A δ Homolog for Dementia Case Finding with Replication in the Alzheimer’s Disease Neuroimaging Initiative

INTRODUCTION

We have pioneered a novel approach to the assessment of dementia. It is conceptually simple. While cognitive impairment is widely held to be the hallmark of dementia, three conditions are necessary to that diagnosis [1]: 1) there must be acquired cognitive impairment(s); 2) there must be functional disability; and 3) the disability must be related to the cognitive impairment(s) that are observed.

This formulation implies that the essential feature(s) of dementing processes can be resolved to the cognitive correlates of functional status. Explicitly measuring that construct can open a way to the completely empirical assessment and diagnosis of dementia.

Surprisingly, the association between many observed cognitive measures/domains and functional status is statistically weak [2]. However, an empirical definition of the “cognitive correlates of functional status” can be approached by confirmatory factor analysis in a structural equation model (SEM) framework. We have found that functional status is linked to cognitive performance through Spearman’s General Intelligence factor “g” rather than through domain-specific cognitive abilities [3, 4]. Our bifactor SEM model parses g into two orthogonal (unrelated) fractions: 1) δ, i.e., “the psychometric correlates of functional status”, and 2) g’, i.e., residual variance in g that is empirically unrelated to Instrumental Activities of Daily Living (IADL). Cognitive variance empirically unrelated to IADL cannot contribute to dementia by our definition. Thus, our method divorces functionally salient cognitive impairment from cognitive impairment per se.

The latent variable δ can be “reified” as a composite “d-score” and applied to individuals as an omnibus dementia severity metric, i.e., a dementia-specific phenotype. Because g is thought to contribute to all cognitive measures, it has proven feasible to construct δ from a wide range of measures, batteries, and samples. δ can be constructed from a comprehensive battery of formal measures, from small batteries of formal measures, from brief batteries of “bedside” measures, and even from the items of a single measure.

We refer to each embodiment as a δ “homolog”. In genetics, a homolog is a gene descended from an ancestral gene in the same species and preserving the original’s function. Similarly, and regardless of their cognitive indicators, all δ homologs (twelve published to date) share the bifactor construction of δ, target a measure of IADL, exhibit strong associations with dementia severity (e.g., as measured by the Clinical Dementia Rating Scale Sum of Boxes (CDR-SB) [5] and achieve high areas under the receiver operating characteristic curve (AUC/ROC) for the discrimination of various dementias from normal controls (NC).

Furthermore, it appears that δ is agnostic to the etiology of dementia [6, 7]. This observation suggests that variance in cognitive performance can be divided into fractions that are necessary to dementia (i.e., δ) and irrelevant to it (i.e., disease-specific variance). To the extent that disease-specific cognitive changes (and domain-specific ones, by definition) and their biomarkers are orthogonal (unrelated) to δ [7], they may be incapable of explaining functionally salient effects. In contrast, the biomarkers of δ (regardless of their disease specificity) may be functionally salient, and intervention on them should produce functionally salient outcomes.

We have been studying δ homologs and their biomarkers in the Texas Alzheimer’s Research and Care Consortium (TARCC) study. TARCC is a large (N ≅ 3,500), well-characterized, ethnically diverse [n ≅ 1,200 Mexican-American (MA)] convenience sample with annual longitudinal follow-up [8].

We have associated δ with a large number of serum proteins, including pro- and anti-inflammatory cytokines [9, 10]. Many of these associations appear to have profound ethnicity effects in TARCC [11, 12]. We have also published the serum protein mediators of specific dementia risks (including age, depressive symptoms, and the apolipoprotein (APOE) ɛ4 allele) [12–16].

All these associations have been replicated in random subsets of TARCC’s large sample. Regardless, TARCC has its limitations. No imaging is available and its protein biomarkers have been obtained in serum. Biomarker associations have been notoriously difficult to replicate and may be impacted by the biofluid in which they are measured [17].

