Partial Least Squares Regression Analysis of Alzheimer’s Disease Biomarkers,Modifiable Health Variables,and Cognitive Change in Older Adults with Mild Cognitive Impairment

Participants

Participants with a diagnosis of MCI from the ADNI1 cohort were included in the current study. Full participant inclusion/exclusion criteria are available in [25] and are summarized here: 6^th grade or higher education, fluent in English or Spanish, minimum Mini-Mental State Examination (MMSE) of 24, Clinical Dementia Rating of 0.5, subjective memory complaint by subject or study partner, impaired episodic memory, and sufficiently preserved general cognition and functional performance not meeting criteria for AD. Participants with missing data for any variables of interest were excluded, as complete data are necessary for PLSR analysis. One participant classified as MCI with an MMSE score of 23 and one participant with an extremely high and improbable triglycerides value (2,084.0 mg/dL) were excluded. Other ADNI cohorts (ADNIGO, ADNI2, and ADNI3) were excluded from the analysis because these data sets did not include numerous biomarkers of interest (e.g., neurotrophic and growth factors were not available).

The final sample included 123 participants who had composite neuropsychological data available at both baseline and two-year follow-up (age: range = 55.1 –88.3, M = 73.9, SD = 7.6; education: range = 6 –20 years, M = 16.0, SD = 3.0; 42 females and 81 males; 119 White, 2 Asian, 2 Black; 121 Non-Hispanic, 2 Hispanic; 54 APOE ɛ4 negative). Longer follow-up times were not assessed due to participant attrition. That is, cognitive composite scores were available in N = 97, N = 54, and N = 39 participants at 36-, 48-, and 60-month timepoints. Previous studies have demonstrated that significant cognitive decline can occur in MCI at one-year [26] or two-year follow-up [27, 28], and the larger dataset available at 24-month follow-up provided greater generalizability of our study sample to the broader MCI population. Cognitively normal older adults were not included in the present study due to a small sample size with biomarkers and neuropsychological data available at baseline and cognitive data at two-year follow-up (N = 34). These participants were excluded to improve the interpretability of the results, because our goal was to identify the physiological variables important for predicting cognitive change in MCI participants. Study procedures were approved by site-specific Institutional Review Boards, and all participants and/or authorized representatives provided written informed consent consistent with the Declaration of Helsinki.

Neuropsychological assessment

Composite scores of episodic memory and executive function published by ADNI were examined [29, 30]. For the episodic memory composite score, factor analytic methods and item response theory were used to create a score obtained as a weighted set of tests from the ADNI neuropsychological battery (for further details, see [29]). This memory composite score was determined to have good validity [29]. For the executive function composite score, the same statistical techniques were employed, and this score was also determined to have good validity (for further details, see [30]).

Cardiometabolic, stress, and inflammation variables

BMI (kg/m²), systolic blood pressure (mmHg), diastolic blood pressure (mmHg), pulse rate (per minute), cholesterol (mg/dL), triglycerides (mg/dL), and serum glucose (mg/dL) data were obtained. Insulin (uIU/mL), cortisol (ng/mL), CRP (μg/mL), and interleukin-6 (IL-6) receptor (ng/mL) data were assessed from fasting plasma blood samples. Data were normalized and checked for the defined least detectable dose during the quality control process.

Growth factors and neurotrophic factors

Insulin-like growth factor binding protein (ng/mL), epidermal growth factor (pg/mL), heparin binding epidermal-growth-factor-like growth factor (HB-EGF-like-GF; pg/mL), hepatocyte growth factor (ng/mL), platelet-derived growth factor BB (PDGF; pg/mL), BDNF (ng/mL), and vascular-endothelial growth factor (pg/mL) were analyzed. Data were normalized and checked for the defined least detectable dose during the quality control process.

AD biomarkers

Plasma Apolipoprotein E (apoE; μg/mL), plasma tau (pg/mL), CSF t-tau (pg/mL), CSF p-tau₁₈₁ (pg/mL), and CSF Aβ₄₂ (pg/mL) were assessed. Only values within the given ranges, which were based on the reported technical limits for each respective assay, were included in the analyses: Aβ₄₂: 200–1700 pg/mL; p-tau₁₈₁: 8–120 pg/mL; t-tau: 80–1300 pg/mL.

