Gray Matter Volume as Evidence for Cognitive Reserve in Bilinguals with MCI

Abstract

Background:

Compared with monolinguals, bilinguals have later onset of mild cognitive impairment (MCI) and Alzheimer’s disease (AD) symptoms and greater neuropathology at similar cognitive and clinical levels. The present study follows from a previous report¹ showing faster conversion from MCI to AD for bilingual patients than comparable monolinguals, as predicted by Cognitive Reserve (CR).

Purpose:

Identify whether the increased CR found for bilinguals in the previous study was accompanied by greater gray matter atrophy than was present for the monolinguals.

Methods:

A novel deep learning technique based on Convolutional Neural Networks (CNN) was used to enhance clinical scans into 1 mm MPRAGEs and analyze the gray matter volume at the time of MCI diagnosis in the earlier study.

Patients:

24 bilingual and 24 monolingual patients diagnosed with MCI at a hospital memory clinic.

Results:

Bilingual patients had more gray matter loss than monolingual patients in areas related to language processing, attention, decision making, motor function, and episodic memory retrieval. Bilingualism and age were the strongest predictors of atrophy after other variables such as immigration and education were included in a multivariate model.

Discussion:

CR from bilingualism is evident in the initial stages of neurodegeneration after MCI has been diagnosed.

1. Introduction

Alzheimer disease (AD) is the most common cause of dementia and poses public health and financial burdens for society, according to the World Health Organization. To address the impact of this situation, several cognitive and lifestyle factors have been studied that potentially counter the effects of neurodegenerative diseases and improve levels of functioning CR². At present, these CR activities are the most effective mitigation against the devastating effects of dementia. One potential source of CR is lifelong bilingual experience³. Bilingualism offers an accessible approach to problems associated with aging because more than half of the world’s population is bilingual^{4, 5}.

The presence of CR in bilinguals has been demonstrated in healthy older adults⁶ and in some studies, accompanied by neuroimaging evidence while participants performed various cognitive tasks⁷. This research has been extended to investigate the potential impact of bilingualism on patients with MCI, AD, or other dementias. In AD samples, the first evidence for protective effects of bilingualism came from a study by Bialystok, Craik and Freedman⁸ in which they found that bilinguals showed first symptoms of dementia 4.1 years later than monolinguals, all other measures, such as immigration, formal education, and employment status, being equivalent. In a similar study, Craik, Bialystok and Freedman⁹ reported that bilingual patients were on average 5.1 years older than monolinguals at age of onset of symptoms of dementia. This pattern has been replicated in Belgium¹⁰, India¹¹, China¹², and elsewhere. Although not all studies have reported these results^13,14, the reliability of the pattern has been confirmed in several meta-analyses^15–17. Similar delays in the appearance of symptoms of MCI have also been reported^18,19.

In contrast to these retrospective studies documenting age of symptom onset or clinical diagnosis in a patient population, prospective studies investigate the incidence rate of disease in healthy populations over time. These results have been inconsistent, frequently indicating no difference in incidence for monolingual and bilingual groups²⁰. However, a study based on population-level samples could potentially be more revealing. If AD is delayed in individuals with a specific experience, then it is possible that a population in which that experience is prevalent may have a lower incidence of AD overall. Since AD is a disease of aging, postponing symptoms long enough will lead to death by other causes, with fewer active cases of AD in the population. To our knowledge, only one study has tested this idea. Klein, Christie and Parkvall²¹ calculated the mean number of languages spoken by the population in 93 countries. Each country was then assigned a multilingualism score that was compared to AD incidence data supplied by the World Health Organization. After controlling for wealth and life expectancy, there was a significant negative relation between AD incidence and population multilingualism, resulting in lower overall incidence in countries with more multilingualism.

