Brain network dynamics in patients with single‐ and multiple‐domain amnestic mild cognitive impairment

Abstract

INTRODUCTION

Brain network dynamics have been extensively explored in patients with amnestic mild cognitive impairment (aMCI); however, differences in single‐ and multiple‐domain aMCI (SD‐aMCI and MD‐aMCI) remain unclear.

METHODS

Using multicenter datasets, coactivation patterns (CAPs) were constructed and compared among normal control (NC), SD‐aMCI, MD‐aMCI, and Alzheimer's disease (AD) patients based on individual high‐order cognitive network (HOCN) and primary sensory network (PSN) parcellations. Correlations between spatiotemporal characteristics and neuropsychological scores were analyzed.

RESULTS

Compared to NC, SD‐aMCI showed temporal alterations in HOCN‐dominant CAPs, while MD‐aMCI showed alterations in PSN‐dominant CAPs. In addition, transitions from SD‐aMCI to AD may involve PSN, while MD‐aMCI to AD involves both PSN and HOCN. Results were generally consistent across datasets from Chinese and White populations.

DISCUSSION

The HOCN and PSN are distinctively involved in aMCI subtypes and in the transformation between aMCI subtypes and AD, highlighting the necessity of aMCI subtype classification in AD studies.

Highlights

• Individual functional network parcellations and coactivation pattern (CAP) analysis were performed to characterize spatiotemporal differences between single‐ and multiple‐domain amnestic mild cognitive impairment (SD‐aMCI and MD‐aMCI), and between distinct aMCI subtypes and Alzheimer's disease (AD).

• The analysis of multicenter datasets converged on four pairs of recurrent CAPs, including primary sensory networks (PSN)‐dominant CAPs, high‐order cognitive networks (HOCN)‐dominant CAPs, and PSN–HOCN‐interacting CAPs.

• The HOCN and PSN are distinctively involved in aMCI subtypes and in the transformation between distinct aMCI subtypes and AD.

1. BACKGROUND

Mild cognitive impairment (MCI) is regarded as a critical intermediate stage between normal cognition and dementia. ¹ Patients with MCI are characterized by cognitive decline greater than that expected for healthy aging but not meeting the criteria for a diagnosis of dementia. ² According to whether memory impairments occur, MCI is commonly divided into amnestic MCI (aMCI) and non‐amnestic MCI (naMCI) subtypes. ³ Patients with aMCI have a higher probability of conversion to Alzheimer's disease (AD), whereas patients with naMCI are more prone to progress toward non‐AD dementia. ⁴ , ⁵ According to the involved domains of cognitive deficits, aMCI can further be classified into single‐domain (SD‐aMCI) and multiple‐domain (MD‐aMCI) subtypes. ⁶

Patients with SD‐aMCI only experience memory deficits, whereas those with MD‐aMCI suffer from other cognitive deficits except for memory. ³ Patients with MD‐aMCI had worse cognitive performance ⁷ and a greater likelihood of conversion to AD than those with SD‐aMCI, ⁸ supporting the notion that MD‐aMCI may constitute a transition stage between SD‐aMCI and AD. ⁹ , ¹⁰ Given a high prevalence and lack of effective treatments for AD, investigating the neuroimaging profiles of these two aMCI subtypes is a critical step toward early recognition, subtype differentiation, and intervention for AD.

Although differences in cortical atrophy have been characterized in SD‐aMCI and MD‐aMCI patients using structural magnetic resonance imaging (MRI), ¹¹ functional alterations have not been fully investigated in these two subtypes. In recent years, resting‐state functional MRI (rs‐fMRI) has been widely used for detecting spontaneous activity changes in populations with cognitive disorders. ¹² , ¹³ , ¹⁴ Using static functional network analysis, studies have revealed alterations not only in high‐order cognitive networks (HOCNs, such as default mode network [DMN] and frontoparietal network [FPN]), ¹² but also in primary sensory networks (PSNs, such as motor‐sensory network [MN] and visual network [VN]). ¹⁵ However, the brain activity is not static but highly dynamic, even during the resting state. ¹⁶ Characterizing dynamics in large‐scale networks allows for detecting cognition‐related network changes associated with neurodegenerative diseases. ¹⁷

A prior study showed that dynamic analysis is superior to static analysis in classifying AD patients from normal controls (NCs). ¹⁸ However, most of the previous studies in dynamic networks segmented the entire scans into short, overlapping time windows to detect MCI‐ and AD‐related alterations. ¹⁹ , ²⁰ These short time windows are not sensitive enough to capture brain patterns at a single time point. In this case, the framework of coactivation pattern (CAP) analysis was developed to characterize brain patterns at a resolution of one time point, ²¹ which allows for a sensitive detection of the brain dynamics of aMCI subtypes.

In CAP analysis, most existing studies constructed CAPs using parcellations defined by group‐level average information, ²² which may neglect individual functional variations across the cerebral cortex. ²³ Due to subtle brain functional differences between SD‐aMCI and MD‐aMCI patients, more sensitive analyses may be required to characterize functional differences in spatiotemporal dimensions. In the spatial dimension, static studies have constructed a personalized functional network parcellation and used it to classify early‐ and late‐stage aMCI ²⁴ or to predict AD symptoms. ²⁵ Through estimating CAPs using an individualized network parcellation, a recent study unraveled changes in the brain dynamics of healthy adults and stroke patients, ²⁶ providing a promising tool for exploring individualized state‐specific alterations in brain networks.

To address the above issues, we performed an individual parcellation and CAP analysis for SD‐aMCI, MD‐aMCI, and AD patients to capture brain fluctuations at the resolution of a single fMRI time point. Based on the results observed in our previous study that healthy aging individuals have HOCN alterations earlier than that in PSNs, which is referred to as the “last‐in‐first‐out” theory, ²² we assumed that alterations in brain dynamics of aMCI subtypes may follow the “last‐in‐first‐out” theory; that is, alterations in the HOCN would be identified in individuals with SD‐aMCI, while these in the PSN would be identified in individuals with MD‐aMCI.

2. METHODS

2.1. Datasets

A total of four datasets were included in this study (Table 1 and Figure 1A). First, Dataset 1 was aimed at constructing a group of CAPs with a large sample size of NCs and served as a reference when analyzing differences in aMCI subtypes in CAPs. Second, Dataset 2 was a private dataset aimed at mapping differences in patients with clinically well‐diagnosed SD‐aMCI and MD‐aMCI. Third, Dataset 3 was aimed at validating the results observed in Dataset 2 by defining SD‐aMCI and MD‐aMCI based on cognitive scales. Finally, Dataset 4 was aimed at validating the progression from SD‐aMCI to MD‐aMCI based on follow‐up fMRI data. Considering the potential influences of ethnicity ¹ and age, ²² Datasets 1 and 2 were launched in China (the Asian population), while Datasets 3 and 4 were part of the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. In addition, according to the main dataset (Dataset 2), the ages of participants in Datasets 1, 3, and 4 ranged from 53 to 85 years. Demographic and clinical data of four datasets are shown in Table 1.

TABLE 1.

Demographic and clinical information of study participants.

