Participants

A total of 5126 participants (1052 participants were mild cognitive impairment and the rest were normal cognition) aged 55 and older were from the Beijing Aging Brain Rejuvenation Initiative (BABRI) database, which is an ongoing community-based prospective cohort study in China [25]. All participants in the present study met the following criteria: no dementia; were native Chinese speakers; were retired (female for 55 years old, male for 60 years old); and had no history of neurologic, psychiatric, or systemic illnesses known to influence cerebral function. They completed structural MRI scans (1117 participants, right-handed) and an extensive cognitive battery within a month (Fig. 1). All participants provided written informed consent for our protocol, which was approved by the ethics committee of the State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University.

Fig. 1

Participant flow chart. BABRI: Beijing Aging Brain Rejuvenation Initiative

The diagnostic criteria for mild cognitive impairment (MCI) based on Petersen’s criteria included (1) the presence of subjective cognitive complaints (self-reported and/or by informants); (2) normal general cognitive ability (a score higher than 23 on the Mini-Mental State Examination [MMSE], which is not the main criterion to distinguish between normal vs abnormal cognition); (3) normal daily living ability (a score of 0 in the Instrumental Activity of Daily Living [IADL] and ADL); and (4) objective cognitive impairment in at least one domain (one of the neuropsychological test scores corresponding to different cognitive domains were less than 1.5 standard deviations below the age- and education-adjusted mean of the Chinese elderly population, see Additional file 1) [26, 27]. The normal cognition participants were non-MCI and the neuropsychological test scores within 1.5 standard deviations.

The measure of lifetime intellectual cognitive reserve

An indicator of lifetime intellectual cognitive reserve (LICR) combined three measures over the life-course: early life education, midlife occupational complexity, and mental leisure activities after retirement. Education level was assessed by recording the number of years of schooling. The matrix derived for U.S. occupations using the 1970 U.S. census was used to measure the complexity of occupation [6, 14]. Each occupation was assigned a score reflecting the level of complexity at which a typical occupation, with Chinese occupation codes matched to the best-fitting category from the U.S. occupational complexity was assessed along the dimensions of data, people, and things with continuous scores ranging from simple (value 0) to complex tasks (values 6, 7, and 8, respectively). To comprehensively examine occupational complexity, the scores of the three dimensions were added together, and then the score was reversed; that is, the higher the total score is, the greater the occupational complexity (0–21) [28]. Mental leisure activities performed for enjoyment were measured using a personal information questionnaire with 23 questions that asked participants to recall their leisure activity in the past year, including reading; writing; participating in a senior citizens’ university; playing chess, poker, or mahjong; and handcrafts. The frequency of each activity was defined as frequent if they participated several times per week and rare if they participated less than once per week, and the total score was based on a weighted score of these 23 questions [24].

Neuropsychological test and personal information questionnaire

All participants underwent a battery of neuropsychological tests which were then used to assess several cognitive domains: 1) the Auditory Verbal Learning Test (AVLT, including N1N3, N4, N5 and N1N5); 2) the Rey-Osterrieth Complex Figure test (ROCF, including copy and delay recall); 3) the Digit Span Test (DST, including forward and backward); 4) the Trail Making Test (TMT, including A time, B time and B-A time); 5) the Symbol Digit Modalities Test (SDMT); 6) the Stroop Color and Word Test (SCWT, including B time, C time and C-B time); 7) the Clock Drawing Test (CDT); 8) the Category Verbal Fluency Test (CVFT, including animal, vegetable and fruit); and 9) the Boston Naming Test (BNT) [24].

The personal information questionnaire included demographic information and medical history. Demographic information included age and sex. Medical history included questions on a series of chronic diseases, including hypertension, hyperlipidemia, diabetes, coronary heart disease and cerebral small vessel disease. The number of these chronic diseases that the elderly suffered from, called disease number, as well as age and sex, were all used as control variables in subsequent statistical analyses.