Imaging studies typically show that bilinguals can withstand more brain atrophy than monolinguals for similar clinical dementia levels. Schweizer, Ware, Fischer, Craik and Bialystok²² compared computed tomography scans of monolingual and bilingual AD patients who were matched for cognitive performance and clinical scores. Monolingual and bilingual patients were the same age and clinical level, but the bilingual patients had more brain atrophy in AD associated regions in the medial temporal lobe. Similarly, Perani, Farsad, Ballarini, Lubian, Malpetti, Fracchetti, Magnani, March and Abutalebi²³ used positron emission tomography to measure brain glucose metabolism in a matched sample of monolingual and bilingual AD patients. The bilingual patients were five years older than monolinguals and had more decline in cerebral glucose metabolism than monolingual patients, indicating more advanced disease. Nonetheless, the bilingual patients outperformed monolinguals on short-and long-term verbal memory and visuospatial tasks. Combining a cross-sectional and prospective approach with MCI patients, Costumero, Marin-Marin, Calabria, Belloch, Escudero, Baquero, Hernandez, Ruiz de Miras, Costa, Parcet and Avila²⁴ reported that monolingual and bilingual patients did not differ in cognitive status but bilingual patients showed less parenchymal volume than monolinguals, consistent with more advanced disease. However, the longitudinal analysis showed higher brain atrophy rates and more cognitive decline for monolinguals than bilinguals over a 7-month period. Similarly, Duncan, Nikelski, Pilon, Steffener, Chertkow and Phillips²⁵ analyzed cortical thickness and tissue density in monolingual and multilingual MCI and AD patients. The multilingual AD patients had thinner cortex and lower gray matter density than monolingual AD patients, consistent with the results of Schweizer, Ware, Fischer, Craik and Bialystok²². However, the multilingual MCI patients had thicker cortex or higher gray matter density than monolingual patients. The authors interpreted the former result as evidence of CR and the latter as evidence of brain reserve. Both brain reserve and CR can make independent contributions to individual differences in resilience. This relationship in bilinguals awaits further investigation.

In all these imaging studies, bilinguals and monolinguals were matched on cognitive level and then brain level was evaluated. However, it is also possible to reverse this conventional approach. Participants can instead be matched on brain measures and then determine if there are differences in cognitive performance or clinical outcomes at specific levels of brain integrity. Using this approach, Berkes, Calvo, Anderson and Bialystok²⁶ matched a group of healthy bilinguals from a previous study²⁷ to monolinguals from the Alzheimer’s disease Neuroimaging Initiative database (adni.loni.usc.edu) by using white matter measures together with age, sex, and education to compare clinical status. All the bilinguals were experiencing healthy aging and performed cognitive tasks to the same level as monolinguals in the previous study, but 41% of the matched monolinguals in the new study had a diagnosis of AD or MCI. The interpretation was that bilinguals could cope better than monolinguals with those levels of brain structure without displaying clinical symptoms.

The evidence reviewed to this point indicates that bilingualism provides resilience both in healthy aging and dementia. However, according to Stern², another prediction that follows from CR is the counterintuitive hypothesis that after diagnosis, high reserve individuals with dementia will experience more rapid decline in cognitive level and clinical status than low reserve individuals. To test this possibility, Berkes, Bialystok, Craik, Troyer and Freedman¹ followed the hospital records of monolingual and bilingual patients diagnosed with MCI and compared the time for the consensus diagnosis to convert from MCI to AD. Consistent with the hypothesis, the bilingual patients converted to AD faster than monolinguals after controlling for other relevant variables.

These findings have implications for the role of bilingualism in the presentation of MCI symptoms. Bilinguals experience symptoms of MCI later in the disease progression than monolinguals but once MCI has been diagnosed, they convert to dementia more rapidly. One explanation for this finding is that the bilinguals were able to compensate for more atrophy than monolinguals and sustain normal levels of cognitive function for longer. It is possible that bilinguals have multiple and stronger routes connecting critical brain regions than monolinguals making the system more resilient to isolated damage³. However, without information about brain structure in these groups of patients, such conjectures remain speculative.

A subset of the sample in the Berkes, Bialystok, Craik, Troyer and Freedman¹ study had MRI scans completed at the time of diagnosis. However, because the images were collected for clinical purposes they were ‘thick-sliced’ and it was not possible to measure the degree of neuropathology. New machine learning approaches offer a solution to this problem. Iglesias, Billot, Balbastre, Tabari, Conklin, Gilberto Gonzales, Alexander, Golland, Edlow, Fischl and Alzheimer’s Disease Neuroimaging²⁸ developed a deep learning technique called SynthSR which allows the reconstruction of high-resolution isotropic volumes from clinical scans with spaced slices acquired with different contrasts, resolution, and orientation. Using this novel technique, it was possible to enhance the clinical scans from the study by Berkes, Bialystok, Craik, Troyer and Freedman¹ into 1 mm MPRAGEs and evaluate the role of reserve in their findings. Therefore, the present study compared the neuropathology levels of the monolingual and bilingual MCI patients in Berkes, Bialystok, Craik, Troyer and Freedman¹. After using the deep-learning convolutional neural network (CNN) technique, the synthetic images were analyzed using voxel-based morphometry (VBM) to compare local concentrations of gray matter between the two language groups. Consistent with reserve theory and previous reports in the bilingualism literature, the prediction was that bilinguals will show more gray matter atrophy than monolinguals at the time of diagnosis despite being matched on clinical cognitive level.