Data processing and analysis

ADNI1 data were scrubbed using the R-language. Raw data files for all blood- and CSF-based biomarkers were checked for imputed values published by ADNI. To ensure data integrity, all analytes with > 10% imputed values were removed. Participants who had missing data or invalid data as indicated by the ADNI manual were excluded.

PLSR R-scripts and functions were developed by co-author H. Abdi and run in R (Version 3.6.1). Several R-packages were used for data analysis (‘PTCA4CATA’, ‘data4PCCAR’, ‘ExPosition’ for PLSR functions, and ‘boot’ for bootstrapping PLSR results) and visualization (‘corrplot’, ‘tidyverse’, ‘ggplot2’). PLSR combines techniques from principal components analysis and multiple linear regression [20]. The goal of a PLSR analysis is to decompose the X-matrix (predictor variables) to create orthogonal components (i.e., optimal linear combinations of the data) that can best predict the outcome (Y-matrix) variable(s). This is done by first creating a correlation matrix between X- and Y-matrices. Next, this covariance matrix is decomposed to generate the X-matrix components. This process generalizes principal components analysis, where the X-matrix is decomposed into optimal linear combinations that are only relevant for the X-matrix, not for both the X- and Y- matrices. Finally, the X-matrix components created from the decomposition of this correlation matrix are used to predict the outcome (Y-matrix) [20]. This made PLSR advantageous over principal components analysis for the present study objectives. Additionally, PLSR was preferred over multiple linear regression because it works well with a large number of multicollinear predictor variables, and uses this multicollinearity to its advantage in its prediction [18, 20]. In multiple linear regression, multicollinear predictor variables are problematic for calculation of regression coefficients and must be removed. However, in PLSR, the researcher can explore how different groups of variables cluster together among different components of the PLSR model. Although PLSR can be run with multiple Y-outcome variables [18, 20], this did not align with the objective of the current study, because our goal was to assess how different physiological variables were weighted and predictive of each cognitive domain (episodic memory and executive function). Therefore, separate PLSR analyses were run per cognitive domain where either slope of change in composite memory or executive function over two years was the Y-matrix, or outcome variable. For both models, physical, health, and AD biomarker variables measured at baseline were the X-matrix (see Table 1 for full list of variables). This approach was implemented to examine how contributions of baseline physiological variables differed in predicting slope of change in episodic memory or executive function over two years.

To create the PLSR models, the X- and Y-matrix variables (where the X-matrix columns were 29 cardiometabolic, inflammatory, neuroprotective, AD biomarker, and demographic variables and the Y-matrices columns were, respectively, change in episodic memory or executive function) were normalized to have a mean of zero and standard deviation of one. Demographic variables (age and education) were entered as continuous variables. A leave-one-out cross-validation procedure (LOO, also called a jackknife) was completed to prevent overfitting and improve generalizability of the PLSR model [20, 21]. In this procedure, each observation is removed in turn from the X and Y-matrices, and a PLSR model is re-created for each of the remaining observations. The model created by the remaining observations and the variables is used to predict the left-out observation. The predicted observations are stored in a new matrix and a “predicted residual estimated sum of squares” (PRESS) value is created to measure the quality of the prediction, where a lower PRESS value indicates a better prediction. This procedure is repeated and iteratively uses a different number of components (from one to the number of X-matrix predictor variables), and the PRESS value is calculated for each possible number of components [20, 21]. The optimal number of components for the PLSR model is chosen based on a pseudo-scree plot (where the inertia extracted by each of the 29 possible components is plotted using the R-function ‘PlotScree’ from PTCA4CATA package) in which the optimal number of components is indicated by an inertia value greater than 1 divided by the number of components (see the Supplementary Material for pseudo-scree plots). Thus, components with an inertia value greater than 0.034 (1/29 = 0.034) were retained in the PLSR model. Next, the PLSR model is run using this optimal number of components.

