Subjects
The data applied to the study were obtained from our in-home database: Nanjing Brain Hospital-Alzheimer's Disease Spectrum Neuroimaging Project 2 (NBH-ADsnp2) (Nanjing, China), which is continuously updated. For information on NBH-ADsnp2, see SI Methods. A total of 256 subjects were included in this study (including 89 CN, 75 SCD, and 92 aMCI). However, due to excessive head movements (cumulative translation or rotation >3.0 mm or 3.0º), 12 of these participants were excluded (3 CN, 3 SCD, and 6 aMCI). Thus, a total of 244 subjects (86 CN, 72 SCD, and 86 aMCI) were included in the study. Of these, 10 patients with SCD and 11 patients with aMCI received follow-up rTMS intervention, fMRI, and clinical cognitive data acquisition. The specific inclusion and exclusion criteria for subjects are described in the SI MethodsS.1.
Neuropsychological assessments
A standardized clinical interview and a comprehensive neuropsychological assessment were conducted in this study. We divided the comprehensive neuropsychological assessment into four cognitive domains: episodic memory (AVLT, LMT, and CFT-20-min-DR), information processing speed (DSST, TMT-A, Stoop-A, and Stoop-B), visuospatial function (CFT and CDT), and executive function (TMT-B, Stoop-C, DST-backward, VFT, and Semantic Similarity). Detailed information for each neuropsychological assessment can be found in the supplement materials (SI MethodsS.2).
MRI data acquisition
All magnetic resonance imaging (MRI) data were acquired at the Affiliated Brain Hospital of Nanjing Medical University using a 3.0 Tesla Verio Siemens scanner and an 8-channel head coil. Detailed information on the parameters of the image acquisition is provided in the supplement materials (SI MethodsS.3).
fMRI data preprocessing
All fMRI data were preprocessed using MATLAB 2013b and DPABI software[29]. Discard the first 10 volumes to reduce the instability of the MRI signal. Then we perform slice timing correction and head motion correction[30,31]. The functional image of each individual is aligned with the structural image. Next, to spatially normalize the fMRI data, we segmented the T1 images into grey matter (GM), white matter (WM), and cerebrospinal fluid (CSF) using the Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) algorithm[32]. Then, functional images were normalized by DARTEL into MNI space (resampling voxel size, 3×3×3 mm3). Finally, spatially smoothed by a Gaussian kernel of 6 mm3 full-width at half maximum (FWHM) to reduce spatial noise. Independent component analysis (ICA) was performed on the pre-processed data.
Group independent component analysis and regions of interest
Using the Group ICA toolbox (http://icatb.sourceforge.net/), we divided the pre-processed data into 27 independent components. Afterward, we spatially correlated these 27 independent components with the corresponding network templates provided by Smith et al[33]and Xue et al[34]. According to the results of the correlation calculation, the independent components were sorted, and the independent components that best fit DMN and CEN templates were selected.
After selecting the independent components of the DMN and CEN, we used the XjView toolbox (http://www.alivelearn.net/xjview) to obtain the peak coordinates of the DMN and CEN core regions. Based on the obtained core peak coordinates, a total of 13 ROIs were defined as spheres with a radius of 6 mm in this study (Table 3 and Figure 2). The eight ROIs of the DMN are: precuneus (PCUN), bilateral angular gyrus (ANG), anterior cingulate gyrus (ACG), left middle frontal gyrus (MFG), left middle temporal gyrus (MTG) and bilateral inferior occipital gyrus (IOG); the five ROIs of the CEN are: right angular gyrus (ANG), left inferior parietal lobule (IPL), left middle temporal gyrus (MTG) and bilateral middle frontal gyrus (MFG).
Spectral dynamic causal modeling
Using SPM (http://www.fil.ion.ucl.ac.uk/spm/software/spm12/), we establish a general linear model (GLM) for each subject and regress the extracted CSF, WM signals and motion parameters as covariates to adjust the GLM. After that, the sphere model is established for the selected ROIs, and the GM mask is used to assist in extracting the time series of each ROI. Then an 8×8 fully connected model for DMN networks and a 5×5 fully connected model for CEN networks were established respectively, and Cross-Spectra was used for parameter estimation. After the parameter estimation is completed, define all possible models, use Bayesian estimation to obtain the corresponding posterior probability of the model, then select the optimal model, and finally obtain the optimal model and its effective connectivity.
rTMS protocol
This study used a Magstim Rapid2 magnetic stimulator connected to a 70 mm figure-8-shaped coil with two coils cross-tipped at the Pz site of the electroencephalogram system 10-20, the PCUN. The stimulus intensity was set to 100% of the resting-state Motor threshold (MT). MT was determined by applying rTMS to the left PCUN (approximately located above the central sulcus) and moving 0.5 cm along the scalp on the left motor cortical area (M1). At the same time, the opposite side (right) was observed to relax the first dorsal interosseous (FDI) muscle. At least 5 of the 10 experiments produced the minimum intensity value of 50 uV motor evoked potential (MEP)[35].
10 SCD patients and 11 aMCI patients all experienced a total of 25 rTMS. Continuous stimulation for 40 times, taking 4s, followed by an interval of 56s, and repeated stimulation for 25 times, with a total pulse number of 1000 times, lasting for 25 min. Each subject received once a day, five times a week (Monday to Friday), two weeks as a course of treatment, a total of four weeks (two courses). Subsequently, imaging data acquisition and neuropsychological evaluation were performed on each subject.
Statistical analyses
The Statistical Package for the Social Sciences (SPSS) software version 22.0 (IBM, Armonk, NY, USA) was used for statistical analyses. The analysis of variance (ANOVA), paired T-test, and the chi-square test were conducted to compare the demographic and neurocognitive data among groups, including the HC, SCD, aMCI, and SCD and aMCI before rTMS and after rTMS. The Bonferroni correction was used for post hoc comparisons. The P-value was set as <0.05 for significant differences.
For the 8×8 fully connected model built in the DMN network and the 5×5 fully connected model built in the CEN network, after obtaining the effective connectivity values of all the subjects, the effective connectivity of the three groups of subjects were tested with a one-sample t-test (p<0.05) to find the effective connectivity that were significantly non-zero to obtain the effective connectivity patterns within the DMN and CEN networks of the three groups of subjects. Then we used variance (ANOVA) to compare the effective connectivity among HC, SCD, and aMCI groups (p<0.05, uncorrected) to find out whether there were differences in the effective connectivity between the groups. Then, after controlling for the effects of age, gender, and level of education, we conducted correlation analyses between effective connectivity changes and cognitive function to reveal their relationships (p<0.05).
After rTMS intervention in 10 SCD patients and 11 aMCI patients, we used paired T-tests to compare the changes in effective connectivity before and after rTMS in SCD and aMCI groups (p<0.05, uncorrected). Similarly, we conducted correlation analyses between effective connectivity and cognitive changes before and after rTMS intervention (p<0.05).