Characteristic	AD	MD‐aMCI	SD‐aMCI	NC	P value	F value
Dataset 1	/	/	/	N = 197	/	/
Age	/	/	/	62.14 ± 6.92	/	/
Sex (M/F)	/	/	/	79/118	/	/
Dataset 2	N = 17	N = 20	N = 18	N = 28	/	/
Age	70.53 ± 9.83	72.3 ± 7.36	70.83 ± 7.8	65.64 ± 6.49	0.02	3.45
Sex (M/F)	8/9	11/9	6/12	11/17	0.55	0.69
Education	8.59 ± 4.36	9.9 ± 3.37	8.67 ± 3.82	12.5 ± 2.96	<0.001	6.19
MMSE	19.71 ± 4.83	24.9 ± 3.81	24.28 ± 3.03	28.68 ± 1.54	<0.001	26.20
MoCA	15.29 ± 3.55	20.85 ± 4.34	19.61 ± 4.38	27.07 ± 1.59	<0.001	44.05
BNT	14.76 ± 5.41	22.75 ± 2.43	27.83 ± 1.29	29.39 ± 0.63	<0.001	107.99
TMT‐A	118.24 ± 21.05	108.55 ± 19.45	73.56 ± 11.08	53.68 ± 9.79	<0.001	82.89
TMT‐B	228.41 ± 53.95	219.45 ± 49.33	155.06 ± 35.76	89.96 ± 29.28	<0.001	54.87
AVLT	17.18 ± 6.96	30.8 ± 12.9	28.3 ± 5.29	53.89 ± 9.41	<0.001	64.03
CDR (0, 0.5, 1‐2)	1 = 13, 1‐2 = 4	0.5 = 20	0.5 = 18	0 = 28	<0.001	196.49
CDT (0, 1, 2, 3)	0 = 1, 1 = 10, 2 = 6	1 = 6, 2 = 9, 3 = 5	2 = 4, 3 = 14	3 = 28	<0.001	50.10
Dataset 3	N = 28	N = 36	N = 27	N = 134	/	/
Age	71.44 ± 6.58	72.59 ± 5.46	69.48 ± 6.69	72.89 ± 4.40	0.015881	3.520706
Sex (M/F)	17/11	25/11	14/13	51/83	0.002735	4.852276
Education	15.64 ± 2.51	16.14 ± 2.58	17.04 ± 2.14	16.69 ± 2.28	0.078865	2.293573
Q1SCORE	6.40 ± 1.46	4.80 ± 1.70	3.25 ± 1.14	2.48 ± 1.18	<0.001	85.40502
Q2SCORE	0.64 ± 0.91	0.36 ± 0.54	0.00 ± 0.00	0.08 ± 0.33	<0.001	14.67455
Q3SCORE	1.07 ± 0.90	0.94 ± 0.83	0.00 ± 0.00	0.25 ± 0.47	<0.001	29.61251
Q4SCORE	8.71 ± 1.46	5.67 ± 2.60	3.44 ± 2.01	2.13 ± 1.62	<0.001	115.873
Q5SCORE	0.39 ± 0.69	0.33 ± 0.53	0.00 ± 0.00	0.01 ± 0.09	<0.001	17.95332
Q5SCORE_CUE	0.29 ± 0.46	0.30 ± 0.53	0.00 ± 0.00	0.01 ± 0.09	<0.001	15.67861
Q6SCORE	0.39 ± 0.63	0.28 ± 0.51	0.00 ± 0.00	0.07 ± 0.25	<0.001	9.642176
Q7SCORE	1.79 ± 1.60	0.72 ± 0.78	0.11 ± 0.32	0.14 ± 0.39	<0.001	44.63121
Q8SCORE	6.57 ± 2.96	3.78 ± 2.66	2.70 ± 2.02	1.78 ± 1.67	<0.001	43.86904
Q9SCORE	0.50 ± 0.96	0.11 ± 0.32	0.00 ± 0.00	0.06 ± 0.27	<0.001	9.513751
Q10SCORE	0.32 ± 0.61	0.08 ± 0.28	0.00 ± 0.00	0.01 ± 0.09	<0.001	12.76143
Q11SCORE	0.93 ± 1.09	0.56 ± 0.84	0.00 ± 0.00	0.05 ± 0.25	<0.001	26.06735
Q12SCORE	0.39 ± 0.74	0.06 ± 0.23	0.00 ± 0.00	0.00 ± 0.00	<0.001	16.36935
TOTSCORE	19.30 ± 6.58	11.96 ± 4.64	6.06 ± 2.60	4.93 ± 2.66	<0.001	133.4228
Q13SCORE	2.07 ± 1.39	1.36 ± 0.93	0.00 ± 0.00	0.38 ± 0.67	<0.001	49.58936
TOTAL13	30.08 ± 8.13	18.99 ± 6.68	9.51 ± 4.28	7.45 ± 4.06	<0.001	168.6956
Dataset 4	/	N  = 8	N  = 8	/	/	/
Age	/	72.88 ± 5.40	76.29 ± 4.62	/	0.196323	1.840871
Sex (M/F)	/	4/4	4/4	/	1	0
Education	/	17.88 ± 1.89	17.88 ± 1.89	/	1	0
Q1SCORE	/	3.04 ± 1.21	3.33 ± 1.49	/	0.6761072	0.182048
Q2SCORE	/	0.00 ± 0.00	0.12 ± 0.35	/	0.3342819	1
Q3SCORE	/	0.00 ± 0.00	0.38 ± 0.52	/	0.0596461	4.2
Q4SCORE	/	3.00 ± 1.69	4.62 ± 2.20	/	0.1198042	2.7447795
Q5SCORE	/	0.00 ± 0.00	0.00 ± 0.00	/	/	/
Q5SCORE_CUE	/	0.00 ± 0.00	0.00 ± 0.00	/	/	/
Q6SCORE	/	0.00 ± 0.00	0.00 ± 0.00	/	/	/
Q7SCORE	/	0.12 ± 0.35	0.50 ± 1.07	/	0.3621751	0.8873239
Q8SCORE	/	2.62 ± 1.51	3.50 ± 3.12	/	0.4863821	0.5111773
Q9SCORE	/	0.00 ± 0.00	0.12 ± 0.35	/	0.3342819	1
Q10SCORE	/	0.00 ± 0.00	0.00 ± 0.00	/	/	/
Q11SCORE	/	0.00 ± 0.00	0.38 ± 0.52	/	0.0596461	4.2
Q12SCORE	/	0.00 ± 0.00	0.00 ± 0.00	/	/	/
TOTSCORE	/	5.79 ± 2.00	8.33 ± 4.09	/	0.1366989	2.4926820
Q13SCORE	/	0.00 ± 0.00	0.50 ± 0.53	/	0.0191876	7
TOTAL13	/	8.79 ± 3.31	13.46 ± 5.88	/	0.0706961	3.82668451

Notes: The one‐way analysis of variance was used for comparing continuous variables, the chi‐square test was used for comparing categorical variables among the four groups, and the paired sample t test was used for comparing SD‐aMCI and MD‐aMCI. Data are shown in mean ± standard deviation.

Abbreviations: AD, Alzheimer's disease; AVLT, Auditory Verbal Learning Test; BNT, Boston Naming Test; CDR, Clinical Dementia Rating; MD‐aMCI, multiple‐domain amnestic mild cognitive impairment; MMSE, Mini‐Mental State Examination; MoCA, Montreal Cognitive Assessment; NC, normal controls; SD‐aMCI, single‐domain amnestic mild cognitive impairment; TMT, Trail‐Making Test.

FIGURE 1.