PLSR whole-model outcomes

Component-specific outcomes (post-hoc analyses)

Individual components of the PLSR model were also analyzed and interpreted. To do this, circles of correlation were produced to explore how the physiological variables were correlated with each component. Each predictor variable is then plotted on a standard coordinate plane created by two components (i.e., a variable positively correlated with both Components 1 and 2 would be in the upper right quadrant of the plot). Circles of correlation also show which predictor variables correlate with one another/cluster together. Additionally, participant observation maps were created to examine the extent to which individual participants were correlated with each component. Participants were either coded by gender or diagnosis at two-year follow-up (where ADNI coded whether participants remained MCI or converted to AD). Bootstrapping (with N = 1,000 iterations) was used to calculate bootstrapped means and bootstrap ratios (BSRs) for each group per component [35]. BSRs indicate whether the group mean, on average, had values significantly different from 0 (|BSR| ≥2.0, corresponding to p < 0.05) on a given component. Bootstrapped means and 95% CIs (95% CIs are represented as ellipses plotted around each bootstrapped mean) per group (either male/female or remained MCI/converted to AD at two-year follow-up) were plotted for each graph. Groups are considered to significantly contribute to a given component if their ellipses do not cross 0. If the ellipses of two groups do not overlap along a given axis, then these two groups are significantly different from one another for the given component.

Whole-model outcomes

Episodic memory

The optimal number of components to include in the model was two (see Supplementary Figure 1A). The PLSR model accounted for 20.9% of variance of the X-matrix, which explained 20.0% of the variance in slope of change in episodic memory over two years (Model R² = 0.200; Fig. 1A). CSF Aβ₄₂ (B_PLS = 0.15; 95% CI = [0.01 –0.29]), CSF t-tau/Aβ₄₂ (B_PLS=–0.11; 95% CI = [–0.19 ––0.03]), and education (B_PLS = 0.17; 95% CI = [0.01 –0.34]) had B_PLS values that were reliable (B_PLS 95% CI did not include 0; see Supplementary Table 1 for B_PLS values for all variables). CSF Aβ₄₂ and education were both positively associated with slope of change in episodic memory over two years while CSF t-tau/Aβ₄₂ ratio had a negative association (based on their positive and negative B_PLS values). CSF Aβ₄₂, CSF t-tau/Aβ₄₂, and education were also considered important (VIP > 1) to the model, along with several other variables: CSF t-tau, CSF p-tau₁₈₁, CSF p-tau₁₈₁/Aβ₄₂, APOE ɛ4 allele status, BDNF, VEGF, and IL-6 receptor (VIP scores for all variables displayed in Fig. 1A). This model identified a pattern in the data suggesting that higher levels of CSF Aβ₄₂, lower CSF t-tau/Aβ₄₂ ratio, and higher education were reliable and positive predictors of slope of change in episodic memory performance, while baseline levels of multiple AD biomarkers, growth factors, education, and an inflammatory marker were important to the overall PLSR episodic memory model prediction.

OPEN IN VIEWER

Fig. 1

VIP Scores represent whether a given input variable is an important predictor in the PLSR model (across all components). Variables with a VIP score > 1 (highlighted by the black horizontal line at VIP value = 1) are considered important to the PLS model’s prediction. A) VIP scores for all variables entered in the episodic memory model. B) VIP scores for all variables entered in the executive function model. BDNF, brain-derived neurotrophic factor; BMI, body mass index; BP, blood pressure; CSF, cerebrospinal fluid; HB-EGF-like-GF, heparin-binding epidermal growth factor like-growth factor; IL-6, interleukin-6; IGF, insulin-like growth factor; p-tau, phosphorylated tau₁₈₁; t-tau, total tau; VEGF, vascular endothelial growth factor.

Component-specific outcomes and post-hoc analyses

Episodic memory

The variance explained per component was, respectively, 13.1% and 7.9%. In Fig. 2A, Components 1 and 2 are shown, respectively, on the horizontal and vertical axes. Input variables to the right of the vertical axis are positively correlated with Component 1, while variables to the left of the vertical axis are negatively correlated with Component 1. Variables above the horizontal axis are positively correlated with Component 2, while variables below the horizontal axis are negatively correlated with Component 2. Input variables are mapped on the circle (which is a standard coordinate plane) based on their correlation values with Components 1 and 2. As visualized in Fig. 2A, Component 1 was largely associated with AD biomarkers, while Component 2 was largely associated with neurotrophic/growth factors, IL-6-receptor, and education. Specifically, CSF t-tau, CSF p-tau₁₈₁, CSF t-tau/Aβ₄₂, CSF p-tau₁₈₁/Aβ₄₂, and positive APOE ɛ4 allele status were negatively correlated with Component 1, while levels of CSF Aβ₄₂ and negative APOE ɛ4 allele status were positively correlated with Component 1. BDNF, VEGF, epidermal growth factor, and PDGF were negatively correlated with Component 2, while IL-6 receptor and education were positively correlated with Component 2. Thus, Component 1 represents how AD pathology contributes to the model, while Component 2 represents how education and neuroprotection contribute to the model. Figure 2A illustrates several AD biomarkers and several neurotrophic factors clustering together, showing that these variables are highly correlated with one another, but contribute to the model differentially.

