Cognitive Variability Predicts Incident Alzheimer’s Disease and Mild Cognitive Impairment Comparable to a Cerebrospinal Fluid Biomarker

INTRODUCTION

As the prevalence of Alzheimer’s disease (AD) rises, strategies to address the disease have shifted toward prevention, targeting the disease when it is active but pre-symptomatic [1]. Effectively monitoring AD in preclinical phases requires the use of disease biomarkers. Among the most promising biomarkers are cerebrospinal fluid (CSF) levels of phosphorylated tau protein (p-tau₁₈₁), amyloid-β 42 (Aβ₄₂) peptide, or the ratio of these analytes (e.g., p-tau₁₈₁/Aβ₄₂). Although CSF analytes are considered among the most direct measures of the disease’s hallmark neuropathological features [2–4], collection of CSF biomarkers is invasive and aversive to many potential research participants, driving a need to develop non-invasive, convenient markers[5, 6].

We recently demonstrated a relationship between a potential cognitive marker of preclinical disease, intra-individual cognitive variability (IICV), and another robust neuroimaging biomarker, longitudinal hippocampal volume loss (HVL) [7]. The proposed cognitive marker represents the within-person standard deviation across cognitive tests completed during a single session. This approached used by Holtzer et al. [8] is grounded on the principle put forward by many others [9–16] that a high level of within-person cognitive variability is a marker of brain pathology.

Like Holtzer et al. [8], our approach to estimating IICV emphasizes detection of differences in cognitive performance across domains measured in a single assessment. We selected indices which could be expected to diverge early in dementia due to AD. For example, performing poorly on executivefunctioning tasks (e.g., Trail Making Test B) in the context of a strong performance on a simple attentional task (e.g., Trail Making Test A) would suggest that an individual is exhibiting executive dysfunction unrelated to simple processing speed and visual scanning [17]. A disparity in performance between the tests would identify someone with very early executive dysfunction yet well-preserved sensory motor abilities, as is typically observed in early AD.

Adding to other investigations of IICV and risk for dementia [8–11, 14], we found that IICV predicted incident AD and mild cognitive impairment (MCI) similar to HVL and Apolipoprotein E (APOE) ɛ4 status [7]. Similarly, using data from the Wisconsin Registry for Alzheimer’s Prevention (WRAP), we noted IICV measured in a younger, i.e., middle-aged cohort, predicted incident cognitive decline occurring approximately a decade later [18].

In the present analysis of Alzheimer’s Disease Neuroimaging Initiative [19–21] (ADNI) data, we contrasted associations of IICV and p-tau₁₈₁/Aβ₄₂ to incident AD and MCI. Consistent with our overall goal to examine the utility of a non-invasive, easily implemented alternative strategy to traditional biomarkers, we hypothesized that IICV would compare favorably to an established biomarker of underlying AD pathology, i.e., p-tau₁₈₁/Aβ₄₂. Secondly, using a larger sample of ADNI participants, not limited to those with CSF analyte data, we examined whether individual IICV predicted risk of conversion over and above the individual components of IICV, hypothesizing that IICV would survive a stepwise backward selection analysis, comparing the summary index to its individual components.

MATERIALS AND METHODS

Participants

Participants were enrolled at ADNI sites in the United States and Canada during three ADNI funding cycles (ADNI-1, ADNI-2, and ADNI-GO) [19–21]. All subjects were between age 55 and 91 at baseline; English or Spanish language speakers; non-depressed; and in one of three diagnostic categories: early AD, MCI, or cognitively healthy. Cognitive status was confirmed using cut-off scores from the Clinical Dementia Rating Scale, Mini-Mental State Examination (MMSE), and Wechsler Memory Scale Logical Memory II. Complete ADNI exclusion criteria are found at http://www.adni-info.org [22]. For the first 24 months, evaluations occurred every 6 months, and every 12 months thereafter, with a mean total follow-up time of 78.2 months (SD = 29.7). Results of cognitive assessments, physical examinations, and MRI scans were considered in determining diagnostic status.

Figure 1 depicts the derivation of the analytic sample. The initial subject pool included 1,729 participants with visits up to 10/15/2016. We excluded subjects with fewer than 2 visits, missing demographic data, or baseline diagnosis of AD. After exclusionary criteria were applied, 1,307 participants remained in the sample. An analysis comparing the contribution of the components of the IICV to the index itself was performed using this sample. Only a limited number of subjects had CSF data available, resulting in a CSF analytic subsample of 349.