The Alzheimer’s Disease Neuroimaging Initiative (ADNI) offers an opportunity to replicate our TARCC findings [18]. Its cognitive battery overlaps substantially with that of TARCC, and both deployed similar blood-based biomarker panels processed by a common vendor [i.e., Rules Based Medicine (RBM) of Austin, Texas]. If we can validate a common δ homolog across both studies, we might be able to confirm blood-based biomarkers of δ across those studies and integrate ADNI’s neuroimaging into the growing literature of δ.

MATERIALS AND METHODS

RESULTS

Descriptive statistics are presented by group in Table 1, and by diagnoses in Tables 2 (TARCC) and 3 (ADNI). Cohen’s d and t tests of significance are reported in Table 1, where estimable, for TARCC (total sample) versus ADNI. These samples differed significantly on all measures. There were no significant differences across the TARCC random splits on any variable. Both splits differed from ADNI on all measures. ADNI appears to have a relatively high fraction of MCI cases, which were recruited explicitly into ADNI-2 and ADNI-GO. TARCC has a much higher prevalence of MA participants.

dT2A’s unadjusted and constrained TARCC models had excellent fit [χ² = 73.6 (20), p < 0.001; CFI = 0.996; RMSEA = 0.028]. All the observed cognitive performance measures loaded significantly on dT2A, ranging from BNT (r = –0.31) to LMII (r = –0.84; all p < 0.001). dT2A exhibited excellent factor determinacy (Table 2). dT2A’s factor weights all replicated in a random split of TARCC’s sample [ΔCHI SQ = 11.8 (6), p > 0.05]. The constrained unadjusted model correlated r = 0.99 with CDR-SB (p < 0.001). For comparison, IADL correlated r = 0.86 (p < 0.001) with CDR-SB across TARCC’s entire sample.

The dT2A composite achieved an AUC = 0.981 (0.976–0.985) for the discrimination between AD and NC. IADL alone achieved an AUC = 0.936 (0.927–0.945) for the same discrimination. This was statistically inferior to dT2A’s by DeLong et al.’s method (p < 0.001). dT2A’s IADL adjusted bivariate AUC was 0.983. Both dT2A (p < 001) and IADL (p < 0.001) made significant contributions. The fully adjusted constrained dT2A correlated r = 0.86 (p < 001) with CDR-SB (Fig. 1).

In ADNI, dT2A’s model had excellent fit [χ² = 12.464 (7), p < 0.001; CFI = 0.999; RMSEA = 0.019]. All the observed cognitive performance measures loaded significantly on dT2A, ranging from BNT (r = –0.53) to MMSE (r = 0.86; all p < 0.001). The unadjusted dT2A exhibited excellent factor determinacy, and correlated r = 0.96 with CDR-SB (p < 0.001). For comparison, the FAQ correlated r = 0.83 (p < 0.001) with CDR-SB. dT2A’s composite achieved an AUC = 1.0 (0.995–1.00) for the discrimination of AD from NC (Table 2). The FAQ alone achieved an AUC = 0.985 (0.974–0.993) for the same discrimination. This was statistically inferior to dT2A’s by DeLong et al.’s method (p = 0.001). dT2A’s FAQ adjusted bivariate AUC was 0.9997. However, only dT2A made a significant contribution (p < 0.001). The FAQ did not enter (p = 0.08). The fully adjusted ADNI dT2A correlated r = 0.96 (p < 001) with CDR-SB (Fig. 2).

OPEN IN VIEWER

Fig. 1.

Fully Adjusted dT2A Homolog and its Association with CDR-SB in TARCC*. *dT2A bifactor δ homolog model in random 50% splits of TARCC’s sample and with parameters constrained across splits. All observed indicators are adjusted for age, education, ethnicity, GDS scores, and gender (paths not shown for clarity). Animals, Animal Naming; BNT, Boston Naming Test; CHI QS, Chi Square; CFI, Comparative Fit Index; EDUC, Education (years); GDS, 30 item Geriatric Depression Scale; IADL, Instrumental Activities of Daily Living; LMI, Wechsler Logical Memory immediate recall; LMII, Wechsler Logical Memory delayed recall; MMSE, Mini-Mental State Exam; RMSEA, Root Mean Square Evaluative Assessment; TARCC, Texas Alzheimer’s Research and Care Consortium; Trails B, Trail-Making Test part B.