OPEN IN VIEWER

Fig. 2

Circle of correlation per model. Figures show how each individual input variable (labelled 1 –29; see figure legend) is correlated with Components 1 (x-axis) and 2 (y-axis) of each model. Each variable is plotted in a standard coordinate plane. For instance, education (purple #20, A) is positively correlated with both Components 1 and 2, which places it on the right, upper half of the coordinate plane. Education was more highly correlated with Component 2, as shown by nearer distance of #20 from the y-axis, relative to the x-axis. A variable that is perfectly correlated with a given component would touch the edge (gray line) of the circle. A) Episodic Memory: AD biomarkers (orange) are largely associated with Component 1, and neurotrophic factors (blue) and education (purple #20) are largely associated with Component 2. B) Executive Function: AD biomarkers (orange) are associated with both Components 1 and 2, while cardiometabolic variables and HB-EGF-like-GF (#16) are largely associated with Component 2. BDNF, brain-derived neurotrophic factor; BMI, body mass index; BP, blood pressure; CSF, cerebrospinal fluid; HB-EGF-like-GF, heparin-binding epidermal growth factor like-growth factor; IL-6, interleukin-6; IGF, insulin-like growth factor; p-tau, phosphorylated tau₁₈₁; t-tau, total tau; VEGF, vascular endothelial growth factor.

Next, BSRs and bootstrapped group means were calculated to explore group comparisons between 1) older adults who remained MCI versus converted to AD at two-year follow-up and 2) males and females on Components 1 and 2 of the episodic memory model. Post-hoc analyses revealed that older adults who converted to AD at two-year follow-up were significantly and negatively associated with Component 1 (BSR_{AD _ Comp1} = –2.98; p = 0.01; Fig. 3A) but not Component 2 (BSR_{AD _ Comp2} = –1.78; p = 0.07; Fig. 3A). Older adults who remained MCI at two-year follow-up were not significantly associated with Component 1 (BSR_{MCI _ Comp1} = 1.69; p = 0.09; Fig. 3A) or Component 2 (p > 0.05; Fig. 3A). Together, as shown by the ellipses plotted in Fig. 3A, older adults who remained MCI were significantly different from those who converted to AD. Females were significantly and negatively associated with Component 1 (BSR_{FshitasenComp1} = –2.24; p = 0.02; Fig. 3B) but not Component 2 (p > 0.05; Fig. 3B). Males were not significantly associated with Component 1 or 2 (all ps > 0.05; Fig. 3B). Males and females were not significantly different from one another in this model, as shown by the overlapping ellipses in Fig. 3B. Therefore, females and those who converted to AD at two-year follow-up were significantly and negatively associated with Component 1, which was largely associated with AD pathology.

OPEN IN VIEWER

Fig. 3

Maps of participant loadings for Components 1 and 2. A) Episodic Memory Model: participants coded by diagnosis at two-year follow-up. B) Episodic Memory Model: participants coded by gender. C) Executive Function Model: participants coded by diagnosis at two-year follow-up. D) Executive Function Model: participants coded by gender.

Episodic memory

Results demonstrated that AD biomarkers were important predictors of slope of change in episodic memory at two-year follow-up. Specifically, our finding that higher CSF Aβ₄₂ was a positive predictor of change in episodic memory is consistent with prior work showing that levels of CSF Aβ₄₂ predicted change in memory over two years in those with early and late MCI [36], as well as research showing that older adults with lower CSF Aβ₄₂ had steeper declines in memory over a three-year follow-up [37]. This result could be due to underlying structural changes in the brain, evidenced by prior work showing that lower CSF Aβ₄₂ at baseline was associated with increased rates of brain atrophy and cortical thinning [38], as well as significant changes in brain volume over one-year follow-up in MCI subjects [39]. Additional work has shown that brain atrophy was a significant mediator between baseline levels of CSF Aβ₄₂ and memory declines in older adults with early and late MCI, an effect suggesting that brain atrophy, specifically in the medial temporal regions of the brain, may be the mechanism underlying changes in cognition associated with baseline levels of CSF Aβ₄₂ [36]. This finding also supports our prior work showing that CSF Aβ₄₂ was a positive predictor of multiple domains of cognition, including episodic memory, cross-sectionally [19].

