also largely unchanged (Figure 5, panels B and C), albeit with some loss in precision (Supplemental Tables 11 and 12).

Figure 5.

Sensitivity analyses - Association of latent classes with external variables in the pathologically confirmed AD sample. The sample for all models was restricted to those with pathologically confirmed AD following NIA-Reagan criteria. In all panels ‘Moderate decline, high’ is the reference class represented by the line at 0. Error bars are 95% confidence bands for the test of mean difference between each class and ‘Moderate decline, high’. (A) Comparisons are log odds ratios (with 95% confidence bands) in relation to the reference class. Some standard errors were estimated with considerable error due to small cell sizes. We recommend interpreting only the direction of association rather than coefficient values. (B) Comparisons represent average burden measure deviations from the reference class. All burden measures are averages of standardized region measures. (C) Units are NFT counts relative to the reference class. Models in panels B and C control for time since first diagnosis, points may be interpreted as coefficients in a regression with the specified measure as the outcome and the ‘Moderate decline, high’ class as the excluded category.

We were also concerned about potential confounding of the results by cohort of origin. We reran all association analyses in the full sample and adjusted for cohort. These results are shown in Supplemental Figure 9 and Supplemental Tables 13, 14, and 15. With the exception of one comparison becoming insignificant (fewer Black/African American individuals in ‘Slow decline, high’ relative to ‘Moderate decline, high’ (p=0.01, q=0.06)), none of our main results changed. Additionally, the comparisons of latent classes with cohorts were not significant when correcting for the hypothesized predictors (Supplemental Table 13).

Discussion

In this paper, we longitudinally model heterogeneity in cognition among individuals with Alzheimer’s dementia while aligning time at dementia diagnosis. Using a joint longitudinal latent class mixed model, we identified five latent classes and found significant associations with APOE ε4 prevalence, self-reported sex and race, years of education, aggregate neuritic plaque burden, and aggregate and regional NFT burden.

Much of the literature on AD subgroups focuses on relative differences either between domains of cognition, regional patterns of atrophy, or regional patterns of pathology accumulation while controlling for the global level. We designed a model with the expectation of finding differences across cognitive domains. For example, our model could have identified a class with the lowest level of memory impairment and/or fastest rate of decline compared to other classes, while also having the highest level and/or slowest rate of decline in language and executive functioning. Instead, our results suggested that the pattern of cognitive decline in any particular cognitive domain was shared across the other domains. There were two exceptions to this broad statement. The ‘Moderate decline, high’ class had, on average, worse memory, but better language and executive functioning at diagnosis compared to the ‘Slow decline, high’ class. The ‘Slow decline, low’ class did not have a statistically significant decline in executive functioning, which was the case for all other classes and domains. These results were weak, and we are hesitant to interpret them as meaningful. Instead, our results indicate consistency in the pattern of decline across cognitive domains.

The heterogeneity we observed in cognitive trajectory likely reflects underlying variation in neuropathological and genetic processes as well as variation in life course exposures. Individuals in the slow-decline, high class, who demonstrated lower neuritic plaque and tangle burden and included fewer APOE ε4 carriers, may represent a group that’s more resilient to the accumulation of amyloid and tau pathology. Additionally, while our results do not seem to be driven by those without AD, this group was less likely to include individuals with pathologically confirmed AD suggesting a larger possible role of co-pathologies. This group had an intermediate risk of death compared to other classes despite their relatively preserved cognition throughout the disease course. This group also included more black/African American participants. Interestingly, increased African ancestry has been associated with a higher prevalence of both APOE ε4 and ε2 alleles as well as an attenuated relationship between APOE ε4 and risk of dementia.^{41, 42} We identified one article that found a slower rate of cognitive decline and a lower level of global cognition at baseline in African Americans compared to non-African Americans⁴³. However, more research is needed to understand how APOE ε4, genetic ancestry, and health disparities experienced by racialized groups interact to influence Alzheimer’s dementia presentation.

While only a few comparisons were statistically significant, the ‘Fast decline’, ‘Moderate’, and ‘Slow decline, low’ classes, tended to have higher average neuritic plaque, average NFT, and regional NFT burden in all but the hippocampus and entorhinal cortex compared to the other two classes. The faster loss of function in ‘Fast decline’ and ‘Moderate’ classes, may correspond to more aggressive Alzheimer’s disease biology marked by interacting amyloid–tau pathology. The association with lower education in the ‘Slow decline, low’ class suggests that reserve-related mechanisms⁴⁴ may modulate how neuropathology translates into clinical decline.

From a translational perspective, these findings suggest the potential benefits of trajectory-based cognitive subtypes for linking clinical presentation to underlying disease mechanisms. If these results are replicated in other longitudinal cohorts and their biological relevance is established, subtypes could be used to refine prognosis so that patients and caregivers can better plan for the future and could eventually be used to inform stratification in clinical trials, and ultimately guide targeted interventions aimed at slowing decline in specific subgroups.