2. Methods

2.1. Participants

The study was approved by the York University Human Participants Review committee, certificate #2014–354. Participants were a subset of the larger sample reported in Berkes, Bialystok, Craik, Troyer and Freedman¹. The full sample consisted of 158 MCI patients who attended a memory clinic in Toronto, Canada, with cognitive complaints. All participants had been diagnosed with MCI using the NIA-AA core clinical criteria²⁹. Patient records were followed until the point that the consensus diagnosis of the medical team converted from MCI to AD. For Berkes, Bialystok, Craik, Troyer and Freedman¹. The relevant variable was the time elapsed until that conversion, a period that was significantly shorter for bilinguals than for monolinguals.

During the intake interview and subsequent neuropsychological assessment, information was also collected about occupational history, education, language history, fluency in English and other languages, place of birth, and immigration status. The primary criteria for classifying participants as bilingual were that they had knowledge of more than one language, regularly used a language other than English, and had lived in a non-English speaking country. From the 158 patients in Berkes et al., 83 were classified as monolingual and 75 as bilingual. The two language groups were further divided by sex and immigration status for the behavioral analysis.

Forty-eight of these patients, consisting of 24 monolinguals and 24 bilinguals, also had MRI images in their clinical record that were taken at the time of MCI diagnosis. The current sample includes 24 monolingual MCI patients and 23 bilingual patients because the brain image from one bilingual was unusable. Aside from English, bilingual patients spoke a variety of languages, the most common being Yiddish (26%), Hungarian (13%), Hebrew (8%), and Italian (8%).

2.2. Data Preparation

The MRI data were acquired for clinical purposes so had large inter-slice spacing. Scans were also acquired in different hospitals using MRI machines with different resolution, contrast, etc. Twelve sites contributed MRI scans which had high x and y coordinate resolution (~1 mm, i.e., high resolution in the sagittal plane), while the z dimension ranged from 3 to 7.5mm (M = 5.84, SD = 1.11). However, as described above, it was possible to turn the clinical scans into 1 mm MPRAGEs using the machine learning approach by Iglesias, Billot, Balbastre, Tabari, Conklin, Gilberto Gonzales, Alexander, Golland, Edlow, Fischl and Alzheimer’s Disease Neuroimaging²⁸. This technique achieves a level of harmonization and has been suggested as a solution to the problem of substantial differences in MR acquisition protocols in multi-site studies²⁸. Nevertheless, as a more stringent test, we also included site as a covariate of no interest in a follow-up analysis which yielded very similar results (see appendix).

SynthSR is the first deep learning technique that uses CNN trained with synthetically generated images²⁸. Convolutional neural networks are trained for super-resolution of different images received as scans with spaced slices, acquired with different contrast, resolution and orientation and producing as output an isotropic scan of canonical contrast- i.e., a 1 mm MPRAGE. Thus, the technique is specifically designed for the type of data used in the present study. Images were processed using the version of SynthSR that was released with the development version of FreeSurfer (Figure 1) and the resulting images were subjected to statistical analysis to examine language group differences. Tissue segmentation was performed on the synthetic images using the CAT12 toolbox running under Statistical Parametric Mapping, Version 12 (SPM12) in a Matlab environment with the default parameters in gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF). CAT12 image homogeneity control resulted in high estimates for all structural images. Following segmentation, normalized subject GM images were smoothed with an 8mm FWHM.

Figure 1.

MRI scans before and after SynthSR. Panel A shows the original MRI clinical scan from one bilingual MCI patient and panel B shows the same image after SynthSR.

2.3. Statistical analysis

For the descriptive statistics, t-tests were conducted for all continuous variables and a chi-square analysis was performed for gender. FSL Randomize, a nonparametric permutation test, was used to infer group differences, and age and intracranial volume (estimated by CAT12) were included as covariates. Statistical values were threshold-free cluster enhanced and FDR corrected at p < .05, using tract-based statistical analysis (TBSS). The ‘atlasquery’ tool from FSL was used to identify regions using ‘Talairach Daemon Labels’ as reference.