OPEN IN VIEWER

Fig.1

CONSORT Diagram –Sample derivation for logistic regression model including subjects with complete longitudinal and APOE ɛ4 and CSF data. Analytic model included baseline age, years of education, APOE ɛ4 status, years of follow-up, and baseline diagnosis. ^aThe distributions of cognitive test scores from 1,324 participants were used to derive IICV estimates. ^bOne subject died during length of follow-up. The subject was included; brain autopsy was used for the final diagnosis instead of clinical examination.

Participants provided written, informed consent per ADNI protocols; study sites obtained approval by the local institutional review boards, ensuring that procedures involving experiments on human subjects were done in accord with the ethical standards.

RESULTS

Participants were average age 72.4 years (SD= 7.09) at baseline, and well-educated (mean = 16.2 years). Table 1 lists participant characteristics by baseline diagnosis. Most participants were white, n = 325 (93.1%), and non-Hispanic n = 341 (98.8%). Compared to the larger pool of ADNI Participants included in the CSF analytic subsample were slightly younger than individuals included in the full IICV analytic sample, and less impaired on the Rey AVLT with lower IICV scores (data not included).

Mean length of follow-up was 2.32 years (SD = 0.682). In total, 42 (12.0%) of the 349 subjects who contributed CSF converted to AD and 9 (2.6%) converted to MCI (Table 1). Converters differed from non-converters in an expected pattern. They were older, performed more poorly on cognitive measures, and demonstrated lower levels of CSF Aβ₄₂ and higher levels of CSF tau₁₈₁. Results from logistic regression models suggest that an increase in IICV elevated an individual’s odds of incident AD and MCI. Figure 2 depicts the positive association of IICV with the proportion of participants who converted diagnostically. The association occurred across the range of IICV values. Regression models in Table 2 indicate an association of CSF with conversion odds in the full sample (n = 349) and when the sample is restricted to individuals with a baseline diagnosis of MCI (n = 244). Specifically, inclusion of IICV alone resulted in ORs between 2.39 (95% CIs 1.60 to 3.56) and 2.77 (95% CIs 1.72 to 4.46), depending on whether cognitively healthy adults were included in the analytic sample.

OPEN IN VIEWER

Fig.2

IICV versus Conversion status with local smoothing trend (LOESS). Data are jittered around conversion status to better illustrate data densities. Panel A includes participants with MCI and those who were cognitively healthy at baseline (n = 349). Panel B includes only those individuals with MCI at baseline (n = 244). Notes: IICV, intra-individual cognitive variability. Used raw IICV score for comparison. Conversion is a dichotomous variable: Converters versus Non-converters. The trend line is smoothed using LOESS in order to illustrate the relationship between conversion and IICV score without linearity assumptions. Additionally, IICV scores are jittered to more clearly see the number of participants obtaining individual IICV scores.

Similarly, estimated OR’s of conversion for 1 standard deviation increase in CSF ratio p-tau₁₈₁/Aβ₄₂ are 2.13 to 2.16, depending upon the subsample (95% CIs span 1.49 to 3.13). Figure 3 depicts the association of the ratio of standardized CSF p-tau₁₈₁/Aβ₄₂ and conversion. Like analysis of IICV and conversion, elevated ratio of p-tau₁₈₁/Aβ₄₂ corresponded to elevations in diagnostic conversion.

OPEN IN VIEWER

Fig.3

Ratio of CSF p-tau₁₈₁/Aβ₄₂ versus Conversion status with local smoothing trend (LOESS). Data are jittered around conversion status to better illustrate data densities. Panel A includes participants with MCI and those who were cognitively healthy at baseline (n = 349). Panel B includes only those individuals with MCI at baseline (n = 244). Notes: Conversion is a dichotomous variable: Converters versus Non-converters. The trend line is smoothed using LOESS in order to illustrate the relationship between conversion and CSF biomarkers without linearity assumptions. Additionally, values for the ratio of CSF analytes were jittered to more clearly see the number of participants obtaining individual ratio values.

