The study was observational and used data from the control group of MCI in the SIMPLE cohort (detailed information is available in [1]). The cohort recruited MCI patients with age between 50 to 80 years from memory clinics or community. After screened by neurologists, patients with cognitive decline caused by other diseases were excluded (including but not limited to cerebrovascular disease, central nervous system infections, Parkinson’s disease, metabolic encephalopathy, deficiency of folic acid/vitamin B12 and hypothyroidism). The patients were diagnosed according to the National Institute on Aging-Alzheimer’s Association (NIA-AA) workgroups [15]. Participants present mild cognitive problems based on the Mini-Mental State Examination and Clinical Dementia Rating scale (0.5). All MCI cases had medial temporal lobe atrophy (MTA) score above Grade II. Further, we exclude patients with MRI artifacts. The imaging and cognitive data were collected in the same day.

In total, we included 118 MCI patients (mean age, 65.39 $\pm$ 6.97; 19 male and 40 female patients). Among them, 14 patients progressed to AD after 12-month follow-up (pMCI group, mean age, 69.14 $\pm$ 5.64; 2 male and 12 female patients), while the other 104 remained stable (sMCI group, mean age, 64.88 $\pm$ 7.02; 36 male and 68 female patients).

After the baseline assessment, patients would be followed up without any intervention for cognition and had the identical neuropsychological assessment after 12 months. AD was diagnosed at the 12-month follow-up if patients met the diagnostic criteria of dementia due to AD (described in [1], according to the National Institute on Aging-Alzheimer’s Association (NIA-AA) workgroups [15]). The AD criteria included decline from previous levels of functioning and significant cognitive symptoms which impaired activities of daily living. The diagnosis was made by a skilled clinician based on the individual circumstances of the patient and were further were confirmed by ${}^{18}$F-Florbetapir PET [16]. These patients were defined as progressed MCI (pMCI). The others were considered as stable MCI (sMCI). Each patient signed written informed consent, and Ruijin Hospital’s ethical committees had approved the study.

Baseline demographics include sex, age, and education level, listed in Table 1. Neuropsychological battery for multiple cognitive domains was administered to all patients, including Alzheimer’s Disease Assessment Scale-Cognitive subscale (ADAS-Cog) [17], the Auditory Verbal Learning Test (AVLT) - Huashan version [18], the Trail Making Test (TMT, including Part A and B) [19], the Rey-Osterrieth Complex Figure Test (CFT) [20], the Stroop Color-World Test (SCWT) and Boston Naming Test. In the present study, we had considered only participants who underwent the neuropsychological assessment both at the first time point and after 12 months.

MRI images were acquired using a 3.0T uMR-890 MRI scanner (United Imaging, Shanghai, China) with a 64-channel coil and simultaneous multislice imaging techniques. The patients were placed in a supine position quietly with their eyes closed but asked to stay awake during the MRI. The MR sequences included high-resolution 3-dimensional T1-weighted imaging and resting-state fMRI. We collected data in Clinical Neuroscience Center, Ruijin Hospital LuWan Branch, Shanghai Jiao Tong University School of Medicine, Shanghai, China.

High-resolution T1-weighted images were acquired as the templates of further functional images for coregistration. The sequences had the following parameters: repetition time (TR) = 7.52 ms, echo time (TE) = 3.4 ms, field of view (FOV) = 256 $\times$ 256 mm, flip angle = 7 degree, thickness = 0.5 mm, no interslice gap, bandwidth = 240 Hz. All of the brain data were acquired in the sagittal plane, yielding 480 continuous slices with an acquired voxel size of 0.5 $\times$ 0.5 $\times$ 0.5 mm${}^3$. Eye-closed resting-state imaging was carried out using gradient EPI with the following parameters: TR = 700 ms, TE = 37.2 ms, FOV = 208 $\times$ 208 mm, matrix size = 104 $\times$ 104, resolution = 2.0 $\times$ 2.0 $\times$ 2.0 mm, slice thickness = 2 mm, number of slices = 70, multiple band = 10, flip angle = 52 degree. The acquisition orientation was axial, and the order of acquisition was output for slice timing in the header information as the protocol was simultaneous multi-slice acceleration. A total of 670 whole-brain volumes were acquired.