Schematic of datasets and single‐frame dynamic CAP analysis. A, Workflow of four datasets. A CAP analysis was performed on each group. In addition, the NCs in Datasets 1 and 3 were used to construct reference CAPs. The Hungarian maximum matching algorithm was then applied to match group‐level CAPs of each group to reference CAPs. B, Individual parcellation and CAP construction. Individual preprocessed fMRI signals were iteratively clustered to map the individual parcellations. The signal of 5124 vertices at all time points for all participants in each group was extracted as features and concatenated into a feature matrix. For each group, temporal K‐means clustering analysis was subsequently performed to classify the fMRI frames into K clusters, forming group‐specific CAPs. AD, Alzheimer's disease; aMCI, amnestic mild cognitive impairment; CAP, coactivation pattern; fMRI, functional magnetic resonance imaging; MD, multiple domains; NC, normal control; SD, single domain.

2.1.1. Dataset 1

A total of 197 Chinese participants (118 females, age ranging from 53 to 85 years) in the Southwest University Adult Lifespan Dataset (SALD; http://fcon_1000.projects.nitrc.org/indi/retro/sald.html) with a strict quality inspection were included. The dataset was selected to increase the sample size of NCs and test the reliability of CAPs in this study. SALD was originally established to study the brain patterns of aging across lifespans. The enrollment criteria for participants were illustrated in a previous publication. ²⁷

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using PubMed and Web of Science resources. While studies have demonstrated large‐scale brain network alterations in Alzheimer's disease (AD), evidence regarding individualized dynamic coactivation patterns is lacking, analysis of which may provide a more sensitive tool to detect neural processes underlying distinct stages of AD, particularly for the single‐ and multiple‐domain amnestic mild cognitive impairment (SD‐aMCI and MD‐aMCI) subtypes.

2. Interpretation: Our findings lead to an integrated hypothesis describing alterations in brain dynamics in distinct aMCI subtypes. That is, the SD‐aMCI stage involves dynamic changes in high‐order cognitive networks, while the MD‐aMCI stage is associated with changes in primary sensory networks.

3. Future directions: Several future directions are noted: (1) The longitudinal changes in dynamic networks within participants require explorations in future prospective follow‐up studies. (2) A dataset with more time points may allow for a more precise mapping of individual functional network parcellations.

MRI data acquisition was performed on a 3‐Tesla scanner (Siemens Medical). 3D T1‐weighted (T1w) anatomical images were acquired with a magnetization‐prepared rapid gradient echo (MPRAGE) method with the following parameters: repetition time (TR) = 1900 ms, echo time (TE) = 2.52 ms, inversion time (TI) = 900 ms, flip angle (FA) = 90°, acquisition matrix = 256 × 256, slices = 176, thickness = 1.0 mm, and voxel size = 1 × 1 × 1 $m{m}^3$. Blood oxygen level–dependent (BOLD) functional images were obtained in an axial orientation using a T2*‐weighted, multi‐slice gradient echo planar imaging (EPI) sequence with the following parameters: TR = 2000 ms, TE = 30 ms, FA = 90°, field of view (FOV) = 220 × 220 $m{m}^2$, slice thickness = 3 mm, gap = 1 mm, 242 volumes, 32 slices, voxel size = 3.4 × 3.4 × 4 $m{m}^3$.

2.1.2. Dataset 2

To map functional differences in aMCI subtypes, we recruited 83 participants (age ranging from 53 to 85 years) from the Department of Neurology, Xuanwu Hospital, Capital Medical University, including 17 AD participants (9 females), 20 MD‐aMCI participants (9 females), 18 SD‐aMCI participants (12 females), and 28 NCs (17 females). The age‐ and sex‐matched NCs were recruited from the local community via advertisements, and all participants received financial compensation. The ethical approval was obtained from the research ethics committee of Xuanwu Hospital, and written informed consent was obtained from all participants or their relatives before participating in this study.

All the AD and aMCI participants underwent a detailed physical and neurological examination. The diagnosis of AD ³ , ²⁸ was confirmed by two experienced neurologists in accordance with the criteria of the Diagnostic and Statistical Manual of Mental Disorders, 5th edition (DSM‐V) for AD and the National Institute of Neurological and Communicative Disorders and Stroke/Alzheimer's Disease and Related Disorders Association (NINCDS‐ADRDA) criteria for AD. ²⁹ , ³⁰ , ³¹ Participants with aMCI were diagnosed according to Petersen's diagnostic criteria ³ and the National Institute on Aging–Alzheimer's Association criteria for MCI due to AD. ²⁸ The inclusion criteria for aMCI were as follows: (1) memory complaints, preferably confirmed by an informant; (2) age‐related memory decline; (3) basically normal performance in cognitive and daily life activities; (4) a Clinical Dementia Rating (CDR) score of 0.5; ³² and (5) the absence of dementia. Exclusion criteria for all participants included: (1) cognitive impairments caused by brain injury; (2) current or lifetime history of any other neurological or mental disorders that may lead to cognitive impairments, such as stroke, depression, or epilepsy; (3) neurological deficiencies such as vision or hearing loss; or (4) MRI contraindications.

All participants underwent extensive neuropsychological evaluations, including the CDR, ³² the Mini‐Mental State Examination (MMSE), ³³ and the Montreal Cognitive Assessment (MoCA). ³⁴ Four specific cognitive domains were assessed: (1) the Boston Naming Test (BNT) ³⁵ to evaluate naming skills, (2) the Trail‐Making Test (TMT) ³⁶ to measure executive function, (3) the Auditory Verbal Learning Test (AVLT) ³⁷ to measure memory function (short delay free recall), and (4) the Clock Drawing Test (CDT; 3‐point) to assess visuospatial ability. ³⁸ The neuropsychological data were collected on the day of the MRI scan. The MMSE, MoCA, BNT, and AVLT were scored by the number of correct responses, while the TMT was scored by the reaction time. A higher number of correct answers indicates better cognitive performance, ² , ³⁷ whereas a longer reaction time indicates worse cognitive performance. ³⁹

MRI data acquisition was performed on a 3‐Tesla scanner (Siemens Medical Solutions). Foam padding and headphones were used to restrict head motion and reduce scanner noise. 3D T1w anatomical images were acquired with the MPRAGE method with the following parameters: TR = 1900 ms, TE = 2.2 ms, TI = 900 ms, FA = 9°, acquisition matrix = 224 × 256 × 176, and voxel size = 1 × 1 × 1 $m{m}^3$. BOLD functional imaging was obtained in an axial orientation using a T2*‐weighted, multi‐slice gradient EPI sequence with the following parameters: TR = 2000 ms, TE = 40 ms, FA = 90°, acquisition matrix = 64 × 64, FOV = 256 × 256 $m{m}^2$, slice thickness = 4 mm, gap = 1 mm, 239 volumes, 28 slices, voxel size = 4 × 4 × 5 $m{m}^3$.