To examine how demographic factors that are known to influence reserve (e.g., age, bilingualism, immigration status, education) are associated with brain measures, behavioral partial-least-squares (PLS) analysis was used³⁰. For the PLS analysis, gray matter volume was considered as the dependent variable, and language group (bilingual | monolingual), immigration status (immigrant | non-immigrant), sex (male | female), education (years), and the scores for the Mini-Mental-Status-Examination (MMSE) were the independent variables.

3. Results

3.1. Descriptive statistics

The mean, standard deviation (SD), and p-values for the descriptive statistics are presented in Table 1. There were no significant group differences in age, t (45) = 1.55, p = 0.06, education, t(45) = −0.26, p = 0.60, gender, X²(1, N = 47) = 0.021, p < 0.88, or MMSE scores, t(45) = −0.87, p = 0.20. Although the statistics for age showed no significant differences between the groups, bilinguals were 3 years older than monolinguals when the first MCI symptoms were reported which is consistent with previous research showing that dementia symptoms appear around 4 years later in bilinguals^8–12.

Table 1.

Demographics and clinical status of participants

Language Group	Monolingual (n= 24)	Bilingual (n= 23)
	M (SD)	M (SD)	P
Age	75.18 (7.2)	78.59 (7.8)	0.06
Education	14.41 (3.0)	14.13 (4.2)	0.62
Gender (n female)	12	11	0.88
MMSE	25.91 (5.9)	24.34 (6.3)	0.34

3.2. Structural analysis

VBM analysis revealed significant language group differences for GM volume in several brain regions (Figure 2). Compared to monolingual MCI patients, bilingual patients showed more significant GM loss in the frontal lobe extending into parietal areas and the limbic lobe with greater involvement of the right hemisphere than the left. Greater GM atrophy for bilinguals was also shown in areas related to language processing (Brodmann area 47), episodic memory retrieval (Brodmann area 7), motor functioning (Brodmann area 4), and decision making (Brodmann area 32). For clarity, only significant peaks (p < 0.05, FDR corrected) are listed in Table 2. Thus, significant clusters of reduced GM density in monolinguals compared with bilingual MCI patients were found to primarily affect the left sub-gyral region, the right precentral gyrus, right cingulate gyrus, and right precuneus.

Figure 2.

Structural analysis showing regions where bilinguals have a higher probability of gray matter loss than monolinguals. The legend refers to 1-p values, and the figure is thresholded at 0.95. Note, this analysis controls for intracranial volume and age.

Table 2.

Clusters of voxels that showed significantly less gray matter volume (mm³) in bilinguals than monolinguals in a VBM analysis (FDR-corrected, p < 0.05). The MNI peak coordinates (x, y, and z) of maximum intensity, and cluster size are listed. Reported clusters contain a minimum of ten contiguous voxels.

Voxels	MAX	X	Y	Z	HEMISPHERE	LABEL	BA*
29264	0.978	−22.9	19.5	28.2	LH	Frontal lobe. Sub-Gyral.	47
5341	0.968	19.7	−46	43.6	RH	Superior parietal lobule. Precuneus.	7
5095	0.972	42.7	7.7	35.9	RH	Frontal lobe. Precentral Gyrus.	4
316	0.958	15.6	−17.8	41	RH	Limbic Lobe. Cingulate Gyrus.	32

^*Brodmann areas

3.3. Partial Least Squares analysis

PLS was used to extract a set of latent variables to explain the covariance between GM volume and the following variables previously associated with CR: bilingualism, age, sex, education, immigration status, and MMSE scores. The analysis yielded a single significant latent variable, p < 0.001, which explained 34% of the cross-block covariance (Figure 3). This LV showed that after age, bilingualism was the strongest factor associated with low gray matter volume.

Figure 3.

PLS results showing the correlation between bilingualism, age, sex, education, immigration status, MMSE scores and GM volumes.

4. Discussion

CR protects against cognitive decline in both healthy aging and dementia. Previous research has shown that lifelong bilingualism may contribute to CR by demonstrating later onset of AD symptoms^8–12 and greater neuropathology for bilingual patients despite having similar cognitive function as their monolingual peers^22,23. This pattern has also been reported during the first stages of neurodegeneration for MCI^18,19. This prodromal phase is crucial to investigate CR factors because it is at this stage when future dementia can be predicted. According to Stern’s CR model, patients with increased CR should convert faster to dementia, a point recently demonstrated for bilinguals¹. However, no brain data were reported in that paper so the potential reasons for earlier conversion from MCI to AD for the bilinguals remained speculative.