The rs-fMRI data were preprocessed using the SPM12 software package (v7487, Wellcome Department of Imaging Neuroscience, London, United Kingdom) based on MATLAB 2020a (Mathworks Inc., Sherborn, MA, USA). Briefly, the preprocessing steps consisted of removal of the first ten volumes of resting-state data for magnetization stabilization, field bias correction, slice time correction, and Friston 24-parameter motion correction. Spatial normalization was performed by Diffeomorphic Anatomical Registration Through Exponentiated Lie algebra (DARTEL) tool for creating a group specific template and for normalizing functional images to the common space [21]. Spatial smoothing was performed in the resultant images using a Gaussian kernel with an 6-mm full-width half-maximum with the modified MATLAB toolbox “Data Processing & Analysis of Brain Imaging (DPABI, version 3.1)” [22]. We excluded patients who had head motion $>$1.0 mm in any directions. Nuisance covariates, including white matter, cerebrospinal fluid and global signals were further regressed out.

Then a group-level independent component analysis (ICA) was used to define the 25 brain nodes of interest (independent component, IC) as described in previous studies [23, 24]. We performed spatial group independent component analysis (ICA) implemented in the Group ICA of functional MRI Toolbox (GIFT v4.0a; http://icatb.sourceforge.net) [25, 26]. Resting-state functional MRI data of all patients were firstly decomposed into principal components for subject-specific data reduction, and then spatially decomposed into 25 ICs, each of which exhibited a unique time course profile, based on the Infomax algorithm [27]. The resultant data was converted to Z scores, and ICs with Z score $>$1.5 can be found in Supplementary Fig. 1. We additionally filtered the fMRI data between frequencies of 0.01 and 0.1 Hz.

As depicted in Fig. 1, the sliding-window approach was used to explore time-resolved dFNC. Resting-state time-series data were extracted and segmented into a 30-repetition time (TR) window with a size of 21 s, which is convolved with Hamming function to mitigate edge artifacts of the windows and attenuate potentially noisy signals [14]. The window was sliding step-wise by 1 TR along the 660-TR length scan, resulting in a total number of 631 consecutive windows. We chose 30 TRs segmented window length as the grade-off between dynamics of functional connectivity [28] and 1/f0 wavelength criterion proposed by Zalesky and Breakspear [29]. Here, the pairwise Pearson’s correlation coefficient between 25 ICs was calculated in each 21s time window. We thus obtained 631 correlation matrix, and positive matrix values entered the next multilayer network switching analysis. Although some identified clusters might be related to noise, we did not exclude ICs from ICA for the following reasons as in the previous studies [14, 30]. In the preprocessing step, nuisance covariates, including white matter, cerebrospinal fluid and global signals were regressed out. Meanwhile, the ICs identified from ICA could not be 100% from noise as they may be in the mid brain (i.e., IC 1). In addition, the switching rate represented the dynamic change of each IC, and it was not greatly influenced by other ICs which might be noise. The elastic net regression analysis automatically selected switching rate of ICs and it could further rule out switching rate of noise.

Fig. 1.

Flow diagram shows data processing and patients grouping for the present study. Abbreviation: MCI, mild cognitive impairment; IC, independent component.

To quantify spatiotemporal network switching, we used an iterative and ordinal Louvain algorithm to track network function over time [31], with the code governing parameters ($\gamma$ = 1 and $\omega$ = 1) and formula suggested by Pedersen, Zalesky [14]:

(1)$$\begin{array}{c}\hfill \mathrm{Q}(\gamma ,\omega )=\frac{1}{2\mu }\sum _{\mathrm{ijsr}}\left[\left({\mathrm{A}}_{\mathrm{ijs}}-{\gamma }_{\mathrm{s}}\frac{{\mathrm{k}}_{\mathrm{is}}{\mathrm{k}}_{\mathrm{js}}}{2{\mathrm{m}}_{\mathrm{s}}}\delta ({\mathrm{M}}_{\mathrm{is}},{\mathrm{M}}_{\mathrm{js}})+\delta (\mathrm{i},\mathrm{j})\cdot {\omega }_{\mathrm{jrs}}\right)\right]\hfill \\ \hfill \delta ({\mathrm{M}}_{\mathrm{is}},{\mathrm{M}}_{\mathrm{jr}})\hfill \end{array}$$