2.1.3. Dataset 3

To test the reliability of findings in aMCI subtypes observed in Dataset 2, we selected 225 participants (118 females, aged from 53 to 85 years) in the ADNI dataset (https://adni.loni.usc.edu/), including 28 AD participants (11 females), 36 MD‐aMCI participants (11 females), 27 SD‐aMCI participants (13 females), and 134 NCs (83 females). The ADNI dataset was established to study changes in different stages of AD patients from multiple visits and was extensively used as the validation dataset in studies of AD. More details are shown in a previous publication. ⁴⁰ Because the ADNI dataset did not label aMCI as SD or MD, we carefully checked the diagnosis of each participant during all the visits and defined the impairment of each domain of memory, language, and praxis based on the scores of the Alzheimer's Disease Assessment Scale‐Cognitive Subscale (ADAS‐Cog). ⁴¹ In detail, NC was defined as no impairments in any domains, SD‐aMCI was defined as damage that only exists in the domain of memory, while MD‐aMCI was defined as impairments that coexist in both memory and any other cognitive domains (language or praxis). The memory, language, and praxis domains were considered impaired when their corresponding ADAS‐Cog item score was > 0. We selected participants according to their ADAS‐Cog scores at baseline. ⁴¹ All the MCI patients due to AD were included and MCI patients due to other diseases were excluded. Due to a lack of screening by experienced neurologists, the results obtained from the ADNI dataset can only partially validate the present findings. Finally, only participants with complete data on T1w, rs‐fMRI, and ADAS‐Cog were included. MRI data were downloaded on May 13, 2024. ADNI data were collected on multiple 3‐Tesla scanners (Siemens, GE, and Philips). See more details in the previous publications. ⁴² , ⁴³ ,  

2.1.4. Dataset 4

Dataset 4 was also a part of the ADNI dataset, which was used to validate the neurodynamic changes in longitudinal development of different aMCI stages. We selected the baseline and follow‐up data of eight patients (four females) who were determined as SD‐aMCI at baseline and were transformed to MD‐aMCI patients subsequently, with SD‐aMCI and MD‐aMCI defined by scales. The MRI data acquisition details were similar to Dataset 3.

2.2. Data preprocessing

All T1w and BOLD images were preprocessed using the fMRIPrep‐22.1.1 pipeline. ⁴⁴ In brief, the T1w image was corrected for intensity non‐uniformity and used as the T1w‐reference. The T1w‐reference was then skull‐stripped using OASIS30ANTs as the target template. Brain tissue segmentation of cerebrospinal fluid (CSF), white matter (WM), and gray matter (GM) was performed on the brain‐extracted T1w using FAST. Cortical surfaces were reconstructed using FreeSurfer. Finally, the T1w‐reference and brain‐extracted T1w were spatially normalized to the Montreal Neurological Institute (MNI) space through non‐linear registration using Advanced Normalization Tools (ANTs).

For BOLD images of each participant, the following preprocessing was performed. First, the BOLD time‐series were slice‐timing corrected, and the transforms were applied to correct for head motion. In addition, several confounding time series were calculated, including the framewise displacement (FD) and three of the region‐wise global signals (including the CSF, WM, and the whole‐brain masks). Second, the preprocessed BOLD images were resampled to the fsaverage6 surface space, which consisted of 40,962 vertices in each hemisphere. Afterward, the removal of non–steady state frames, spatial smoothing (an isotropic, Gaussian kernel of 6 mm full‐width half‐maximum), and motion artifacts were performed. Finally, we performed de‐mean, band‐pass filtering (0.01 to 0.08 Hz), as well as regression of nuisance signals such as FD and region‐wise global signals. The maximum values of head motion were required to be < 3 mm/degree. To reduce the computational cost, the preprocessed data were downsampled to the fsaverage4 surface space, which included 2562 vertices in each hemisphere, resulting in 5124 vertices in the whole brain.

2.3. Individual parcellation and CAP construction

To parcellate cortical functional networks at the individual level, we used the HFR‐ai toolbox (https://github.com/MeilingAva/Homologous‐Functional‐Regions) ⁴⁵ , ⁴⁶ to map 5124 vertices of the whole brain onto individual brain parcellations with seven networks, including the attentional network (AN), DMN, FPN, MN, limbic network (LN), salience network (SN), and VN. ⁴⁷ , ⁴⁸ The seven network parcellations were used for the definition of CAP in an individualized pattern, which allows for inter‐study comparisons.

In CAP construction, we extracted signals from vertices in fsaverage4 space as the feature matrix for each participant. Compared to signals extracted from 7‐network atlas, ⁴⁸ Brainnetome atlas, ⁴⁹ or Glasser atlas, ⁵⁰ signals extracted from 5124 vertices were more  suitable for CAP constructions with a high spatial resolution. As shown in Figure 1B, the dimension of the feature matrix in each participant is M × 5124, in which M denotes the number of time frames and 5124 denotes the number of networks. Afterward, we concatenated the feature matrices of participants within the same group and constructed the group‐wise feature matrix. The group‐wise feature matrices were constructed for NCs in Dataset 1, NCs in Dataset 2, SD‐aMCI in Dataset 2, MD‐aMCI in Dataset 2, AD in Dataset 2, NCs in Dataset 3, SD‐aMCI in Dataset 3, MD‐aMCI in Dataset 3, AD in Dataset 3 and aMCI in Dataset 4.

In the construction of reference CAPs, group‐wise feature matrices of NCs in Dataset 1 (SALD, N = 197) and Dataset 3 (ADNI, N = 134) were used as the input of the K‐means clustering algorithm, ⁵¹ with K ranging from 2 to 21. This algorithm was applied to assign BOLD images into different clusters based on their spatial similarity. The spatial similarity was calculated by the Pearson correlation coefficient, which has been previously used to temporally decompose resting‐state networks into multiple CAPs. ²¹ To avoid random effects, the K‐means clustering was replicated 500 times. After clustering, we averaged the volumes with the same labels based on the time series of participants to produce cluster centers (reference CAPs, as shown in Figure 1A). To validate the construction of reference CAPs and determine the optimal value of K, we used the clustering results of the NC group in Datasets 1 and 3 as references, respectively. We evaluated multiple parameters, including the widely used silhouette coefficient and explained variance. In addition, because the clustering analysis in this study used the Pearson correlation distance, the magnitude of the correlations between CAPs can be used as a measure of the clustering effect. That is, the absolute value of the negative correlations between distinct CAPs should be as large as possible. According to the criteria, we selected the value of K when the absolute value of negative correlation between CAPs was the largest. Meanwhile, the silhouette coefficient and explained variance were large.

Based on the optimum K value, the construction of group‐wise CAPs and individual‐wise CAPs was performed with group‐wise feature matrices of each group in each dataset as the input of the K‐means clustering algorithm. After that, the Hungarian matching algorithm was performed for inter‐group matching. The Hungarian matching algorithm was used to solve the assignment problem between group‐wise CAPs and reference CAPs, considering the maximization of correlations among all matched pairs of CAPs. After matching group CAPs to reference CAPs, volumes with the same label were averaged to produce K individual CAPs based on individual time series.

2.4. Spatiotemporal characteristics calculation of CAPs

The clustering analysis usually uses the label vector and cluster centroids to describe findings. In clustering analysis, the dynamics of each participant were described with a vector of labels ranging from 1 to K. The vector suggests that the participant showed different CAPs at different time points. Signals in the same CAPs were averaged to obtain the clustering centroids. In this study, the dimension of CAP features was the total number of vertices, 5124. To define CAPs at the network level, the 5124 features were mapped to individualized brain network parcellation (including AN, DMN, FPN, MN, LN, SN and VN), as described in Section 2.3.

To quantify dynamic differences among groups, we calculated multiple spatiotemporal characteristics at the individual level, including the amplitude of the network in each CAP, the occurrence of CAPs, and transitions between CAPs. In detail, the amplitude characterized the activation and deactivation of networks in different CAPs and was defined as the average intensity of BOLD signals in the seven networks under the K individual CAPs. The occurrence characterized the probability of occurrence and was defined as the proportion of the specific label found in the label vector. In addition to occurrence, the transition time was also an important measure on the time scale, which characterized the transfer probability between two CAPs and was defined as the number of transfers between different labels.