OPEN IN VIEWER

Fig. 2.

Fully Adjusted dT2A Homolog and its Association with CDR-SB in ADNI*. *All observed indicators are adjusted for age, education, ethnicity, GDS scores, and gender (paths not shown for clarity). ADNI, Alzheimer’s Disease Neuroimaging Initiative; Animals, Animal Naming; BNT, Boston Naming Test; CHI QS, Chi Square; CFI, Comparative Fit Index; EDUC, Education (years); FAQ, Functional Abilities Questionnaire; GDS, 30 item Geriatric Depression Scale; LMI, Wechsler Logical Memory immediate recall; LMII, Wechsler Logical Memory delayed recall; MMSE, Mini-Mental State Exam; RMSEA, Root Mean Square Evaluative Assessment; Trails B, Trail-Making Test part B.

In TARCC, dT2A contributed to 5-year prospective AD conversion risk independently of other demographic risks (Table 5). dT2A’s effect attenuated all the unadjusted risks, and fully attenuated the effect of depressive symptoms (Table 6).

DISCUSSION

We have employed SEM to construct a latent dementia-specific phenotype engineered to allow direct comparisons between TARCC and the ADNI. dT2A fits well in either dataset, is strongly associated with CDR-SB in both, and has a high AUC for discrimination of AD from controls in both. Thus, it may now be feasible to replicate TARCC’s proteomic findings in ADNI.

Replications across these studies would be highly desirable because ADNI offers access to multimodality neuroimaging and AD-specific cerebrospinal fluid biomarkers while TARCC does not. ADNI’s biomarkers could be used to test the relevance of TARCC’s proteomic findings to AD-specific biomarkers and explore the neuroimaging mediators of their associations with cognitive performance. TARCC has the advantage of a larger and more ethnically diverse sample. Moreover, a third well-characterized cohort, i.e., the Australian Imaging, Biomarker & Lifestyle Flagship Study of Ageing (AIBL), also has RBM proteomic data. Like TARCC, its cognitive battery is based on ADNI’s. Like ADNI, it measures proteomic data in plasma. It should be possible to construct dT2A in the ABIL dataset and provide yet other opportunities for the independent replication of TARCC’s and ADNI’s findings.

However, obstacles remain to these goals. First, while TARCC and ADNI share dT2A’s indicators, those are but a fraction of the measures available in either study. Second, TARCC and ADNI are convenience samples with differing case-mixes and demographic profiles. Third, although they share a common proteomic profile, TARCC and ADNI’s blood-based biomarkers have been measured in two different biofluids (i.e., serum and plasma).

dT2A’s indicators are but a fraction of the measures available in TARCC and ADNI. We have observed that the diagnostic accuracy of a δ homolog can be affected by its indicators. In general, accuracy improves with larger and more comprehensive batteries. Such batteries are less likely to be compromised by domain-specific variance, and more likely to attain Spearman’s g. Had there been a greater direct correspondence across these studies, we might have engineered a better homolog.

Regardless, δ homologs seem to be robustly “indifferent” to their indicators. The current homolog still exhibits a high degree of diagnostic accuracy. The same feature is widely acknowledged to be a property of g, and is a notable feature of “intelligence” as measured by these constructs. This property explains one major advantage of our approach. δ homologs can be engineered to accommodate a wide range of agendas. δ can be constructed as easily from bedside assessments as from formal psychometrics [1, 43]. The resulting loss of diagnostic accuracy may be justified by lesser administrative burdens and an increase in the number and availability of qualified administrators.

δ can be even be constructed from the items sets of individual cognitive measures [44] or from verbal batteries, i.e., in anticipation of telephone administration [45]. However, the loss of diagnostic accuracy may be consequently greater, presumably due to contamination by domain-specific variance, which would be orthogonal to g (and δ) when derived from a more comprehensive battery.