Modularity was quantified by Q ranging from 0 (low network segregation) to 1 (high network segregation). $\delta (M_{is},M_{js})$ and $\delta (M_{is},M_{jr})$ were 1 if nodes belonged to in the same module and 0 if they did not belong to the same module (M). This process was iterated before the inherent heuristics of the multilayer modularity algorithm converged, which also determined the number of modules per individual. A${}_{\text{ijs}}$ is the sliding-window correlation matrix between node i and j for time point s. The k was node degree at time point s, and m = sum degree of all nodes at time point s. The $\gamma$${}_{\text{s}}$ is the topological resolution parameter of time point. $\omega$${}_{\text{jrs}}$ is the temporal coupling parameter for node j between time window r and s.

Networks had average modularity (Q) of 0.536 $\pm$ 0.026 SD. The final output of the multilayer modularity algorithm was a 25 $\times$ 631 array with integer values denoting modules that each IC was assigned to in the specific time window. The switching rate for each IC was estimated as the percentage of time windows when a brain node transitions between different module assignments.

We used elastic net regression to test the association between network switching rate of 25 ICs and the individual’s cognitive performance at baseline, as well as its changes (12 months - baseline), adjusted for age, sex, and education, and APOE genotype. Elastic net enabled data-driven regression analysis by enforcing sparsity of regression output values and it provided automatic variable selection by removing all ICs not related to cognitive performance. Network switching rates, cognition data were normalized into z-scores to ensure all data were scaled equally. According to previous study, we set the $\alpha$ value to 0.5 to take advantage of the relative strengths of LASSO and Ridge regression approaches, providing a non-sparse solution with low variance among several correlated independent variables [14]. In each regression, we calculated elastic net over a range of different $\lambda$ values between 0 and 1 with increments of 0.001 using 10-fold cross-validation (see equation in the Supplementary Material), The selected threshold had lowest mean square error over all possible $\lambda$s across the 10-folds.

Given the $\beta$ from elastic net regression equation in the Supplementary Material, we defined prediction as:

(2)$$\text{Prediction}=\mathrm{X}\cdot \beta +{\beta }_0$$

Where X is the original values of our neuropsychological variables, and $\beta$${}_0$ is the intercept of the elastic net regression model. The significance and Spearman’s Rho between prediction and X were calculated for regression evaluation.

In relationship with MMSE, AVLT, STT, CFT, and Boston naming test, positive $\beta$ suggests a positive relationship between switching rate and cognitive performance. On the contrary, positive $\beta$ found in the regression between switching rate and ADAS-Cog or SCWT suggests a negative effect on performance due to the character of the two measurements.

For assessing the association between groups (sMCI and pMCI) and ICs’ switching rate, a Bayesian belief network was introduced. Package Bnlearn (version 4.4, Marco Scutari, Lugano, Switzerland) and Caret (version 6.0-81, Max Kuhn, New London, CT, USA) in RStudio software (the R Project for Statistical Computing, R software, version 3.1.0, R Core Team, Boston, MA, USA) was introduced for building a Bayesian belief network and directed acyclic graph. To quantitively define their relationship, we calculated the mean receiver operating characteristic (ROC) curve for the results from the Bayesian belief network. For multiple comparisions, we used Bonferroni correction to minimize the chance of false positive rate.

The age, sex, years of education, and ApoE genotype were of no difference between the two groups (Mann-Whitney U-test for continuous data or Chi-square test for categorical data). The pMCI group had significantly declined performance mainly in SCWT and ADAS-Cog test (p 0.05, repeated measures analysis, Table 1). By contrast, sMCI had statistically better cognitive performance in AVLT, CFT and ADAS-Cog after 12-month follow-up. It suggested sMCI of negative trend for cognitive decline. Thus, sMCI could be the control for pMCI patients who would progress to AD with AD clinical and pathological characters.

