Longitudinal change-change associations of cognition with cortical thickness and surface area

Abstract

Age-related changes in cortical volumes are well established but relatively few studies probed its constituents, surface area (SA) and thickness (TH). Here we analyzed 10-year, 3-waves longitudinal data from a large sample of healthy individuals (baseline age = 55–80). The findings showed marked age-related changes of SA in frontal, temporal, and parietal association cortices, and Bivariate Latent Change Score models revealed significant SA-associations with changes in speed of processing in both the 5- and 10-year models. The corresponding results for TH revealed a late onset of thinning and significant associations with reduced cognition in the 10-year model only. Taken together, our findings suggest that cortical surface area shrinks and impacts information-processing capacity gradually in aging, whereas cortical thinning only manifests and impacts fluid cognition in advanced aging.

Introduction

The aging brain undergoes a variety of changes, at multiple scales, that have been explored in relation to cognitive aging (e.g., [15]). In human longitudinal studies, which allow assessment of brain-cognition change-change associations, magnetic resonance imaging (MRI) markers have been linked to various cognitive measures. The number of studies is still limited, and the methodological variation is substantial, but converging evidence has emerged for correlated changes between episodic memory and gray-matter integrity in medial-temporal lobe regions (i.e., more memory decline relates to more atrophy, e.g., [7], [10], [23]; for a review, see [19].

Brain-cognition change-change relations have also been observed for cortical regions (see [19]). Cortical volumes were examined in most published studies to date (e.g., [1], [18]), but it might be crucial to decompose volume into surface area (SA) and thickness (TH). This is because there is evidence that SA and TH have distinct genetic architectures [8]. Moreover, recent findings indicated differential change patterns and associations with cognitive measures [2], with TH more strongly related to aging and SA more strongly related to cognition. A stronger TH-age relation is supported by some (e.g., [24]), but not other (e.g., [23]) studies, and inconsistent findings have also been reported in relation to cognition (i.e., more robust cognition associations for TH, [23].

The mechanisms that explain links between cognition and cortical configuration remain unclear. In a study of 740 young individuals from the Human Connectome Project, SA and TH were related to measures of fluid (e.g., working memory, speed, episodic memory, visuospatial abilities) and crystallized (e.g., general knowledge, vocabulary) cognitive abilities [25]. Tadayon et al. suggested that their findings of a link between better fluid cognition and larger SA in frontal, temporal, and parietal areas may reflect more cortical columns and thereby higher information-processing capacity (fluid cognition was unrelated to TH). Crystallized abilities were related to larger SA along with thinner TH (cf., [26]), possibly reflecting pruning of unnecessary and weak neural connections. Extending the findings of Tadayon et al. to older age, a prediction for SA is that preserving or at least limiting SA reduction will be advantageous for fluid cognition in aging. Indeed, there is evidence for a consistent relation of SA with fluid cognition across the lifespan [27]. A more complex model seems necessary to capture a putative shift from no or even a negative relation of TH with fluid cognition in younger age to a positive relation in older age (cf., [5]).

In sum, the evidence from existing longitudinal studies of age-related changes in SA and TH and their relation to cognitive changes is inconsistent. The purpose of the present study was therefore to compare SA and TH on magnitudes of age-related changes as well as brain-cognition change-change associations, using longitudinal MRI data and fluid cognition (verbal memory, speed-of-processing, verbal fluency, visuospatial ability) across 10 years and 3 waves from the Swedish Betula project [17]. Previous Betula studies of this sample have demonstrated age-related decreases in cortical volumes (e.g., [10]), but this is the first Betula study of SA and TH. We compared three regions across the cortical mantle (rostral middle frontal, middle temporal, inferior parietal cortex) where past longitudinal studies revealed age-related changes in SA and TH (e.g., [24]) and where measurement properties were excellent (see Methods). For SA we predicted substantial age-related changes and significant brain-cognition associations. For TH, we predicted weaker age-related changes and brain-cognition associations.