Another potential obstacle is that TARCC and ADNI are convenience samples with differing case-mixes and demographic profiles. However, the latent variable approach minimizes non-systematic measurement bias, including cultural and linguistic bias. TARCC has a relatively high prevalence of Hispanic participants, many of whom were tested in Spanish, and with significantly lower educational attainment. Regardless, δ’s factor loadings, obtained from data collected in TARCC, have been exported to Japanese data without loss of diagnostic accuracy [47], and δ homologs perform similarly in well-characterized European [43] and Asian [48] samples.

Therefore, we did not attempt to demonstrate factor equivalence across these samples. We reasoned instead that the effects of study-specific demographic differences would be pushed up into δ’s residuals, potentially changing its indicator loadings, but without necessarily compromising its performance. This seems to be validated by δ’s similar performance across cohorts.

TARCC and ADNI also differ in their case-mixes. Since d-scores are standardized to each study’s distribution, the optimal diagnostic threshold may not replicate across samples. However, we have reported that factor weights derived post hoc (as in this application) in poorly-characterized epidemiological samples can be exported into well-characterized samples for validation and the resulting thresholds exported back to the original sample without performance compromise [46].

Moreover, there may be cross-sample variation with regard to the within diagnosis prevalence of AD-specific biomarkers. Even though both the current samples are likely to be highly enriched with AD cases, no AD-specific biomarkers are available in TARCC, while a clinical diagnosis of “AD” is not confirmed by amyloid imaging in a substantial fraction of cases [49, 50]. We cannot be sure that a clinical diagnosis of “AD” is bioequivalent across these cohorts.

Regardless, δ appears to be “agnostic” to dementia’s etiology [6, 7]. Just as δ distinguishes functionally-salient cognitive variance from cognitive impairment per se, its biomarkers (including any disease-specific ones) should distinguish dementia-salient biological processes from non-dementing ones, e.g., the disease-specific cognitive changes recently shown to orthogonal to δ [7]. Thus, the distinction between δ and disease-specific cognitive changes may allow for the identification of functionally-salient dementia-specific biomarkers.

Moreover, clinical dementia in either cohort is likely to be “overdetermined” by multiple contributors to δ’s variance [51]. Age, the APOE ɛ4 allele, and depressive symptoms contribute variance to δ independently of each other [12]. Each of their effects is mediated by largely non-overlapping sets of serum proteins [13–16]. Cross-study differences in the prevalence of these risks could alter δ’s biomarker associations.

dT2A attenuates multiple prospective “AD risks” in TARCC, and fully explains the nearly twofold conversion risk associated with depressive symptoms. However, to the extent that dT2A failed to fully attenuate the other risks, they can be said to contribute to conversion independently of disabling cognitive decline. That insight implies demographic bias in clinicians’ estimates of cognitive status. Second, AD-specific neurodegeneration may not mediate either age’s or depression’s effects on δ [43, 52]. Moreover, without either autopsy or AD-specific cerebrospinal fluid or neuroimaging biomarkers, it remains unclear from TARCC data whether AD-specific neurodegeneration is involved in any of these dementing processes.

We have associated an ethnicity equivalent δ homolog (i.e., “dEQ”) with multiple serum proteins in TARCC [13]. Similarly, O’Bryant et al. have identified yet another set of proteins related to clinical “AD” [53]. The latter have not yet been tested for their relationship to δ, but it seems very likely that they will be related, given their relatively high AUC for the discrimination of AD from controls in TARCC [54] and other cohorts [55, 56]. The relationship of these biomarkers to those we have associated with δ [11] is yet untested, but there is little overlap across the two sets. O’Bryant et al. find evidence for their panel’s association with dementia in Hispanics [57], while we do not for ours [11, 12].

Finally, TARCC and ADNI’s blood-based biomarkers have been measured in two different biofluids (i.e., serum and plasma). Some observers opine that blood biomarker studies are plagued by failure to replicate findings. Regardless, none have employed our (latent variable) SEM methods. Those can be applied equally to biomarkers as to cognitive batteries. Moreover, few studies have employed samples as large as TARCC’s, which is twice as large as ADNI’s. We remain optimistic that this approach will finally allow for biomarker replications cross-cohort and across biofluids.