2.5. Statistical analysis

Statistical analysis was performed in MATLAB 2022b (MathWorks) by T.L. and J.Z., with 7 and 13 years of experience, respectively. In the statistical analysis of demographic and clinical variables, a one‐way analysis of variance was used to test group differences in neuropsychological scale scores, and a chi‐square test was used to test differences in categorical variables.

Pearson correlations were used to evaluate the similarities among CAPs, and Wilcoxon signed‐rank tests were used to identify the dominant networks of CAPs and compare the occurrence of CAPs within groups. In this study, CAPs were constructed within the group of each dataset. The Hungarian matching algorithm was used to match the CAPs extracted from each group. The Pearson correlation coefficient between CAPs of NC in Dataset 1 and CAPs of NC in Dataset 2 was calculated to evaluate the matching rate of CAPs between data of different sample sizes. In addition, the Pearson correlation coefficient between CAPs of NC in Dataset 1 and CAPs of NC in Dataset 3 was calculated to evaluate the consistency of CAPs between different populations (Chinese and White populations).

Independent two‐sample t tests were performed on Datasets 2 and 3 to assess differences in the FD values and the spatiotemporal characteristics (including the amplitudes and transition times) of the CAPs among the four groups. Paired‐sample t tests were performed on Dataset 4 to assess differences in the spatiotemporal characteristics of the CAPs during the progression from SD‐aMCI to MD‐aMCI. Furthermore, linear relationships between the spatiotemporal characteristics and neuropsychological scores were estimated by Spearman partial correlations. ⁵² , ⁵³ , ⁵⁴ In the analysis of spatiotemporal dynamics, age, sex, education, FD, and GM volume ⁵⁵ served as covariates. p < 0.05 was considered to indicate the statistical significance level, and the false discovery rate correction (FDR correction) was applied for multiple comparisons correction.

3. RESULTS

3.1. Definition of CAPs

In head motion analysis, differences in FD among NC, SD‐aMCI, MD‐aMCI, and AD are shown in Table S1 in supporting information. In Dataset 2, no significant differences were found among the four groups (all p > 0.05). In Dataset 3, significant differences were found between NC and MD‐aMCI (t = 2.857, P = 0.029). Although there were no significant differences in the FD values among groups, we still included the FD as a covariate in the group comparison analysis to control the head motion.

The determination of K values (corresponding to K CAPs) and CAP definition were performed based on the clustering analysis results of the NC group in Dataset 1 and Dataset 3, which served as a reference. The clustering centers of other datasets were then matched to the well‐defined CAPs of the NC group. The number of brain states was selected mainly based on the absolute values of the negative correlations among CAPs, the silhouette coefficient of the K‐means clustering, and explained variance (see Figures S1 and S2 in supporting information). The clustering analysis showed that the eight‐cluster solution yielded high similarities within CAPs and remarkable differences among CAPs. In addition, when K is > 8, the silhouette coefficient of the K‐means clustering decreases slowly, and the variance gain is close to 0.01, indicating that the clustering effect is not significantly improved. In this way, eight CAPs were ultimately identified.

In each CAP of the NC group in Dataset 1, the BOLD signal intensity of seven networks is shown in Figure 2A. The results of Dataset 2 and Dataset 3 are shown in Table S2 in supporting information. The larger the absolute values of the network signals, the stronger the activation or deactivation of the network. Each CAP had specific brain networks with relatively stronger activation or deactivation than others. Within each CAP, the Wilcoxon rank‐sum test among seven networks revealed the dominant networks, which are in PSN or HOCN (Figure 2B). The P‐ and z values of the Wilcoxon rank‐sum tests are shown in Table S3 in supporting information. In summary, the VN is the dominant network in CAP 1 and CAP 2, the dominant networks of CAP 3 and CAP 4 were the MN and SN, the dominant networks of CAP 5 and CAP 6 were the FPN and SN, and the dominant networks of CAP 7 and CAP 8 were the DMN and LN.

FIGURE 2.

Definition of CAPs based on clustering results of the NC group in Dataset 1. A, Distribution of signal intensities. The MRI signal values of each vertex in the NC group (197 participants) were averaged to obtain the signal intensities of different brain networks in distinct CAPs in the form of spine maps. The vertical axis represents seven networks, and the horizontal axis represents the strength of the MRI signals (see Table S2 in supporting information for more details). B, Definition of functional CAPs. The Wilcoxon rank‐sum test was used to evaluate the absolute signal intensities among the seven networks, and the dominant network of the CAP was subsequently determined. The absolute signal intensity of the dominant network is significantly different from the signal intensity of most of the other remaining networks. The boxplots present the median, interquartile range, minimum, and maximum values, where * indicates p < 0.05, ** indicates p < 0.01, and *** indicates p < 0.001. “V” or “VN” indicates the visual network, “A” or “AN” indicates the attentional network, “D” or “DMN” indicates the default mode network, “F” or “FPN” indicates the frontoparietal network, “L” or “LN” indicates the limbic network, “S” or “SN” indicates the salience network, and “M” or “MN” indicates the motor‐sensory network. The “+” symbol indicates activation, and the “–” symbol indicates deactivation. CAP, coactivation pattern; MRI, magnetic resonance imaging; NC, normal control.

Considering the positive and negative signs of the signal values, CAP 1 showed strong deactivation in the VN (V–), while CAP 2 showed strong activation in the VN (V+). CAP 3 showed a strong activation in the MN (M+) and SN (S+), while CAP 4 was strongly deactivated in the MN (M–) and SN (S–). CAP 5 was strongly activated in the FPN (F+) and SN (S+), while CAP 6 showed a strong deactivation in the FPN (F–) and SN (S–). CAP 7 showed strong activation in the DMN (D+) and LN (L+), while CAP 8 was strongly deactivated in the DMN (D–) and LN (L–). Finally, the eight CAPs were defined as CAP V−, CAP V+, CAP M+S+, CAP M–S–, CAP F+S+, CAP F–S–, CAP D–L–, and CAP D+L+, which were named by their peak activation (shown by a plus sign) and deactivation (shown by a minus sign). Similar to the previous CAP study, ²² the eight CAPs were grouped into four pairs of opposite CAPs (Figures S1 and S2).

A similar analysis was performed on Dataset 3 and the results are shown in Table S4 in supporting information. The eight CAPs of Dataset 3 were defined as CAP V−, CAP V+, CAP V+S+, CAP V–S–, CAP F+, CAP F–, CAP A–, and CAP A+. Therefore, two dependent datasets of NC (Dataset 1 and 3) jointly revealed three kinds of CAPs, including PSN‐dominant CAPs (CAP V– and CAP V+), HOCN‐dominant CAPs (CAP F+S+, CAP F–S–, CAP D–L–, and CAP D+L+; or CAP F+, CAP F–, CAP A–, and CAP A+) and PSN‐HOCN interacting CAPs (CAP M+S+ and CAP M–S–; or CAP V+S+, CAP V–S–).