Additionally, the CSF t-tau/Aβ₄₂ ratio was a negative predictor of change in episodic memory. To our knowledge, this may be one of the first reports linking CSF t-tau/Aβ₄₂ ratio to longitudinal change in objective assessment of episodic memory performance in MCI. These data strengthen the findings of a recent cross-sectional study, which demonstrated that older adults with a higher CSF t-tau/Aβ₄₂ ratio had worse verbal episodic memory performance [7], as well as earlier work showing that a higher CSF t-tau/Aβ₄₂ ratio predicted change in cognition as assessed by the clinical dementia rating scale [40]. Other research exploring plasma rather than levels of CSF t-tau and CSF Aβ₄₂ in those with amnestic MCI found that using both of these biomarkers led to better predictions of current and future episodic memory performance than exploring either alone, although they did not explore the ratio [41]. Multiple studies have shown that CSF t-tau/Aβ₄₂ ratio is a reliable predictor of conversion from MCI to AD [42], including one study which found that older adults who converted from MCI to AD at two-year follow-up had higher levels of CSF t-tau/Aβ₄₂ at baseline [43]. CSF t-tau/Aβ₄₂ ratio may represent a potentially promising biomarker for identifying older adults with MCI who are most at risk for a decline in episodic memory.

CSF t-tau, CSF p-tau₁₈₁, CSF p-tau₁₈₁/Aβ₄₂, and APOE ɛ4 allele status were additional variables related to AD pathology that were considered important (VIP > 1) but not reliable (B_PLS values not significantly different from zero at p < 0.05) predictors in the episodic memory model. This result suggests when explored together, these variables were important but not independent predictors of change in episodic memory. Research has established that these variables are important in exploring trajectories of normal aging to AD [44, 45], and APOE ɛ4 allele status is a well-known predictor of cognitive decline. Additionally, prior cross-sectional work in our laboratory revealed that CSF t-tau and CSF p-tau₁₈₁ were negative predictors of multiple cognitive domains including episodic memory [19]. However, our results revealed that when these AD-related variables were explored simultaneously, CSF t-tau, CSF p-tau₁₈₁, CSF p-tau₁₈₁/Aβ₄₂, and APOE ɛ4 allele status were relatively inferior predictors of change in episodic memory when compared to CSF Aβ₄₂ and CSF t-tau/Aβ₄₂.

Post-hoc analyses of episodic memory revealed that participants who were female or converted to AD at two-year follow-up significantly and negatively loaded on Component 1, which was largely correlated with AD biomarkers indicative of worse pathology (Fig. 2A). This pattern is consistent with current literature exploring AD biomarkers across the cognitively normal to AD spectrum, which has found that CSF Aβ₄₂ declines first, then CSF t-tau and p-tau₁₈₁, followed by changes in memory [36, 45]. Thus, one could speculate that participants who negatively loaded on Component 1 may have been at a higher risk of converting to AD at two-year follow-up due to their baseline AD biomarker values, which was potentially followed by brain atrophy leading to steeper declines in memory and thus a diagnosis of AD at follow-up [36, 39]. These post-hoc results are also consistent with work showing that in older adults with MCI, females with levels of CSF Aβ₄₂ and CSF t-tau indicative of worse pathology had steeper declines in memory as well as hippocampal volume when compared to men [46]. This suggests that women may have a higher likelihood of converting to AD than men, even with similar levels of pathology.