We identified 25 ICs from baseline resting-state fMRI using the group ICA. Some ICs are topologically consistent with previous results on the networks by visual inspection [24, 32]: default mode network (DMN, IC 15), salience network (IC 9), auditory network (IC 10, 14), sensorimotor network (SMN, IC 17), visual network (IC 3, 11), executive control network (ECN, IC 5, 13, 16) and cerebellum network (IC 19, 22). The detailed spatial maps of independent components are listed in Supplementary Fig. 1.

The switching rate of each IC was calculated, as shown in Supplementary Table 1, and normalized further. The regularization parameter in elastic regression, $\lambda$, and the minimum mean square error was listed in Table 2. With each pair of parameters, the switching rate of several ICs contributed to specific cognitive domains after adjusting the effect from age, sex, education, and ApoE genotype. The mapping of IC 15 was comprised of the posterior cingulate and medial prefrontal cortex, suggesting its major role in the DMN. In the models above, a higher switching rate of IC 15 was positively related to global and word recall of ADAS-Cog performance (Table 2).

Besides, the switching rate of the auditory network (IC 14) involved in the AVLT immediate and 5-min recall, and it also positively contributed to ADAS-Cog word recall. DMN (IC 15) also positively contributed to AVLT 5-min recall and ADAS-Cog word recall. The switching rate of SAN (IC 9) and ECN (IC 16) negatively associated with performance in ADAS-Cog word recall. Another component in ECN (IC 5) also negatively associated with MMSE and SCWT-interference (Fig. 2).

Fig. 2.

Cross-sectional switching rate and cognition. (A) Contribution of switching rate in each IC to cross-sectional cognitive performance. The positive effect suggested that a higher switching rate correlated with better cognition and vice versa. (B) The fitting between original cognitive performance and predicted data from the elastic regression model, with the input of demographics and switching rate. Abbreviation: IC, independent component; MMSE, Mini-mental State Examination; AVLT, Auditory Verbal Learning Test; CFT, Complex Figure Test; SCWT, Stroop Color-Word Test; TMT, the Trail Making Test; BNT, Boston naming test; ADAS-Cog, Alzheimer’s Disease Assessment Scale–Cognitive Subscale.

We compared the switching rate of each IC between sMCI and pMCI groups at baseline, and IC 15 had a significantly higher switching rate in the sMCI group (4.1% $\pm$ 1.4% vs. 2.3% $\pm$ 1.5%, p = 0.0017). Further, we estimated the predictive effect of switching rates in each ICs on the disease progression. In the elastic regression model used above, lower switching rate of IC 15 predicted the MCI progress ($\lambda$ = 0.169, mSE = 0.109, r = 0.339, p = 0.009) and faster MMSE decline ($\lambda$ = 0.067, mSE = 0.038, r = 0.554, p 0.001). By contrast, lower switching rate in SMN (IC 17) contributed to slower cognitive decline in ADAS-Cog, AVLT-5 min and AVLT-20 min recall (Table 3, Fig. 3).

Fig. 3.

Switching rate and longitudinal cognition changes. (A) Contribution of switching rate in each IC to longitudinal cognitive changes. The positive effect suggested that a higher switching rate correlated with better cognition endpoints. (B) The mapping of IC 15 and the ROC curve for MCI grouping by IC 15’s switching rate. (C) The fitting between original cognitive performance changes and predicted changes from the elastic regression model. Abbreviation: IC, independent component; MMSE, Mini-mental State Examination; AVLT, Auditory Verbal Learning Test; CFT, Complex Figure Test; SCWT, Stroop Color-Word Test; TMT, the Trail Making Test; BNT, Boston naming test; ADAS-Cog, Alzheimer’s Disease Assessment Scale–Cognitive Subscale.

To make the results more consolidated, we used the Bayesian belief network and found that only a lower IC 15 switching rate has an associated effect on the pMCI group (Coefficients of IC 15: –8.135). Since we found the topological association between progress and IC 15, we built a classifier just using IC 15 with a random forest model. The area under the curve was 0.820 $\pm$ 0.094.