When IICV and the ratio of CSF analytes were included in the model together, IICV continued associate with conversion odds (See Table 2). When the ratio of CSF analytes was added to models including IICV, the association between IICV and conversion appears to increase, with estimated ORs between 2.81 (95% CIs 1.83 to 4.33) and 3.41 (95% CIs 2.03 to 5.73), depending on whether cognitively healthy individuals were included in the analysis.

Lastly, we examined whether the individual components of IICV outperformed the overall estimate of variability (i.e., IICV) in predicting conversion to MCI or AD, using a larger analytic sample, not restricted to participants with CSF data (n = 1,307). Subject characteristics for this group of subjects are provided in Supplementary Table 1. An initial examination comparing conversion status with mean test performance revealed that conversion was not related to overall performance on cognitive tests (see Supplementary Figure 1). Nonetheless, the importance of IICV over and above its individual test components were compared using an AIC backward selectionprocess.

Table 3 displays the backwards selection results comparing IICV with its component tests. Overall, it appears that TMT A does not provide additional explanatory information compared to other items. Importantly, IICV was not excluded in this iterative analysis, and the final model was restricted to ANART, RAVLT, TMT B, and IICV. IICV was associated with odds of conversion over and above the tests used to calculate IICV. Thus, IICV appears to provide additional explanatory information beyond its component tests when predicting conversion.

DISCUSSION

In these analyses of ADNI data, IICV, an indicator of dispersion between cognitive test scores, was associated with elevated odds of conversion from MCI to AD or from a cognitively healthy state to MCI or AD. The IICV index used here characterized dispersion in cognitive abilities including speeded visual motor sequencing with and without an additional divided attention component, crystalized verbal abilities, and verbal memory. Whether used with or without its individual component tests, or the ratio of CSF p-tau₁₈₁/Aβ₄₂, IICV obtained at a single time point was associated with increased odds of developing incident AD and MCI.

We have previously shown in survival analyses that IICV predicted time to diagnostic conversion. Specifically, our models suggested that an estimate of IICV predicted time to incident AD and MCI even after accounting for a neuroimaging biomarker, HVL and a genetic risk factor, APOE ɛ4 status [7]. Similarly, in a middle-aged cohort enrolled in WRAP study, we noted IICV measured at midlife predicted incident cognitive decline approximately a decade after baseline [18]. In the current analyses we sought to examine how IICV compared to CSFbiomarkers.

As anticipated, the ratio of CSF values of phosphorylated tau (p-tau₁₈₁) and Aβ₄₂ were associated with elevated odds of conversion. These findings are consistent with previously published ADNI analyses, revealing associations between clinical symptoms, incident disease and CSF analytes [27, 28]. A recent summary analysis of multiple studies examining CSF biomarkers for AD-associated neurodegeneration suggested that CSF analytes are robust indicators of underlying disease pathology [2]. In particular, CSF Aβ₄₂, p-tau, total tau (t-tau), and neurofilament light protein emerged among several candidate CSF biomarkers as the most useful for differential diagnosis. While unquestionably valuable, these data can only be obtained via lumbar puncture. This invasive procedure must be performed by trained medical personnel in a clinical research setting. Moreover, based on historical research abuses, like the Tuskegee Study of Untreated Syphilis in the Negro Male [29], the lumbar puncture procedure is associated with one of the most egregious and symbolic cases of research misconduct for African Americans [30]. For participants unwilling to complete a lumbar puncture, IICV may offer an alternative.

Obtaining a reliable IICV is dependent upon having trained examiners. This would require concerted but not insurmountable effort. Moreover, it is possible that IICV would lend itself to laptop or tablet platforms, reducing the skill level needed to collect data. Altogether, there is justification to explore IICV and other alternatives to CSF biomarkers, especially low-cost, easily disseminated alternatives, not associated with burdensome procedures, especially aversive for African Americans.

When models included both the ratio of CSF p-tau₁₈₁/Aβ₄₂ and IICV, the odds of incident AD or MCI associated with elevations in IICV remained relevant. Indeed, odds ratios for these predictors remained relatively stable in iterative models, suggesting unique predictive contributions from both IICV and ratio of CSF analytes. Indeed, there appears to be little correlation between the two predictors. Supplementary Figure 2 shows a plot of IICV to the ratio of CSF p-tau₁₈₁/Aβ₄₂. Although studies examining the association of cognitive performance and CSF analytes in pre-clinical or cognitively healthy subjects have reported somewhat mixed findings, Pettigrew et al. [31] and others [4] found associations between cognitive factors and CSF amyloid and CSF tau. Analyses presented here suggest that the processes underlying cognitive variability and those sub-serving CSF amyloid and tau abnormalities may differ, possibly explaining inconsistences in findings, and why others have reported that CSF amyloid and especially tau findings were uncoupled from cognition in the preclinical stages (e.g., Li et al. [32] and Rolstad et al. [33]). Moreover, the finding that neither IICV and CSF biomarkers appeared to suppress the effect of the other when both were included in the model, suggests that the two predictors of risk could be used together in models examining risk for AD.