In addition to AD biomarkers, education was considered an important (VIP > 1) and reliable positive (B_PLS significantly different from zero at p < 0.05) predictor of slope of change in episodic memory. This finding supports recent work showing that higher education positively impacted future memory and language declines in White older adults without dementia at baseline, as our sample was predominately White [47]. Prior work in our laboratory exploring cross-sectional associations between education and cognition using PLSC also showed that education was positively related to multiple cognitive domains including episodic memory [19]. This result also aligns with multiple studies, including a meta-analysis, that show that lower levels of education put participants at greater risk for developing dementia [48, 49]. Several studies have shown that education may be used as a proxy of cognitive reserve [48, 50]—defined as an individual’s characteristics that preserve cognitive function despite age-related or pathology-related brain changes in aging [51]. However, education is one of many factors that can contribute to an individual’s cognitive reserve [49]. Thus, our finding that education was positively associated with trajectories of episodic memory performance suggests that older adults with higher levels of education have attenuated decline in episodic memory over two years, even at the MCI stage—an effect which may be due to higher education increasing cognitive reserve.

Neurotrophic/growth factors BDNF and VEGF were important (VIP > 1) but not reliable (B_PLS values not significantly different from zero at p < 0.05) predictors of episodic memory. Thus, BDNF and VEGF were not reliable predictors of change in episodic memory but are important enough to the model that removing these two markers would affect the model prediction. Recent work has suggested that plasma levels of BDNF might not be useful to explore in MCI, because plasma BDNF levels could not distinguish between healthy older adult and MCI groups, and did not correlate with changes in multiple cognitive domains including memory and executive function [52]. The current results were inconsistent with these findings, as plasma BDNF was important in predicting episodic memory change in MCI. But this finding does align with our prior analysis using PLSC, which revealed that plasma BDNF was significantly and positively associated with cognition cross-sectionally. Other cross-sectional work has shown that lower serum levels of BDNF were associated with memory impairment but not executive dysfunction in community dwelling older adults [8]. Our results are consistent with those findings and further strengthen and extend this pattern to longitudinal changes, as we found that BDNF was important in predicting change in episodic memory but not executive function. Prior work has also shown that VEGF was one of six biomarkers in a profile that distinguished AD participants from cognitively normal participants and correlated with current but not future severity of cognitive impairment [11]. Other work has shown that increased VEGF levels in AD were associated with poorer memory and executive function, but only in APOE ɛ4 carriers [53]. Thus, our finding that VEGF was an important but not a reliable predictor of episodic memory in MCI could be due to differences between APOE ɛ4 genotype or its relevance to cognition in AD, which may create differences in findings for those with MCI due to AD rather than MCI due to another type of dementia [53]. Given the conflicting results regarding plasma BDNF and VEGF for predicting cognition in MCI, it is clear that additional studies are needed to clarify the relationships between these growth factors and decline in episodic memory.

Finally, IL-6 receptor was considered an important (VIP > 1) but not reliable (B_PLS value did not significantly differ from zero at p < 0.05) predictor of slope of change in episodic memory at two-year follow-up. This is consistent with research exploring CSF levels of IL-6-receptor in ADNI, which found that three of 83 assessed CSF proteins were significant predictors of global cognition in MCI patients at four-year follow-up [54]. However, IL-6-receptor was no longer significant after controlling for multiple comparisons. Similarly, our results show that plasma IL-6 receptor is one of the 12 important variables out of 29 predictors of change in episodic memory, but did not have a reliable B_PLS value, a pattern indicating that plasma IL-6 receptor was not a reliable predictor on its own. Other work has shown decreases in serum levels of IL-6 receptor in AD patients compared to healthy controls, suggesting that higher levels IL-6-receptor may be beneficial [55]. However, few studies have explored plasma levels of IL-6-receptor in the context of predicting episodic memory or global cognition, highlighting the need to explore this variable further.