The present work should, however, be situated within the large existing literature on subtypes of Alzheimer’s dementia. This literature can be organized according to the type of method used (theory versus data-driven) and the type of data used (cross-sectional versus longitudinal). The most widely applied classification scheme, originally proposed by Murray et al.⁴, uses regional NFT data from autopsy and a theory-based subtyping algorithm to classify individuals into hippocampal sparing (HS), limbic predominant (LP), and Typical (TAD) groups. These subtypes have been associated with regional volumetric measures such as the ratio of hippocampal to cortical volume¹² as well as the level of cognitive impairment at symptom onset and the rate of cognitive decline.⁴⁵ The classification algorithm has also been applied to cross-sectional data from MRI¹¹ and tau measured with Positron Emission Tomography (tau-PET).⁸ As research on these subtypes has progressed a fourth, minimal atrophy (MA) subtype has been added.^{2, 7} Some similarities have been found between this and a second theory-based, cross-sectional subtyping method that uses the relative cognitive domain impairment at dementia diagnosis to categorize individuals¹. Crane et al.¹⁴ found a concordance between those with a relative memory impairment and the LP group, as defined using volumetric measures from MRI.

More recently, data-driven subtyping methods have become more common, and some have utilized longitudinal data. Poulakis et al.⁵ used a longitudinal Bayesian clustering method and identified five subtypes using volumetric MRI data that seem to correspond to the HS, MA, and two versions of the LP class, and additionally identified a diffuse atrophy class. Vogel et al.⁶ used a machine learning algorithm applied to cross-sectional tau-PET data and found limbic predominant and medial-temporal lobe sparing phenotypes as well as subtypes characterized by predominant posterior and left-lateralized temporal lobe patterns. A few data-driven methods have used cognitive data for subtyping^{13, 18, 19}, however, only Geifman et al.¹³ developed groups based on longitudinal measures of cognition. They looked at global cognition using the ADAS-Cog and identified three groups differentiated by level of impairment and rate of decline. However, time was aligned at study baseline so heterogeneity may reflect differences in disease stage rather than disease presentation.

Overall, the existing subtyping literature suggests the existence of limbic predominant, hippocampal sparing, and typical groups with some evidence for a minimal atrophy group, a group with left-lateralization of temporal pathology and atrophy^{6, 15}, and a posterior predominant group^{6, 15}. There has been some convergence around the characteristics of these groups. However, Mohanty et al.¹⁰ found that, while group characterizations seemed similar across subtyping schemes, there was large disagreement in terms of individual subtype assignment.

A comparison of our findings with the rates of decline and patterns of association found in the previous literature suggests that longitudinally modeling cognition revealed different patterns of heterogeneity compared to those identified using data from neuroimaging or neuropathology and cross-sectional studies using cognitive data. For example, although a minimal pathology group has not been described, the lower pathology burden in the ‘Slow decline, high’ group suggests a possible concordance with the minimal atrophy group. However, our finding of fewer APOE ε4 carriers in this group has been considered a characteristic of the hippocampal sparing group.^{2, 11} Our methods are most similar to those of Geifman et al.,¹³ who also found heterogeneity in the rate of cognitive decline, and our results share some similarities with Uretsky et al.⁴⁵ who found variation in the overall rate of decline, but no relative cognitive domain impairments between the neuropathologically defined HS, LP, and TAD groups.

A few limitations of the present work should be noted. First, the ROS sample is highly selected, highly educated, and predominantly white. Additionally, the ROS and MAP samples required brain donation, while MARS did not. Given that none of the three cohorts included in our sample are meant to be representative of the general elderly population, it is unlikely that the relative sizes of the latent classes would be found in replication studies. Some of our cognitive scores were at the floor of what was measured in the neuropsychological batteries. We could have modeled this using a link function, but preliminary analyses indicated that no link was needed. We were still, however, concerned about these floor scores and treated scores that occurred after a floor as missing; therefore, we are unable to observe scores that fluctuated back above the floor. Of the excluded scores, 18 percent (45) from memory, 26 percent (44) from language, and 17 percent (52) from executive functioning fluctuated back above the floor. Two of our classes were proportionally small, each making up less than 10 percent of the sample. Our sample was not large enough to consider interactions between latent class predictors, notably self-reported race and APOE allele status, when assessing the associations with latent classes. We used dementia diagnosis to align individuals at similar disease stages. This assumes that the annual visits were frequent enough for timely diagnosis and that diagnosis represents a comparable disease stage across subtypes. Latent class analysis is known to find classes in homogenous but non-normal distributions.⁴⁶ Our classes were well separated; however, it is still possible that the classes do not represent truly distinct groupings and are instead categorizations of a continuous distribution of heterogeneity. Finally, it is possible that some classes may reflect the effects of differing underlying pathological processes. Our sensitivity analyses indicate that it is unlikely that the pattern of any one class is driven by individuals with no or little AD pathology. However, the latent classes may reflect different combinations of AD and other pathologies. Future work should assess if this is the case.

The main strengths of this work are that we employed longitudinal modeling and anchored time at dementia diagnosis so that individuals were at similar disease stages. The joint aspect of our model accounted for survival bias after diagnosis. We used methods to account for uncertainty in latent class assignment and adjusted for multiple hypothesis testing. Finally, we modeled three domains of cognition rather than looking at a global measure.

Our results suggest that heterogeneity in the global pattern of cognitive decline may be relevant for gaining a deeper understanding of heterogeneity in Alzheimer’s dementia. However, these findings should be considered in the context of the existing literature on relative differences in regional patterns of biomarker accumulation and cognitive impairment.

Data Availability Statement

The data supporting the findings of this study are available on request from Rush Alzheimer’s Disease Center’s Research Resource Sharing Hub. The authors are not able to directly provide the data due to privacy and ethical restrictions.