Cognitive dedifferentiation as a function of cognitive impairment in the ADNI and MemClin cohorts

Clinical perspective

To our knowledge, this is the first time cognitive impairment as explanatory variable for dedifferentiation is found in two large and separate cohorts in which each patient’s cognition has been thoroughly examined and diagnosed. The present study therefore puts particular emphasis on the dedifferentiation phenomenon in the context of geriatric care and its patient population in abnormal cognitive decline, i.e. patients with MCI or dementia.

One important issue that the present study highlights pertains to the cognitive profile of a patient, and whether this profile of performance on psychometric tests is relatively similar across cognitive domains gauged by the tests. Clinicians often reason that an “even” profile is a sign of healthy cognitive functioning, and vice versa. The present study problematizes such reasoning as it suggests that an even cognitive profile can de facto be due to dedifferentiation of cognitive abilities. Most clinicians are well-aware that a cognitive test profile that is similar across domains but significantly worse across these domains relative to normative test scores usually signifies some form of cognitive decline. If however dedifferentiation presents itself early on for a patient, it might obscure a serious condition if the clinician applies the “even profile” heuristic when presented with a cognitive profile that is a bit lower in performance relative to applicable test norms but similar across cognitive domains. Simply because AD is very common, and because the classic AD profile is uneven with fairly distinct underperformance on episodic memory tests, does not allow for diagnostic heuristics to be applied insofar that possible dedifferentiation is ignored. Further, if a cognitive profile of a patient is even because of dedifferentiation and then also coupled with inadequately estimated premorbid cognition, proper diagnosis of MCI could be substantially delayed for patients initially classified as cognitively healthy or with only subjective cognitive impairment.

Cognitive status conditions are difficult to diagnose and require thorough examination by a multidisciplinary team at a memory clinic to achieve sufficient diagnostic accuracy for the most difficult to separate patients. In clinical practice, stronger dedifferentiation with progressing neurodegeneration means that impaired patients depend to a greater extent on their general intelligence because their task specific skills developed earlier in life are deteriorating. This suggests that important lessons are to be learned regarding how these patients become, for instance, cognitively overburdened in concrete daily situations, during which they previously could offload their cognitive ability through skill heuristics but which are no longer accessible, or not as easily accessible to them.

Limitations and strengths

All observational research is limited by possibly remaining residual confounding. This includes our study, since there may be other factors than cognitive impairment that produces dedifferentiation of cognitive abilities in late life. Factors such as biological age [15], cognitive reserve [17], or brain compensation [18] might also produce dedifferentiation, and we cannot exclude the possibility that compensatory mechanisms co-exist with cognitive impairment, especially in the early phases. We were not able to directly control for all these factors in the present study, but through propensity score matching on age, education, and sex, it is reasonable to assume that we control for the bulk of such possible confounding indirectly. Control for premorbid ability was not deemed necessary as years of education was controlled for and is considered a good proxy for premorbid ability. Moreover, the different prevailing retrospective methods for estimating premorbid ability; using education and demographics, or semantic knowledge test performance (e.g. WAIS-IV Information), or a specific pronunciation type test (e.g. NART [19], ISW [20]), have their respective limitations such as overestimation and conceptual biases. More research is likely needed in which other potential confounders are controlled for, simultaneously being wary of the modelling pitfall of overadjustment bias, i.e. controlling for intermediary variables situated on the suggested causal path from cognitive impairment to dedifferentiation. Our study was also limited by data being cross-sectional. There are longitudinal measurements on some variables in both ADNI and MemClin but a more complete recording of psychometric performance across time is needed to study dedifferentiation over time within individuals. Another important feature in our study was the multiple dimensions for which the ADNI and MemClin cohorts differ. For instance, ADNI is a North American research database, employing in part different psychometric tests, brain imaging techniques, diagnostic methods, and patient inclusion procedures compared to the Swedish MemClin database. Because of these differences, the two datasets could not be combined and analysed as a whole. We could however use the strengths of this study feature by investigating, and also finding, dedifferentiation in the two datasets separately, leveraging their individually fairly large size and generally high quality of measurements. Another important difference relates to differing sampling procedures for which ADNI deliberately sampled healthy controls whereas MemClin controls where SCI patients. The pattern of dedifferentiation was also more pronounced in ADNI compared to MemClin. Importantly, MemClin is a new large-scale naturalistic database with high ecological validity and generalizability to the clinical geriatric population in its specialized care setting provides both a rare strength, and a complement to the excellent ADNI database for conducting such studies as the present one.