MRI data acquisition

MRI data were acquired using a Siemens TRIO 3T scanner at the Imaging Center for Brain Research at Beijing Normal University and Beijing Tiantan Hospital, Capital Medical University (Beijing, China). Participants laid supine with their head fixed snugly by straps and foam pads to minimize head movement. T1-weighted, sagittal 3D magnetization prepared rapid gradient echo (MP-RAGE) sequences were acquired and used to cover the entire brain, and the scan parameters were as follows: for Beijing Normal University, 176 sagittal slices, repetition time (TR) = 1900 ms, echo time (TE) = 3.44 ms, slice thickness = 1 mm, flip angle = 9°, field of view (FOV) = 256 × 256 mm2, acquisition matrix = 256 × 256; for Beijing Tiantan Hospital, 192 sagittal slices, repetition time (TR) = 2300 ms, echo time (TE) = 2.32 ms, slice thickness = 1 mm, flip angle = 8°, field of view (FOV) = 256 × 256 mm2, acquisition matrix = 256 × 256.

DTI scans were obtained in the axial plane with a single-shot, spin-echo, echo planar sequences in the axial plane, and the scan parameters were as follows: for Beijing Normal University, a total of 70 slices covered the entire hemisphere and brainstem without a gap. TR = 9500 ms, TE = 92 ms, FOV = 256 × 256 mm2, matrix = 128 × 128, slice thickness = 2 mm; for Beijing Tiantan Hospital, a total of 75 slices covered the entire hemisphere and brainstem without a gap. TR = 8000 ms, TE = 60 ms, FOV = 256 × 256 mm2, matrix = 128 × 128, slice thickness = 2 mm. To increase the signal-to-noise ratio, two or three repetitions were performed. Diffusion sensitization gradients of each scan were applied in 30 non-collinear directions with a b value of 1000 s/mm2 and a non-diffusion weighted direction with a b value of 0 s/mm2.

Macro- and micro-structural MRI data processing

T1-weighted images were processed using the Computational Anatomy Toolbox (CAT12, http://dbm.neuro.uni-jena.de/cat12/) of Statistical Parametric Mapping software (SPM12, http://www.fil.ion.ucl.ac.uk/spm) implemented in MATLAB (R2014a). First, the NIfTI files converted from the raw DICOM T1-weighted images were segmented and spatially normalized into gray matter (GM), white matter (WM) and cerebrospinal fluid in standard MNI space using optimized shooting algorithm [29]. After segment, the total intracranial volume (TIV) and the volume of gray matter (GMV)/white matter (WMV) were estimated from xml files for each subject which contain these raw values using the “Estimate Total Intracranial Volume” module. Finally, Gaussian kernel of 8 mm full-width-half-maximum (FWHM) was used to smooth the image of gray matter tissue components.

DTI images were processed using the Pipeline for Analyzing braiN Diffusion imAges (PANDA, http://www.nitrc.org/projects/panda/). First, the DICOM files of all subjects were also converted into NIfTI files. The preprocessing included the following steps: (1) Brain extraction, in which the parameter for extracting brain tissue was 0.25 and the cropping gap was 3 mm; and (2) Eddy current and head motion correction, which was corrected by applying an affine alignment of each diffusion-weighted image to the b = 0 image. Accordingly, the b-matrix was reoriented based on the transformation matrix. (3) The average of the three scans was taken. (4) The diffusion tensor was estimated, and calculated the fractional anisotropy (FA), which is the most sensitive index to reflect the integrity of white matter. Six elements of the 3 × 3 diffusion tensor were determined via multivariate least-squares fitting of diffusion-weighted images. The tensor was diagonalized to obtain three eigenvalues (λ1−3) and three eigenvectors (ν1−3). (5) Tract-based spatial statistics: The FA images of each subject were nonlinearly registered to the FMRIB58_FA standard space, the average FA images of all subjects were calculated, and the average FA skeleton was generated at the same time. Finally, the FA images of all subjects were projected onto the average FA skeleton for subsequent analysis. (6) Smooth: the above images were smoothed using 6 mm FWHM.

Statistical analysisBehavior data statistical analysis

First, to acquire more representative cognitive abilities, exploratory factor analysis (EFA) and confirmatory factor analysis (CFA, Mplus8.3) were performed via the cross-validation method to derive cognitive domain scores and general ability according to neuropsychological tests. Model fit effect was assessed using chi square (χ2) goodness of fit, the Comparative Fit Index (CFI), the Tucker-Lewis Index (TLI), the Root Mean Square Error of Approximation (RMSEA), and the Standardized Root Mean Square Residual (SRMR). The predicted values of the latent variables were derived via the maximum likelihood method and determined via Savedata statements based on the weights of their respective indicators.