Materials and methods

Participants and cognitive data

The present analyses are based on data from the longitudinal Betula study on memory, health, and aging [17]. The research was approved by the local ethics board at Umeå University, Sweden. All participants provided written informed consent and were compensated monetarily for their participation. Some of the participants were also part of an imaging subsample scanned for the first wave in 2009–2010 and then at two additional test waves within 10 years (see [17]). Here, of 292 individuals from the first imaging session, 283 had valid brain and cognitive data and could be included in the analyses (Mean age at wave 1 = 67.2, range 55.2–81.3 years, 148 females). At the second imaging session, 163 individuals were available for longitudinal analyses (Mean age at wave 2 = 69.4, range 59.3–84.8 years, 74 females). At the third and final imaging session, data from 90 individuals could be included in the analyses (Mean age at wave 3 = 72.3, range 63.0–87.9 years, 39 females).

Dropout analyses (Fig. S1) first compared SA and TH for individuals who took part in both wave 1 and 2 (N = 163, mean age at wave 1 = 65.4) with individuals who took part in wave 1 only (N = 57, mean age at wave 1 = 70.8; 63 individuals were not included in this analysis since they took part in the Betula health and cognitive test sessions but could not be scheduled for scanning within the time window, cf., [16]. Second, individuals who took part in waves 1, 2, and 3 (N = 90, mean age at wave 1 = 65.1) were compared with individuals who took part in waves 1 and 2 only (N = 72, mean age at wave 1 = 65.7; 1 case of missing cognitive data at wave 7). The dropout analyses indicated a weak dropout effect between the 1st and 2nd wave for TH (p = 0.05) and a trend for SA (p = 0.09), with no significant effects between waves 2 and 3 (Fig. S1).

At each test wave, the participants completed a test battery including several tests of fluid cognition: Visuospatial ability (block design), speed of processing (letter-digit substitution), verbal memory (verbal free recall), and semantic memory (verbal fluency). For a detailed description of the cognitive test battery, see Nyberg et al. [17].

MRI scanning and image processing

The same 3 T General Electric scanner (equipped with a 32-channel head coil) was used to collect images at all imaging sessions. T1-weighted data (T1w) were acquired in axial orientation with a 3-dimensional fast spoiled gradient echo sequence (176 slices with 1 mm thickness, TR: 8.2 ms, TE: 3.2 ms, TI: 450 ms, flip angle: 12°, field of view: 25 × 25 cm, matrix: 256 × 256, reconstructed to 512 × 512). All T1-images were processed through the longitudinal pipeline in Freesurfer ver. 7.11 [20].

Cortical surface area and thickness average were extracted for the ROIs of interest and summed respective averaged over hemispheres. The following Desikan-Killiany FreeSurfer ROIs were used; (i) rostral middle frontal, (ii) middle temporal, and (iii) inferior parietal cortex. All segmentations were manually checked for the three regions and 6 subjects were excluded due to misclassifications close to the rostral middle frontal area.

Statistical analyses

We used the Pearson correlation between consecutive timepoints as one measure of stability for the SA and TH measurements in the Betula sample, along with reliability measures of repeated scanning reliability of the scanner with respect to SA and TH (further described in a separate section below). To investigate longitudinal age effects, we used a Generalized Additive Mixed Model (GAMM, [28]) approach to estimate the age trajectories of SA and TH without assuming a particular shape of the trajectories.

To define the breakpoint-age at which thickness started the decline in this sample, a model with thickness as a piecewise linear function of age was applied, where the first line is set to be horizontal and the slope of the second line is estimated from data after the breakpoint. The breakpoint was chosen such that the piecewise model explained most variance of the estimated GAMM-smooth.

For the investigation of longitudinal Change-Change associations between the various cognitive domains and SA/TH, we used Bivariate Latent Change Score (BLCS) models [12]. The BLCS models were used for two- and three test wave data, employing full-information maximum likelihood estimation for utilizing information from subjects with missing data at one or two test wave(-s), under the assumption that the missing data mechanism is Missing-at-random (MAR, [13]). For the 3-wave BLCS models, we investigated models with and without an invariance restriction on the change-change covariances across timepoints, where the restriction is that they are set to be equal across timepoints. As the restricted model is nested within the non-restricted, we could use a standard Likelihood Ratio Test for model comparison (cf., [12]).

Principal Component Analysis (PCA; [9]) was used to identify a general cognition/SA/TH score. Separate PCAs were done on the first test wave-data for: (i) the four cognitive tests, (ii) SAs in the three cortical ROIs, and (iii) THs in the three cortical ROIs (rostral middle frontal, middle temporal, inferior parietal cortex). Next, the obtained loadings for the first PC were used to calculate scores for the first PC across all test waves. Independent group t-tests on age-residualized SA and TH scores were used to compare individuals who remained versus dropped out across test waves (Fig. S1). When reporting p-values, we consider p < 0.05 as statistically significant. All calculations were performed in R (https://www.R-project.org/) or Matlab (ver R2014b, https://www.mathworks.com/).