Executive function

Our findings demonstrated that AD biomarkers were important and negative predictors of slope of change in executive function at two-year follow-up. Specifically, lower CSF t-tau/Aβ₄₂ ratio was the only reliable and positive predictor of change in executive function (B_PLS significantly different from 0 at p < 0.05). Similar to our finding in the episodic memory model, we believe this is one of the first reports linking CSF t-tau/Aβ₄₂ ratio to longitudinal change in objective assessment of executive function performance in MCI. These results support recent work which revealed that in those with late MCI, lower levels of CSF Aβ₄₂ and higher CSF t-tau (which would produce a higher ratio) were significant predictors of change in executive function, with lateral temporal lobe atrophy mediating this relationship [36]. Additionally, recent work in those with subjective cognitive decline demonstrated that lower baseline levels of CSF Aβ₄₂ and higher CSF t-tau were associated with steeper declines in tests of attention and executive functions [56]. Cross-sectional work exploring both CSF t-tau and Aβ₄₂ found that older adults with MCI that had lower CSF t-tau and higher CSF Aβ₄₂ had better performance in attention tasks, which is a component of executive function [57]. Prior research has shown that CSF t-tau/Aβ₄₂ ratio in MCI subjects was significantly different between normal and control participants, was significantly and negatively related to longitudinal changes in brain volume in age-vulnerable regions [39], and predicted conversion from amnestic MCI to dementia [42]. Thus, one could infer that participants with a higher baseline CSF t-tau/Aβ₄₂ ratio (indicative of worse AD pathology) might have negative slopes of change in executive function at two-year follow-up due to being in the later stage of MCI [36], decreases in brain volume [39], or being at greater risk of converting to AD [42].

CSF Aβ₄₂, CSF t-tau, CSF p-tau₁₈₁, CSF p-tau₁₈₁/Aβ₄₂, and APOE ɛ4 allele status were additional variables related to AD pathology that were considered important (VIP > 1) but not reliable (B_PLS values not significantly different from zero at p < 0.05) predictors for executive function. As noted in the episodic memory model discussion, prior research has shown that these variables were important predictors of change from cognitively normal to AD status [44, 45]. Additionally, as discussed in the episodic memory model discussion, prior work in our lab using PLSC indicates that CSF Aβ₄₂, CSF t-tau, and CSF p-tau₁₈₁ are significantly associated with several cognitive domains including executive function cross-sectionally [19]. However, only one executive function task (Trail Making Test Trail B) was significant in this prior analysis, an effect which is not as robust as using the ADNI executive function composite score employed for the present study. The present findings show that, when explored simultaneously, CSF t-tau/Aβ₄₂ is a better predictor of executive function than the other explored AD-related variables (CSF Aβ₄₂, CSF t-tau, CSF p-tau₁₈₁, CSF p-tau₁₈₁/Aβ₄₂, and APOE ɛ4 allele status).

The executive function model revealed that HB-EGF-like-GF was important (VIP > 1) but not reliable (B_PLS value not significantly different from zero at p < 0.05) predictor of executive function at two-year follow-up. Interestingly, our previous cross-sectional study showed that HB-EGF-like-GF was positively associated with multiple cognitive domains, including executive function, in a cohort of participants with MCI, a pattern suggesting potential neuroprotective benefits [19]. Very few studies have explored HB-EGF-like-GF in relation to human aging and cognition [17, 58]. In one study exploring cognitively normal and MCI participants, higher levels of CSF HB-EGF-like-GF were associated with greater AD pathology, demonstrated by lower levels of CSF Aβ₄₂ and higher levels of CSF t-tau [58], which conflicted with our prior finding. Other studies exploring mice supported this relationship, demonstrating that HB-EGF-like-GF caused a cascade of cellular events leading to neuroinflammation and subsequent increases in Aβ in the brain [59, 60]. However, other work exploring mice has shown that lack of HB-EGF-like-GF was associated with decreased neurogenesis in the hippocampus, which aligns with our prior cross-sectional finding that plasma HB-EGF-like-GF may have a neuroprotective effect [61]. The current study showed that HB-EGF-like-GF was not a reliable predictor of executive function on its own, although it was important to predicting change in executive function in this MCI cohort. Unfortunately, the directionality of this relationship cannot be probed further, as the B_PLS value was not significantly different from zero, and thus cannot be interpreted. The available literature has demonstrated that it is possible that plasma HB-EGF-like-GF contributes to cognitive changes in humans, but future human subjects research should further explore this biomarker to assess its potential role in predicting cognitive change.

Modifiable cardiometabolic variables including diastolic blood pressure, systolic blood pressure, and cholesterol were important (VIP > 1) but not reliable (B_PLS values not significantly different from zero at p < 0.05) predictors of change in executive function at two-year follow-up. This is consistent with research showing that metabolic syndrome symptoms as a whole, rather than individual symptoms, predicted small but negative effects on executive function and language but not episodic memory in community dwelling older adults [62]. In terms of individual cardiometabolic variables, a recent longitudinal population-based study of those ages 5 to 95 years found that mean systolic blood pressure over time was associated with significant declines in executive function at follow-up (average of 12.4 years; [63]).