3.2. Spatial distributions and temporal occurrences of CAPs

Following the definition of CAPs in the NC group, we set K = 8 for the clustering analysis of other datasets. The CAPs of each group were Hungarian‐paired with those of the NC group. The matching results are shown in Table S5 in supporting information. In addition, spatial distribution maps of Dataset 2 were computed by averaging the fMRI frames assigned to the same brain state and displayed on an inflated cortical surface (Figure 3A). The Pearson correlation coefficients between the four pairs of CAPs within each of the four groups were calculated separately, as shown in Table S3. As mentioned above, the CAPs of NC, SD‐aMCI, MD‐aMCI, and AD in Dataset 2 are matched to the reference CAPs, which are calculated by Dataset 1. In Dataset 3, the CAPs of SD‐aMCI, MD‐aMCI, and AD are matched to the CAPs of NC in the same dataset. Pearson correlations between these matching CAPs are shown in Tables S4 and S5.

FIGURE 3.

Eight CAPs in the four groups and their occurrences in Dataset 2. A, Cluster centers for the NC, SD‐aMCI, MD‐aMCI, and AD groups. The preprocessed fMRI signals of each group were iteratively clustered individually, and the cluster centers of NC, SD‐aMCI, MD‐aMCI, and AD in Dataset 2 were matched to the cluster centers of NCs in Dataset 1 using the Hungarian matching algorithm. Each column represents four pairs of CAPs. The red color indicates positive activation, and the blue color indicates negative deactivation. B, Average number of occurrences in each CAP for the four groups. Wilcoxon signed‐rank tests were used to examine within‐group differences in the occurrence, and post hoc analysis was performed using false discovery rate correction, where * indicates p < 0.05, ** indicates p < 0.01, and *** indicates p < 0.001. The boxplots present the median, interquartile range, minimum, and maximum values. “V” indicates the visual network, “A” indicates the attentional network, “D” indicates the default mode network, “F” indicates the frontoparietal network, “L” indicates the limbic network, “S” indicates the salience network, and “M” indicates the motor‐sensory network. The “+” symbol indicates activation, and the “–” symbol indicates deactivation. AD, Alzheimer's disease; aMCI, amnestic mild cognitive impairment; CAP, coactivation pattern; fMRI, functional magnetic resonance imaging; MD, multiple‐domain aMCI; NC, normal control; SD, single‐domain aMCI.

As shown in Table S6 in supporting information, we calculated the occurrence of the eight CAPs in the four groups and compared the occurrence of the eight CAPs by the within‐group Wilcoxon rank sum test. As shown in Figure 3B (results of Dataset 2), CAP D+L+ had a high percentage of occurrence in the NC and SD‐aMCI groups, while CAP M+S+ and CAP M–S– had a high percentage of occurrence in the MD‐aMCI and AD groups. In Dataset 3, CAP A+ had a high percentage of occurrence in the NC, SD‐aMCI, MD‐aMCI, and AD groups. In addition, CAP V– is highly involved in AD, and CAP V–S– is highly involved in MD‐aMCI. Overall, these findings suggest a dominant involvement of HOCN in the SD‐aMCI stage and that of PSN in the MD‐aMCI stage.

3.3. Differences in network amplitudes of CAPs among four groups

As depicted above, the eight CAPs can be grouped into four pairs (Dataset 1: CAP V− and CAP V+; CAP M+S+ and CAP M–S–; CAP F+S+ and CAP F–S–; CAP D–L– and CAP D+L+; Dataset 3: CAP V− and CAP V+; CAP V+S+ and CAP V–S–; CAP F+ and CAP F–; CAP A– and CAP A+), and the 5124 vertices are individually parcellated into seven networks, namely, the VN, LN, DMN, MN, SN, FPN, and AN. Based on individual parcellations, the amplitudes of each network in each CAP were compared among the NC, SD‐aMCI, MD‐aMCI, and AD groups by using two‐sample t tests, with age, sex, education, FD, and GM serving as covariates. Group differences derived from Dataset 2 are shown in Figure 4A and Table S7 in supporting information (P < 0.05, FDR‐corrected). In brief, significant differences between SD‐aMCI and MD‐aMCI were found in the signal strength of DMN in CAP V– (t = −4.503, p < 0.001), VN in CAP V– (t = −3.373, p = 0.0077), and VN in CAP M+S+ (t = −4.569, p < 0.001). Compared to NC, SD‐aMCI showed a significant decreased signal strength of DMN in CAP V– (t = −6.709, p < 0.001) and SN in CAP V– (t = −5.696, p < 0.001), while MD‐aMCI showed a significant increased signal strength of VN in CAP M+S+ (t = 3.496, p = 0.005). Compared to AD, aMCI subtypes showed significantly increased signal strength of DMN in CAP M+S+ (SD‐aMCI: t = 4.110, p = 0.002; MD‐aMCI: t = 4.742, p < 0.001), and FPN in CAP D+L+ (SD‐aMCI: t = 5.567, p < 0.001; MD‐aMCI: t = 4.980, p < 0.001), and significantly decreased signal strength of VN in CAP V– (SD‐aMCI: t = −7.395, p < 0.001; MD‐aMCI: t = −6.661, p < 0.001), DMN in CAP V– (SD‐aMCI: t = −9.753, p < 0.001; MD‐aMCI: t = −6.198, p < 0.001), SN in CAP V– (SD‐aMCI: t = −8.700, p < 0.001; MD‐aMCI: t = −7.905, p < 0.001), DMN in CAP M–S– (SD‐aMCI: t = −4.493, p < 0.001; MD‐aMCI: t = −7.378, p < 0.001), VN in CAP F+S+ (SD‐aMCI: t = −6.497, p < 0.001; MD‐aMCI: t = −6.787, p < 0.001) and FPN in CAP D–L– (SD‐aMCI: t = −4.128, p = 0.002; MD‐aMCI: t = −3.901, p = 0.002).

FIGURE 4.

Changes in network amplitudes of CAPs. A, Significant differences in network amplitudes among groups in Dataset 2. The violin plots present the median, interquartile range, minimum, and maximum values. * indicates p < 0.05, ** indicates p < 0.01, and *** indicates p < 0.001. All the above P‐values are FDR corrected. B, Correlations between network amplitudes and neuropsychological scales (more details are shown in Table S6 in supporting information). Age, sex, education, FD, and GM served as covariates, and all P values were FDR corrected. The Y axis and X axis show the amplitudes of networks (including the AN, DMN, FPN, LN, MN, SN, and VN) and neuropsychological measurement scores (including the MoCA, AVLT, BNT, TMT, and MMSE), which were adjusted by covariates such as age, sex, education, FD, and GM. The small gray circle indicates no significant correlations (p > 0.05), while the purple circles indicate significant correlations (p < 0.05). Larger and darker circles indicate stronger correlations. C, Summary of results in network amplitudes across Dataset 2 and Dataset 3. Significant network changes are labeled in the line connecting groups. Consistent results between Dataset 2 and Dataset 3 are shown in red. AN, attentional network; AVLT, Auditory Verbal Learning Test; BNT, Boston Naming Test; CAP, coactivation pattern; DMN, default mode network; FD, framewise displacement; FDR, false discovery rate; FPN, frontoparietal network; GM, gray matter; LN, limbic network; MMSE, Mini‐Mental State Examination; MN, motor‐sensory network; MoCA, Montreal Cognitive Assessment; SN, salience network; TMT, Trail‐Making Test; VN, visual network.

Furthermore, we estimated the correlations between neuropsychological scale scores and network amplitudes for different CAPs using Spearman partial correlations. As shown in Figure 4B and Table S8 in supporting information, significant correlations between amplitudes and neuropsychological scores survived after FDR correction. A similar analysis was performed on Dataset 3 and the results are shown in Table S8. The results of Datasets 2 and 3 were summarized in Figure 4C, in which validated results are shown in red and the results found only in Dataset 2 are shown in black. In addition, comparisons on Dataset 4 also showed significant differences in VN in CAP F– between SD‐aMCI and MD‐aMCI (t = −2.541, p = 0.005), suggesting the disturbance of VN in aMCI subtypes. Overall, these results characterize changes in networks and CAPs from SD‐aMCI to MD‐aMCI and AD.