Reliability and stability estimates

Four volunteers (not part of the Betula study) were scanned 3 times in a row on the same scanner using the same T1 sequence as in the main study. They exited the scanner between scans 1–2 and 2–3. SA and TH were compared across repeated scans and within-persons using IntraClass Correlation (ICC) for agreement and consistency as measures of scanner reliability (see e.g. [14]). The reliability estimates were high for both SA and TH (Table S1), but higher for SA (mean consistency = 0.996; mean agreement = 0.997) than for TH (mean consistency = 0.844; mean agreement = 0.855).

Assessment in the Betula sample of stability of individual differences over time revealed higher correlations across test waves in all regions for SA (mean r = 0.99) than for TH (mean r = 0.89). All pairwise correlations are given in Table S2, and an example is presented in Fig. S2 for temporal SA and TH.

Results

Age-related changes in SA and TH

GAMM analyses (536 cross-sectional and longitudinal observations) revealed significant changes in SA in all three regions (Fig. 1a; all p’s < 2 × 10⁻¹⁶). For TH, the change patterns were weaker, but still significant in parietal (p = 2.3 × 10⁻⁶), frontal (p = 0.03) and temporal (p = 0.01) cortex (Fig. 1b). Corresponding age-change analyses were done on global cortical factors for SA (R² = 78.4) and TH (R² = 80.3), as defined by PCA. Significant age effects were observed for both global factors (Fig. 1c), but less strongly for TH (p < 0.001) than for SA (p < 2 × 10⁻¹⁶).

Fig. 1.

Longitudinal trajectories in frontal, temporal, and parietal SA (a) and TH (b). (c) Longitudinal trajectories of factors for global SA (left) and TH (right).

As illustrated in Fig. 1, for SA the age-change trajectory was approximately linear across the examined age range, whereas onset of change occurred at a higher age for TH (after age 70). This impression was substantiated by a breakpoint analysis (see Methods), which revealed that cortical thinning occurred after age 70 (Fig. S3).

Longitudinal brain-cognition relations for SA and TH

A global Cognition factor was identified by PCA (R² = 51.0; Fig. 2a) and related to SA and TH. The 2- and 3-wave BLCS models were non-significant (Table 1) for global cognition in relation to both the global TH and SA factors illustrated in Fig. 1c.

Fig. 2.

Age-trajectories for global cognition (a) and speed-of-processing (b). Change in speed was related to change in temporal SA (c) and change in cognition with TH change at the last time point (d).

Table 1.

Brain-cognition change-change relations.

	SA		TH
	BLCS 2-w	BLCS 3-w	BLCS 2-w	BLCS 3-w
Cortex-Cognition	0.13	0.03	−0.02	0.1*
Cortex-Speed	0.19	0.16	0.06	0.13
Cortex-Visuospatial	0.02	−0.03	−0.02	0.08
Cortex-Verbal memory	0.04	−0.06	−0.02	0.07
Cortex-Semantic	0.04	0.08	−0.03	0
Temporal-Speed	0.21	0.19	0.11	0.13
Parietal-Speed	0.13	0.12	0.1	0.14
Frontal-Speed	0.14	0.08	−0.05	0.05*
Temporal-Visuospatial	0.02	0.04	−0.03	0.06
Parietal-Visuospatial	0.02	−0.05	−0.01	0.1
Frontal-Visuospatial	0.02	−0.03	−0.01	0.05
Temporal-Verbal memory	0.04	0.03	−0.02	0.09
Parietal- Verbal memory	0.09	0.1	−0.06	−0.02
Frontal- Verbal memory	−0.02	−0.19	0.03	0.09
Temporal-Semantic	0.08	0.08	−0.06	−0.05
Parietal-Semantic	0.09	0.11	−0.04	0
Frontal-Semantic	−0.03	−0.01	0.04	0.05

BOLD = p < 0.05; * Models where the invariance test was significant (p < 0.05); Cortex = PCA-score for all 3 regions; Cognition = PCA-score for all 4 tasks; BLCS = Bivariate Latent Change Score Models; w = wave.