Post-hoc analyses of executive function revealed that participants who were female or converted to AD at two-year follow-up significantly and negatively loaded on Component 1 (Fig. 2B). AD biomarkers indicative of worse pathology were approximately equally correlated with both Components 1 and 2 of this model. However, post-hoc analyses for gender and diagnosis at two-year follow-up were not significant for Component 2. Thus, there was not a clear interpretation for a significant and negative loading on Component 1 (females and those converting to AD at two-year follow-up), and as such we will not interpret these results further in order to avoid an over-interpretation of the data.

Limitations and strengths

The present study had several limitations. First, the ADNI sample lacks racial diversity and is highly educated—a configuration which may limit generalizability. This work also did not explore a few biomarkers of interest such as IL-6 (rather than IL-6 receptor) and tumor necrosis factor-alpha due to data either not being available in the ADNI sample or due to high amounts (> 10%) of imputed biomarker values in the MCI sample who had neuropsychological data available. The current report did not explore other modifiable cardiometabolic/metabolic syndrome variables of interest in aging such as high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, waist circumference, body fat percentage, etc., as these data were not available in this cohort. Bidirectional associations (e.g., does cognitive decline precede change in a given modifiable variable?) were not assessed because several of the biomarkers examined at baseline were either only assessed at baseline (cardiometabolic variables such as triglycerides, cholesterol and serum glucose and AD biomarker plasma tau) or were only assessed at baseline and 12-month follow-up but not 24-month follow up (inflammatory markers such as insulin, interleukin-6 receptor, c-reactive protein, and neurotrophic/growth factors such as BDNF and VEGF). However, these variables are likely modifiable and not simply a consequence of disease progression as several large population-based studies have shown that higher levels of cardiovascular, metabolic, and inflammatory variables were predictors of memory decline [64], executive function decline [65], and incident dementia at follow-up [64]. In the few studies that have assessed bi-directional relationships between modifiable variables and cognition in older adults, higher BMI predicted longitudinal declines in cognition, but cognition did not predict change in BMI [66]. Lastly, the present study had a relatively small sample size (N = 123) and a large number of predictors (I = 29), which may limit the generalizability of the current findings to the broader MCI population. In particular, important but less reliable predictors such as cardiometabolic variables and neurotrophic/growth factors that exhibited domain-specific findings for episodic memory and executive function respectively, may not generalize to the broader MCI population.

Overall, the current study had multiple strengths. A unique multivariate approach, PLSR, was implemented, which allowed for the simultaneous exploration of several physiological and demographic variables of interest without violating traditional statistical assumptions needed for other multivariate analyses such as multiple linear regression. This approach is new relative to much of the extant literature, which examines a single variable or class of variables and examines global cognition or a single cognitive domain. This work also explored longitudinal cognitive change in MCI, where older adults are at greatest risk of converting to dementia.

Conclusions

In summary, the simultaneous assessment of AD biomarkers, neurotrophic/growth factors, cardiometabolic variables, and demographics revealed that AD biomarkers were important and reliable negative predictors of change in both episodic memory and executive function over two years in older adults with MCI. Specifically, CSF Aβ₄₂, CSF t-tau/Aβ₄₂, and education were reliable predictors of change in episodic memory whereas CSF t-tau/Aβ₄₂ was a reliable predictor of change in executive function. There was also evidence of domain specificity. Variables associated with neuroplasticity and cognitive reserve were important for change in episodic memory, whereas variables associated with cardiometabolic health were important in predicting executive function. These findings provide preliminary evidence for the suggestion that interventions targeting neurotrophic/growth factors and cognitive reserve may be relatively more effective in slowing decline in episodic memory, whereas interventions targeting cardiometabolic health variables may be relatively more effective in slowing decline in executive function. Together, variables associated with neuroplasticity and modifiable cardiometabolic health variables were important for predicting change in episodic memory and executive function respectively but were not reliable predictors on their own. AD biomarkers, however, were reliable predictors on their own for both cognitive domains, and thus may carry relatively more importance in predicting these two cognitive outcomes in older adults with MCI, which should be considered when designing future research.