3.4. Differences in CAP transitions between groups

In each group, we calculated the transition matrix (dimension: 8 CAPs × 8 CAPs) for each participant, and the number of transitions was summed for all participants (Figure 5A). Each transition in the transition matrix was compared among four groups using an independent two‐sample t test, and the results are shown in Figure 5B (Dataset 2) and Table S9 in supporting information (Dataset 3). A significant difference in the transition from CAP F+S+ to CAP F–S– was observed in the comparisons between SD‐aMCI and MD‐aMCI (t = −3.420, p = 0.035), between NC and SD‐aMCI (t = 4.403, p = 0.005), between MD‐aMCI and AD (t = 4.360, p = 0.007), and between NC and AD (t = 4.224, p = 0.006). Compared to NC, MD‐aMCI showed more transitions from CAP V– to CAP V+ (t = 3.107, p = 0.037), while AD showed more transitions from CAP F+S+ to CAP V+ (t = 4.224, P = 0.005), CAP V– to CAP F+S+ (t = 3.342, p = 0.034), CAP M–S– to CAP V– (t = 3.038, P = 0.042), but fewer transitions from CAP V+ to CAP D+L+ (t = −4.495, P = 0.008), CAP D+L+ to CAP V– (t = −3.146, p = 0.039), CAP D+L+ to CAP V+ (t = −4.069, p = 0.007). Compared to AD, aMCI showed more transitions from CAP D+L+ to CAP V+ (SD‐aMCI: t = 3.373, p = 0.037; MD‐aMCI: t = 3.230, p = 0.038), but less transitions from CAP F+S+ to CAP V+ (SD‐aMCI: t = −3.222, p = 0.038; MD‐aMCI: t = −3.829, p = 0.016).

FIGURE 5.

Changes in transitions between CAPs among groups in Dataset 2. A, The number of transitions among the eight CAPs. The X and Y axes of the matrix represent the eight CAPs; the elements in the matrix represent a kind of transition; and the values of the elements represent the number of transitions, where the darker and larger circles indicate the larger transition. B, Differences in CAP transitions among groups. The violin plots present the median, interquartile range, minimum, and maximum values, where * indicates p < 0.05, ** indicates p < 0.01, and *** indicates p < 0.001. “V” indicates the visual network, “A” indicates the attentional network, “D” indicates the default mode network, “F” indicates the frontoparietal network, “L” indicates the limbic network, “S” indicates the salience network, and “M” indicates the motor‐sensory network. The “+” symbol indicates activation, and the “–” symbol indicates deactivation. AD, Alzheimer's disease; aMCI, amnestic mild cognitive impairment; CAP, coactivation pattern; MD, multiple domains; NC, normal control; SD, single domain.

We additionally estimated the correlations between the transitions and neuropsychological scales by Spearman partial correlation analysis, with age, sex, education, FD, and GM serving as covariates. As shown in Figure 6A and Table S9, strong correlations between interstate transitions and clinical performance were observed after pairwise comparisons with Bonferroni correction. We drew a schematic diagram, as shown in Figure 6B, to represent significant differences in transitions between CAPs among groups, which combined the differences in CAP transitions between groups and the correlations between CAP transitions and neuropsychological scales in Datasets 2 and 3.

FIGURE 6.

A schematic representation of changes in CAP transitions among groups. A, Correlations between transitions and neuropsychological scales. The X axis represents the number of transitions, and the Y axis represents the neuropsychological scale scores. Each point represents a CAP pair, and the shading represents 95% confidence intervals. The correlation between the number of transitions in CAP and the scale was calculated for four groups of participants, with age, sex, education, FD, and GM serving as covariates. All the P values were false discovery rate corrected. B, A summary of correlations and changes in transitions among groups. Combining the significant differences and scale‐correlated findings, a schematic diagram of the trends in CAPs during the progression from SD‐aMCI to MD‐aMCI and AD was obtained. “V” indicates the visual network, “A” indicates the attentional network, “D” indicates the default mode network, “F” indicates the frontoparietal network, “L” indicates the limbic network, “S” indicates the salience network, and “M” indicates the motor‐sensory network. The “+” symbol indicates activation, and the “–” symbol indicates deactivation. AD, Alzheimer's disease; aMCI, amnestic mild cognitive impairment; AVLT, Auditory Verbal Learning Test; BNT, Boston Naming Test; CAP, coactivation pattern; FD, framewise displacement; fMRI, functional magnetic resonance imaging; GM, gray matter; HOCN, high‐order cognitive network; MD, multiple‐domain aMCI; MMSE, Mini‐Mental State Examination; MoCA, Montreal Cognitive Assessment; NC, normal control; PSN, primary sensory network; SD, single‐domain aMCI; TMT, Trail‐Making Test.

4. DISCUSSION

To the best of our knowledge, this study is the first to track large‐scale network dynamics underlying distinct aMCI subtypes and the differences between them and AD. We performed resting‐state individual parcellations and constructed four pairs of recurrent CAPs. Validated by multiple datasets, temporal alterations in HOCN‐dominant CAPs were found in SD‐aMCI compared to NC. As for MD‐aMCI, a temporal alteration in PSN‐dominant CAPs was found compared to NCs. In addition, the transitions from SD‐aMCI to AD may involve PSNs, while MD‐aMCI to AD was associated with both PSNs and HOCNs, suggesting a differential neural mechanism underlying the transformation between distinct aMCI subtypes and AD.

4.1. CAPs of PSNs and HOCNs

Consistent with previous studies, ²² , ²⁶ the present study revealed HOCN‐dominant CAPs, PSN‐dominant CAPs, and PSN–HOCN‐interacting CAPs based on two independent datasets of Chinese and White populations, and a strong pairwise negative correlation was present between these CAPs. In these two independent datasets, the PSN‐dominant CAPs are highly consistent because the dominant network in both datasets is VN. The PSN–HOCN‐interacting CAPs showed consistent activation in SN, while the SN was coactivated with the PSN (VN or MN), which varied as datasets. As for HOCN‐dominant CAPs, activation in FPN is observed in both datasets, but the activation of FPN is also accompanied by that of SN in the Chinese dataset. Finally, activation of AN is revealed in the ADNI dataset, while activation of DMN and LN is revealed in the Chinese dataset, suggesting a large variability in HOCN‐dominant CAPs. The consistency of PSNs and variation of HOCNs are supported by a previous study on brain functional differences between Chinese and White populations, which also showed a high consistency of PSNs between these two populations. ⁵⁶ Using individual parcellations, a prior study also revealed a high consistency of PSNs and high variability of HOCNs across individuals. ⁴⁶ In addition, it is of note that brain function could be affected by racial background, and the brain patterns vary among populations of different races. ⁵⁷ Hence, the CAP inconsistency observed in these two datasets may be partially attributed to ethnic backgrounds (Asian or White populations).