Study population

We included patients from the Alzheimer’s Disease Neuroimaging Initiative (ADNI, n = 1282) [13] and the Memory Clinic Project (MemClin, n = 1223) [12]. The diagnostic procedures for ADNI [21] and MemClin have previously been described [12]. For ADNI, the diagnosis procedure relied on subjective and objective cognitive measures but was independent of biomarker information. For MemClin, diagnosis was determined at multidisciplinary consensus meetings based on cognitive testing, neurological examination, brain imaging, biomarkers and other routine clinical measures as available [12, 22, 23]. The diagnosed cognitive statuses of dementia, MCI, and SCI were collapsed into two classes: CI and NCI. CI represents impairment of clinically diagnosed severity whereas NCI represents no objective impairment.

Cognitive variables

The administered neuropsychological batteries of ADNI and MemClin are described elsewhere [12, 14] and covered similar cognitive domains and subdomains with some differences by original design. Importantly however, across main analysis comparisons of dedifferentiation in on CI vs NCI variables were identical in each database. Unless further defined, psychometric tests are total raw scores. Some psychometric variables were reverse coded as needed. Psychometric tests are separated with semi-colon (;) and subtests belonging to the same test are separated with dash (/) as follows.

From ADNI we selected patient performance on 15 variables: the Mini Mental State Examination (MMSE); Clock test; Copy test; Rey Auditory Verbal Learning Test (AVLT) trial 1/2/3/4/5/total score/delayed recall/recognition; semantic fluency (animals); Trail-Making Test (TMT) A/B; and Boston Naming Test spontaneous recall.

From MemClin we selected 20 variables: MMSE; Wechsler Adult intelligence Scale 4^th edition (WAIS-IV) Information; Block design; Digit Span total/forward/backward; AVLT 1/2/3/4/5/total score; Rey Complex Figure copy; Delis-Kaplan Executive Function System (D-KEFS) Verbal fluency FAS/semantic fluency/shifting (correct answers)/shifting (correct shiftings); and TMT 1/2/3. TheHiveDB was used for MemClin data management [24].

Statistical analyses

All analyses were mirrored and run separately in both ADNI and MemClin. Due to missing data some variables were excluded (cut-off > 25% missing). Remaining missing data sorted by % missing of total (in ADNI: TMT B [1.9], NART [0.7], sex [0.3], age [0.3], edu [0.3] and in MemClin: AVLT total [23.8], D-KEFS TMT 3 [22.8], RCFT copy [22.6], WAIS-IV Block design [20.9], D-KEFS TMT 1 [20.7] was deemed acceptable, assumed to be missing at random (MAR) and imputed with k Nearest Neighbour (kNN) [25, 26] applying the unweighted Gower distance metric [27] with k set to three.

To control for confounding of the Dedifferentiation ~ CI association, propensity score matching [28, 29] was performed prior to main analysis. A logit model was constructed estimating the probability for CI with age at examination, years of education, and sex as predictors. Exposed (CI) where thereafter propensity score matched 1:1 with unexposed (NCI) without replacement on this probability applying SD = 0.05 caliper width. Figure 3 plots the matching diagnostics and resulting across-group balance in the confounders.

Figure 3. Propensity score matching diagnostics in the ADNI and MemClin datasets. (A) Patient counts after matching are shown as a function of their individual propensity score and overlayed with density plots, stratified by level of cognitive impairment. (B) Mean distance followed by single-covariate balance by group calculated before (Unadjusted) and after (Adjusted) matching. ADNI matched: n CI= 392, n NCI = 392. MemClin matched: n CI = 143, n NCI = 143. ADNI, Alzheimer’s Disease Neuroimaging Initiative; MemClin, Memory Clinic Project; CI, cognitive impairment; NCI, no cognitive impairment.

Main analysis

After matching, principal component analysis (PCA) was performed on the mean centered and unit variance scaled psychometric variables through single value decomposition. Percentage explained variance in psychometric variables by the first unrotated component (PC1), a proxy for general cognitive ability, was compared across the matched CI and NCI groups expecting a smaller % explained variance by PC1 among scores in the CI vs the NCI group. We also report n components with eigenvalues > 1 (Kaiser criterion) expecting fewer selected components by this criterion for CI vs NCI. Loadings on PC1 by key psychometric tests that tap central cognitive domains were expected to be higher in CI vs NCI groups. Supplementary Tables 1, 2 show PCA analysis by subgroups after propensity score matching for both the ADNI and the MemClin cohort.

Correlation coefficients calculated across all informative pairings of psychometric variables were thereafter sorted from weakest to strongest and assessed by t-test and also by visual inspection comparing the strengths of correlations across matched CI and NCI groups expecting average test-pair correlations to be higher in the CI vs the NCI group.

Additional statistics

Unless further explained, bivariate summary statistics are presented as mean (SD) for numeric variables and count (%) for factors. Statistical significance was set to 5%.

Analyses were performed in R version 4.0.1 using packages caret, cobalt, corrplot, data.table, dummies, factoextra, FactoMineR, foreign, ggplot2, gridExtra, haven, MatchIt, matrixStats, mice, readxl, tableone, and VIM.