Using novel deep learning techniques, the present study reconstructed clinical images to analyze GM atrophy in a subset of the participants in the earlier study by Berkes, Bialystok, Craik, Troyer and Freedman¹. Results revealed significantly more gray matter loss for the bilingual MCI patients than for the monolingual patients with similar clinical levels. As shown in Figure 2, this decline in gray matter density for bilinguals was found in the frontal and middle parts of the brain, primarily in the frontal gyri, the right cingulate gyrus and the precuneus. Overall, two main brain patterns emerged: 1) greater GM atrophy for bilingual MCI patients, and 2) significant covariance between bilingualism and gray matter reduction. These findings will be discussed below.

Consider first the GM atrophy pattern found for MCI. Although the cause of MCI is still uncertain, prior work suggests that it is associated with volumetric loss, neuroinflammation, synaptic dysfunction, and vascular pathology, especially in the temporal, prefrontal, and insular cortices³¹. Research using MRI has identified the medial temporal lobe as the earliest affected area³². This is the stage in which patients typically present the first cognitive deficits. Later, polymodal association areas (prefrontal, inferior, parietal and superior temporal regions) are affected and dementia symptoms become more evident³³. During conversion from MCI to AD, the parietal cortex has been suggested as a crucial region in structural, functional, perfusion and metabolic neuroimaging studies³⁴.

In the present study, there were no language group differences in the temporal cortex, but bilinguals exhibited more gray matter loss than monolinguals in parietal regions, consistent with the neuroimaging data on conversion to dementia. Moreover, bilinguals exhibited more gray matter loss than monolinguals in some areas of the frontal cortex. Both the parietal and the frontal cortex are involved in bilingual processing³⁵. Greater atrophy for bilinguals was also seen in areas related to episodic memory, a set of processes that has also been impacted by bilingual experience³⁶. Although there were no specific memory measures obtained from these patients, performance on neuropsychological tests that incorporate some aspects of memory were equivalent for the two groups. Thus, the neuroplastic changes to these areas induced by bilingualism may make them more robust even in the presence of brain atrophy.

Atrophy patterns may be understood in terms of the variants of MCI as each subtype shows a different GM pattern. These variants, namely amnestic MCI, multiple domain MCI, and single domain amnestic or non-amnestic MCI, typically present different prognoses of progression. It is generally proposed that amnestic MCI progresses to AD while non-amnestic varieties can progress to vascular cognitive impairment or AD³⁷. Single domain non-amnestic MCI, which includes attention/executive deficits, might progress to frontotemporal dementia, or Lewy Body dementia³⁸. It has been suggested that the CR effects of bilingualism may be more specific to the single-domain amnestic MCI¹⁹. Thus, it would be useful to examine the GM atrophy patterns of MCI subtypes in our sample, but it was not possible to corroborate these distinctions due to the small sample sizes in each group. Future studies might usefully map gray matter loss in terms of these different prodromes of dementia.

There was greater volume loss in the right hemisphere than the left for bilinguals. It is believed that the left hemisphere is dominant for language functioning in right-handed adults^39, but it is not clear if this pattern of lateralization differs in bilinguals. Some studies have suggested bilateral engagement and increased interhemispheric activity for bilinguals³⁹, lesion studies have shown compensation mechanisms in the right hemisphere of bilinguals after extensive lesions⁴⁰, and neuroimaging studies in healthy adults have revealed involvement of the right hemisphere during language switching and inhibitory control³⁵. Therefore, bilingualism leads to a reconfiguration of language processing extending to areas of the right hemisphere that are not specific to linguistic function.

The PLS analysis extended these results by including the contribution of age, sex, education, and immigration to the outcomes. All these factors have been shown to be associated with CR. However, bilingualism was the second strongest predictor for MCI and GM atrophy after age. Taken together, the results indicate that bilingualism contributes to increased reserve in the initial stages of neurodegeneration after MCI has been diagnosed. Because bilinguals can withstand more neuropathology than monolinguals, they already have more atrophy at the time of diagnosis. However, at some point the accumulated neuropathology will become impossible for the system to accommodate and the cognitive decline will be more rapid. Thus, assessing GM atrophy in MCI bilingual populations offers new perspectives to better understand the role of CR in the progression of dementia.

Data availability

The imaging data reported in this study are available on request from the corresponding author (EB). The clinical data are not publicly available due to privacy/ethical restrictions.