4.2. Spatial and temporal patterns of CAPs in aMCI subtypes

In spatial analysis, our study revealed a decrease of SN in PSN‐dominant CAPs in SD‐aMCI patients compared to NCs, which was basically consistent across the Chinese and ADNI datasets. The spatial and temporal analysis in both datasets revealed significant differences between SD‐aMCI patients and NC. Specifically, compared to NC, SD‐aMCI patients showed fewer transitions from CAP F+S+ to CAP F–S– in the Chinese dataset. In the ADNI dataset, SD‐aMCI patients showed more frequent transitions from CAP A+ to CAP A– and from CAP A– to CAP A+. All of the above‐mentioned networks related to SD‐aMCI belong to HOCN, suggesting a relatively reliable change in HOCN during the SD‐MCI stage, whereas for the comparisons of MD‐aMCI patients and NCs, the patients showed increased spatial signal strengths of VN in PSN–HOCN‐interacting CAPs and more transitions from CAP V– to CAP V+ in the temporal analysis, which was substantially validated by an independent dataset. Collectively, these findings suggest alterations of HOCN in SD‐aMCI patients but alterations of PSN in MD‐aMCI patients.

Furthermore, a direct comparison between the SD‐aMCI and MD‐aMCI subtypes showed a significant difference of VN in PSN–HOCN‐interacting CAPs. In detail, MD‐aMCI patients showed a greater signal strength in VN than did the SD‐aMCI patients, indicating the importance of the VN in distinguishing aMCI subtypes. The analysis of follow‐up data from the validation dataset also revealed a difference in VN between SD‐aMCI and MD‐aMCI. Prior studies ⁵⁸ , ⁵⁹ that have not specified subtypes of MCI indicate abnormalities in both the PSN and HOCN. Further, by distinguishing subtypes, we found significant differences in these brain networks between SD‐aMCI and MD‐aMCI. The HOCN and PSN are closely connected, and the primary sensorimotor cortex lays the foundation for various cognitive functions. ⁶⁰ When the PSN is damaged, cognitive functions may be affected. ⁵⁸ The increase in dynamic transformation of HOCNs observed in SD‐aMCI patients may indicate compensation for memory loss before extensive cognitive declines occur. When the cognitive declines stretch to multiple domains, the sensory network architecture may be damaged, which may constitute a basis for various cognitive functions to decline. Based on the “last‐in‐first‐out” theory proposed by our previous study that PSN changes earlier than HOSN during development and that PSN changes later than HOCN during healthy aging, ²² the current results may also suggest that the dynamics of HOCN change are earlier than PSNs in the development of AD.

4.3. Spatial difference between AD and aMCI subtypes

The comparison analysis of AD and SD‐aMCI based on the two datasets failed to identify significant differences between groups in temporal scales, while a spatial difference was detected in MN of PSN‐dominant CAPs in the SD‐aMCI group. In the comparison analysis of MD‐aMCI and AD, we observed differences at the spatial scale with disturbances of LN, MN, VN, and SN in PSN‐dominant CAPs, and DMN in PSN–HOCN‐interacting CAPs observed in the MD‐aMCI group. These results collectively suggest that the transition from SD‐aMCI to AD may involve PSN, while MD‐aMCI to AD involves both PSN and HOCN, and the neural basis of the transformation between distinct clinical subtypes of aMCI and AD is different, suggesting a necessity of aMCI subtyping.

Both the abnormalities in PSN and HOCN have been reported across distinct stages of AD. ²⁰ , ⁶¹ , ⁶² Among these networks, the DMN is strongly associated with episodic memory ⁶³ and has shown disruptions in populations with cognitive decline. ⁶⁴ The AN and FPN are activated during goal‐directed tasks and show opposite patterns of activation to those of the DMN. ⁶⁵ During the goal‐directed task, the FPN and AN are activated and the DMN is deactivated to serve the demand of effortful control. ⁶⁶ Both static and dynamic functional changes in the DMN, AN, and FPN have been observed in individuals with the preclinical and clinical stages of AD. ²⁰ , ⁶¹ More specifically, the interactions between the DMN and FPN in specific brain states have been correlated with working memory performance. ⁶² The LN is among the other important networks that were disrupted in AD. ⁶⁷ The functional change in LN has been associated with healthy aging, ⁴⁷ and was considered a biomarker classifying aMCI ⁶⁸ and AD subtype. ⁶⁹ Because AD is the stage of further development of cognitive decline in MCI patients, our results further support the “last‐in‐first‐out” theory on the progression of AD.

4.4. Limitations and future directions

The strengths of this study are the individual parcellation and time‐resolved analysis, which enable the capture of brain fluctuations at specific time points and the brain dynamics of distinct aMCI subtypes. The independent sample validation also facilitates the identification of reliable neural markers. However, several issues still should be addressed in future studies. In view of the scarcity of imaging data on SD‐aMCI and MD‐aMCI, we used the samples selected from the ADNI dataset to verify the results. According to the criteria that SD‐aMCI has memory decline and MD‐aMCI has decline in other cognitive domains besides memory, we classified SD‐aMCI and MD‐aMCI exclusively on the basis of the scale whereas the strict classification should be made by the neurologist combining the patients’ complaints, neuropsychological tests, and scales. The results should be thus interpreted cautiously. In the future, collecting samples with strictly diagnosed aMCI subtypes will allow for a more adequate validation of the results. In addition, the present study used limited time points obtained from a relatively short scan to perform individual parcellations. Based on the evidence that brain parcellations could be affected by the number of scanning time points, ⁷⁰ a collection of datasets with more time points may allow for a more precise identification of the individual boundaries of functional networks. ²⁶

5. CONCLUSIONS

In summary, this study constructed CAPs based on individual parcellations to characterize neurodynamic patterns in distinct aMCI subtypes, and performed comparisons to AD to infer neural differences mediating their transformation to AD. The results indicate that the SD‐aMCI stage involves dynamics in HOCNs, while the MD‐aMCI stage is associated with those in the PSN. The neural networks may be differentially involved in the transformation between aMCI subtypes and AD. Ultimately, this study highlights the importance of the aMCI subtype classification in monitoring the disease progression and clinical strategy of AD.

AUTHOR CONTRIBUTIONS

Tianyi Yan, Tiantian Liu, and Mingjun Wang contributed to the conception of the work, and drafting and revision of the manuscript for intellectual content. Jianghong Liu and Li Wang contributed to the development of the protocol. Jian Zhang conducted the literature search and coordinated the screening process. Tiantian Liu and Mingjun Wang extracted data from eligible papers. Tianyi Yan, Tiantian Liu, Mingjun Wang, and Jian Zhang contributed to the analysis and interpretation of the data. Tianyi Yan, Jianghong Liu, Li Wang, Tiantian Liu, and Mingjun Wang contributed to the writing of the manuscript. Chuyang Ye and Shintaro Funahashi provided study supervision. Tiantian Liu and Tianyi Yan conceived and organized the workshop for this paper and others in the series; obtained funding, contributed to the conception of the work, revised the intellectual content of the manuscript, and coordinated the manuscript with the other papers in the series. Jian Zhang revised the manuscript for intellectual content and harmonized it with other papers in the series. All authors read and approved the final manuscript.

CONFLICT OF INTEREST STATEMENT

All authors report no biomedical financial interests or potential conflicts of interest. Author disclosures are available in the supporting information.

CONSENT STATEMENT

All participants gave written informed consent prior to their participation in the study.

Supporting information

Supporting Information

Supporting Information

Liu T, Wang M, Zhang J, et al. Brain network dynamics in patients with single‐ and multiple‐domain amnestic mild cognitive impairment. Alzheimer's Dement. 2024;20:7657–7674. 10.1002/alz.14227