The same confirmatory factor analysis was used to determine the lifetime intellectual cognitive reserve (LICR) score, which included the total number of years of education in the early life, the total occupational complexity score in middle life and the total frequency of mental leisure activities after retirement. Later, reliability and validity tests were carried out. Each numerical variable in the questionnaire was included in the reliability analysis; that is, the Cronbach-α coefficient was calculated. Given that we have not been able to exhaust all the proxy measures of cognitive reserve, we have not been able to compare these measures with the LICR. Previous researchers have often regarded a single indicator such as education level as a proxy index of cognitive reserve, so we compared the regression coefficients of LICR and its single cognitive reserve index on cognitive function using permutation test to test the validity of LICR. Finally, another hierarchical regression analysis was used to explore whether LICR was a significant positive predictor of each cognitive ability, and the difference in regression coefficients was tested to compare whether the protective effect was cognitive domain specific or domain general. Furthermore, logistic regression was used to explore the relationship between LICR and the risk of MCI. All of the above analyses were controlled for age, sex and disease number.

Macro-structural MRI data statistical analysis

For T1 data analysis, SPM12 and CAT12 in MATLAB were used to perform voxel-based morphometry (VBM) analysis. A general linear model was used, with LICR as the independent predictor, sex, age, disease number, MRI site (Beijing Normal University or Beijing Tiantan Hospital), and total intracranial volume (TIV) only for VBM analysis as covariates. The significance level was set to p < 0.05 after false discovery rate (FDR) correction, and the cluster size was set to > 50. In addition, to verify the relationships between whole-brain volume and cognitive reserve found in previous studies, partial correlation was used to explore the relationships between LICR and these whole-brain volume indicators, including TIV, GMV, WMV, relative GMV and WMV (GMV and WMV divided by TIV). Finally, we extracted regional gray matter volume which were significantly correlated with LICR in the above regression analysis, and partial correlation was used to explore the relationship between them and cognitive function.

Micro-structural MRI data statistical analysis

For DTI data analysis, the derived FA data were analyzed using tract-based spatial statistics (TBSS, FA threshold = 0.2) analysis and region of interest (ROI) analysis. For TBSS analysis, all the FA images of the elderly were projected on the average FA skeleton, and then the “randomize” tool in the FSL toolkit and the GLM were used for statistical analysis. The GLM statistical model used the average FA data as the statistical template, cognitive reserve as the independent variable, sex, age, disease number and MRI site as covariables to construct a design matrix in which continuous variables were centralized, and the statistical results were obtained by 5000 permutation tests. The significance level was set to p < 0.05 after threshold-free cluster enhancement (TFCE) transformation and family wise error (FWE) correction. For the ROI analysis, the JHU-ICBM-DTI-81 atlas, which included 50 fiber fasciculi, was used for the anatomic labels. According to previous studies, associative fibers were more closely related to cognitive ability and aging [30]. In the present study, 19 associative fibers of interest were selected from the above atlas, including the fornix (column and body of fornix) (FX), sagittal stratum (including the inferior longitudinal fasciculus and inferior fronto-occipital fasciculus) (SS), external capsule (EC), cingulate gyrus (CGC), cingulum (hippocampus) (CGH), fornix (cres)/stria terminalis (FX/ST), superior longitudinal fasciculus (SLF), superior fronto-occipital fasciculus (SFOF), inferior fronto-occipital fasciculus (IFOF) and uncinate fasciculus (UNC). Then, partial correlation was used to explore the relationships between the FA values of the above fibers and cognitive function, with the above factors used as control variables. The significance level was set to p < 0.05 after family wise error (FWE) correction (p = 0.0026).

To further explore whether the protection of LICR on cognitive functions occurred through a larger gray matter volume or more complete white matter microstructure, we further used structural equation model to construct a mediation model of “LICR-brain structural characteristics-cognitive function”, with the above factors used as control variables.