The concept that dispersion in performance signifies underlying disease is an established and deep-rooted theory in cognitive psychology [9–16]. Many others have examined the relationship between cognitive variability and neurological and psychiatric illnesses, including dementia [7–11, 34–36]. Explanations for the link between inconsistency and brain disease include impaired neural networks, impaired functional connectivity, as well as executive dysfunction [37, 38]. The premise of these theories is that disease results in a disrupted ability to maintain consistent mental efforts.

Like Holtzer et al. [8], our approach to estimating IICV focused on consistency across cognitive domains, rather than consistency between trials of the same task. We intentionally selected indices which could be expected to diverge early in the AD course, e.g., TMT B was contrasted to TMT A. TMT A requires a person to connect a series of numbered circles as quickly as possible; whereas in TMT B, the individual must alternate between numbers and letters, connecting the circles as rapidly as possible. A disparity in performance between the tests would identify someone with very early executive dysfunction yet well-preserved sensory motor abilities, as is typically observed in AD.

The remaining two cognitive tests contributing to the IICV index used here were included to maximize detection of early AD-related disparities in cognitive performance. A list learning test (RAVLT), selected as a measure of hippocampal-based learning [39], was contrasted with a test of reading ability estimating baseline crystallized intelligence (e.g., semantic knowledge) [40]. In total, the cognitive indices included in our measure of variability were selected if they were likely to either reflect a typically well-preserved ability or a typically diminished ability, representing the peaks and valleys in a cognitive profile.

The older participants included in these analyses were followed for an average of 2.32 years. It is unknown how early cognitive variability appears prior to a diagnosis. Holtzer et al. [8] found that IICV predicted incident dementia on average 3.3 years after initial evaluation. Participants were ∼79 years of age and cognitively normal at baseline. In our two previously published papers, we demonstrated IICV’s predictive utility in two different groups, over disparate time periods. Similar to Holtzer et al., we found that older ADNI subjects (∼74 years) who were cognitive normal or diagnosed with MCI on average evidenced cognitive variability 30.81 months prior to diagnosis [7]. In contrast, Koscik et al. demonstrated the utility of IICV in a much younger cohort (∼53 years) over a period of 8 to 10 years [18]. In total, variability in cognitive performance appears to indicate elevated risk for later cognitive impairment and/or dementia. The exact time frame for prediction is not yet clear and may depend on how IICV is estimated and/or the characteristics of the cohort being examined.

In our final analysis, we sought to examine the overall contribution of IICV to elevated risk versus the risk for conversion associated with the components of IICV. IICV survived an iterative examination, wherein elements were individually removed from models, suggesting that IICV informed risk for conversion, over and above the contribution of its components alone.

Among the limitations in our analyses, we highlight that incident cognitive impairment in this ADNI sample primarily represented conversion from MCI to AD, i.e., conversion within a clinical population. Given the low rate of conversion in cognitively healthy subjects (n = 10), we are unable to make definitive statements regarding detection of preclinical risk. However, in our previous analysis of WRAP data, our sample was middle-aged and cognitively healthy at baseline [18]. IICV predicted incident cognitive decline in the preclinical WRAP cohort. Nonetheless, we emphasize the need for replication, especially in younger cohorts, more remote from conversion. Notably, ADNI participants with MCI were selected for inclusion based on their poor performance on memory measures in the presence of spared global cognition [22]. At the outset, individuals with MCI included in this analysis would demonstrate greater cognitive variability than the general population. Finally, an additional limitation is the relative homogeneity of the sample. Subjects were predominantly non-Hispanic and white, reflecting a selection bias toward those willing to enroll in a longitudinal study involving lumbar puncture. IICV may not be associated with diagnostic conversion for Hispanic, non-white groups or for cases of mixed or non-AD dementias.