Analyses at the task-specific level revealed that cortical SA change was significantly related to speed (Fig. 2b) in both the 2- and 3-wave BLCS models. The relations of the global SA factor with visuospatial ability, verbal memory, and semantic memory were all non-significant (Table 1). Cortical TH change was unrelated to changes on all cognitive tasks, with the exception for speed in the 3-wave BLCS model (Table 1).

Analyses at the region- and task specific levels converged on brain-cognition associations for speed. Speed was significantly associated with temporal SA in both the 2- and 3-wave BLCS models (Fig. 2c), with parietal SA in the 3-wave BLCS model (Table 1), and with temporal and parietal TH in the 3-wave BLCS models (Table 1). With only 2 exceptions out of 36 comparisons, no associations were seen for visuospatial ability, verbal memory, or semantic memory (Table 1). The invariance tests did not reject the assumption of equal change-change covariance across time points for most comparisons, but significant invariance was seen for TH in two 3-wave BLCS models (Table 1). In both cases, it reflected weak brain-cognition relations at the first timepoint along with a stronger relation at the second timepoint which reached significance in the cortex-cognition model (p = 0.04, Fig. 2d).

Discussion

The present longitudinal analyses revealed pronounced SA changes in frontal, temporal and parietal cortex along with less marked changes in TH. Specifically, at the regional as well as global levels, SA declined approximately linearly across the examined age span, whereas TH was significantly reduced only after age 70 (Fig. 1 & S3). These findings are consistent with a recent study in which it was concluded that: “one has to be careful in declaring area as being less susceptible to the effects of aging than thickness. In fact, our data suggest that it may be the other way around in older age“ [23].

The second and main goal of this study was to compare changes in SA and TH in relation to cognitive changes. The estimates of SA and TH had good reliability and long-term stability, with slightly higher values for SA, which provided a solid basis for relating these brain metrics to cognitive changes. The analyses of brain-cognition change-change associations yielded mostly weak associations, which is consistent with much prior work (e.g., [19]). Still, some significant associations were observed, notably in relation to the measure of speed which in Betula and other longitudinal studies has been found to be highly age sensitive [17], [22]; cf., [21].

Our observation of a significant brain-cognition association for SA agrees with previous findings (e.g., [2], [3]). At the regional level, temporal SA had the strongest association with speed, but similar trends were seen for frontal and parietal SA. Also, the global SA score was significantly related to speed in both the 2- and 3-wave models. Thus, older adults who maintained cortical SA showed less slowing, which supports and extends the suggestion that larger SA in younger age reflects more cortical columns and higher information-processing capacity [25].

TH had a late onset of age-related changes and the change-change relation with speed was only significant in the 3-wave BLCS-models. These observations indicate that cortical thinning becomes salient in advanced aging, likely due to factors such as reduced synaptic density and/or cell-body size [6]. At this late stage, degree of thinning is positively related to cognitive change. That is, while thinner cortices in younger ages tend to be related to superior cognition, late-onset cortical thinning is associated with cognitive decline (cf., [5]). It should be noted that although the TH-cognition change-change association was strongest for the speed task, the 3-wave Cortex-Cognition BLCS model revealed a significant correlation at the last time point. Hence, late-onset cortical thinning may more generally relate to decline in fluid cognition.

A strength of the present study is the longitudinal design, but it is well known that inferences from longitudinal studies can be affected by selective attrition (e.g., [11]). The dropout analyses revealed that individuals who dropped out after the first imaging session tended to have lower TH (and to some degree also SA) than those who remained in the study (Fig. S1). Thus, a positive selection bias could have limited the magnitude of longitudinal TH changes. The fact that marked changes still were seen in SA supports the view that SA is susceptible to the effects of normal aging. A potential limitation is that we restricted the analyses to three cortical regions. These large ROIs covered the anterior-posterior axis (frontal-temporal-parietal) and had very good measurement properties. Moreover, individual differences in volume changes have been found to be quite homogenous across different cortical zones [4], [18]. Still, we cannot completely rule out the possibility that inclusion of other ROIs could have yielded different results, such as a stronger age-relation for TH than for SA (cf., [2]).

Conclusion

The findings of the present study show that cortical SA is markedly reduced in normal aging, and that heterogeneity in the degree of SA reductions to some extent explain individual differences in speed of processing. Preservation of SA in aging may reflect sparing of cortical columns and thereby higher information-processing capacity.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: