Abnormal functional connectivity of dynamic brain network in toddlers with autism and its correlation with symptoms

Abstract

Background and objective

The pathogenesis of autism spectrum disorder (ASD) remains unclear. Diagnosis relies on symptoms and relevant scales, which are highly subjective. Moreover, most diagnoses are made after the age of 3, delaying treatment opportunities. This study aimed to explore dynamic functional network connectivity (dFNC) in toddlers with ASD aged 1–3 years, to identify group-level differences in dynamic functional connectivity (FC) patterns and investigate their associations with autistic symptoms, with the goal of providing preliminary neuroimaging insights into the neural mechanisms underlying ASD during this critical developmental window.

Methods

Magnetic resonance imaging (MRI) data of 41 toddlers diagnosed with ASD and 23 age- and sex-matched normal development control (NC) individuals were collected. We examined group differences in dFNC properties, including temporal characteristics of brain states, mean dwell time (MDT), state occurrence rate (OR), and state transition frequency. Correlation analyses were performed to assess the relationships between these dFNC metrics and the severity of autistic symptoms.

Results

Compared with the NC group, toddlers with ASD spent more time in a brain state characterized by globally weak connectivity and a notable lack of between-network functional connectivity, with a significantly higher OR and nominally longer MDT. They also showed a nominally reduced MDT in a brain state marked by strong within-network connectivity and enhanced modular organization across multiple functional networks. Exploratory correlation analyses between autism metrics and dFC measures suggested several nominal associations (uncorrected p < 0.05).

Conclusion

Toddlers with ASD exhibit abnormal dFNC patterns, characterized by altered temporal properties of brain states and disrupted state-related connectivity profiles. These results provide empirical evidence for altered dynamic brain organization in early ASD, contributing to a better understanding of its neural underpinnings during this critical developmental window.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12888-026-08152-4.

Introduction

ASD is one of the fastest-growing neurodevelopmental conditions worldwide [1]. The primary characteristics thereof include difficulties in social interactions, challenges with language communication, and inflexible and repetitive behaviors [2]. ASD may also be accompanied by mental development disorders, sleep disorders, sensory and perceptual abnormalities, and other symptoms [3], which can result in lifelong mental disability.

Children with severe ASD often require substantial caregiving support, imposing significant economic and psychological burdens on families and society. ASD symptoms typically emerge in early childhood and can be identified as early as 2 years of age [4]. These symptoms may persist; however, many children with ASD are misdiagnosed with general intellectual disability or speech impediments [5], thereby missing a critical period for intervention treatment. Early identification, diagnosis, and structured interventions can substantially improve development outcomes in children with ASD [6]. However, as a congenital neurodevelopmental disorder, the current median age at ASD diagnosis remains approximately 49 months across large population-based studies, highlighting persistent delays in early identification [7], which greatly surpasses the optimal period for ASD treatment. Therefore, early identification of ASD in toddlers is imperative.

Advances in neuroimaging are essential for elucidating the neurophysiological mechanisms of ASD and enhancing the accuracy of early diagnosis. Individuals with ASD exhibit both structural and functional brain abnormalities [8–11]. Using region-of-interest and independent component analysis (ICA), FC in brain areas linked to the “social brain” has been shown to be diminished among individuals with ASD [12]. Further, the functional connections between the amygdala and the temporal lobe, as well as between the prefrontal lobe and the striatum, are reduced during the resting state [13], with a decrease in brain-wide FC [14]. Nevertheless, studies have reported inconsistent results. Large-scale consortia such as EU-AIMS, ABIDE, and InFoR have greatly improved the reliability of ASD neuroimaging research. Holiga et al. (2019) analyzed multiple large cohorts and confirmed highly reproducible FC alterations in ASD [15]. However, these large-scale studies mainly used static FC and focused on older children and adults, rather than the critical early period of 1–3 years.

Individuals with ASD exhibit unusual dynamics, particularly in brain function [16–19]. The abnormal brain dynamics of ASD individuals are closely related to their core symptoms. Traditional resting-state FC approaches typically estimate static, time-averaged connectivity patterns and thus fail to capture temporal fluctuations in brain network interactions. In contrast, dFNC explicitly characterizes time-varying changes in functional interactions across the scanning period, enabling a more comprehensive and in-depth assessment of the spatiotemporal organization of functional brain networks in individuals with ASD [20]. Abnormalities in dynamic functional interactions between the posterior cingulate gyrus and sensorimotor cortex in individuals with ASD are significantly associated with social difficulties [21]. Further, when the dynamic FC between the right anterior insula and the default mode network (DMN) decreases, social interaction ability is affected in ASD individuals [22]. Studies have used entropy analysis to quantify the complexity of the brain and information flow, and have found abnormalities in toddlers with ASD [23]. These findings suggest that from the perspective of dynamic FC, further exploration of the time-varying characteristics of brain networks and their association with early symptoms is an important current direction. Although dFNC has been investigated in older individuals with ASD, its application in toddlers during the 1–3 year critical window remains largely unexplored, and the present study aims to address this gap.

This study focused on toddlers aged 1 to 3 years old. By combining spatial ICA and dynamic FC correlation, it evaluated the dynamic changes in the whole-brain functional network connectivity (FNC) of ASD toddlers aged 1 to 3 years old. K-means clustering was then applied to the correlation matrices generated by the sliding windows to identify the dynamic patterns of FC. We further employed a network-based statistical (NBS) approach to analyze the differences in the FNC between the groups under each connectivity state. Toddlers with ASD and age- and sex-matched NC children were compared to investigate the temporal variability of FC between the two groups. Our results enhance understanding of the functional organization and information processing of the brain in the resting state, and offer fresh perspectives on the neural mechanisms underlying ASD in toddlers, with potential implications for the future development of neuroimaging markers that could contribute to early identification.

Materials and methods

Study participants

Forty-one toddlers with autism (27 males, age 19–45 months) treated in the Child Health Care Department of Hubei Maternal and Child Health Hospital, and 23 NC individuals (15 males; age 15–45 months) matched by age, sex, and handedness were enrolled.

Children were included if they met the benchmark for diagnosis based on the Diagnostic and Statistical Manual of Mental Disorders (DSM-V) [24–25] and had a CARS score > 30. Individuals with epilepsy, cerebral palsy, various organic brain disorders, and other neurodevelopmental conditions were excluded. All participants diagnosed with ASD were verified not to have any recognized genetic syndromes linked to ASD, such as Fragile X syndrome. The admission requirements for the NC group were as follows: no DSM-V related diagnosis; development quotient (DQ) ≥ 85; no record of psychiatric hospitalization, medication treatment, or any diagnosis of ASD, schizophrenia, psychosis, or bipolar disorder in first-degree relatives.

Measurements of behavior

The ABC was applied to assess symptoms of ASD, focusing on sensory issues, relating, body awareness, language, and social self-help. All ASD toddlers scored at least 68 on the ABC, with an average score of 84.3 ± 7. All children further exhibited significant clinical symptoms, such as stereotypical behaviors, communication difficulties, and/or language regression.

The Gesell Scale for Children was applied to evaluate DQ, including areas of adaptability, gross motor skills, fine motor skills, language development, and social-emotional responses. The following classifications were used to mark the development of toddlers based on the full-scale DQ: normal (DQ ≥ 85), deficient (DQ < 75), and borderline (75–85). A DQ < 75 in any area indicated a deficiency. The CARS was applied to assist in the diagnosis of ASD and assess severity. ASD can be diagnosed at a CARS score ≥ 30, while a higher CARS score indicates more severe ASD.

MRI protocol

All children underwent MRI scanning under sedation. Oral 10% chloral hydrate was administered at a dose of 50 mg/kg (maximum dose 100 mg) approximately 30 min prior to scanning. MRI data were acquired only after the child had fallen asleep and achieved stable sedation. An identical sedation protocol was applied to both the ASD and control groups, with consistent dosing, timing, and monitoring criteria across all participants. No differential requirement for additional sedation was observed between groups. Resting-state functional MRI (fMRI) data were gathered using a 3T MRI scanner (United Imaging u780), equipped with an 8-channel phased-array head coil. A total of 240 volumes of echo-planar images were captured using the following settings: repetition time (TR)/echo time (TE) = 2,000/30 ms, field of view (FOV) = 240 × 240 mm (64 × 64 matrix), flip angle (FA) = 90°, and 40 sequential ascending axial slices with a thickness of 3 mm.

Data preprocessing

Resting-state fMRI data were preprocessed using the Data Processing and Analysis of Brain Imaging (DPABI) toolbox [26]. The steps were as follows: (1) removal of the first 10 volumes to allow for signal equilibration; (2) slice-timing correction and head-motion realignment, participants with > 20% flagged volumes, translation > 2.5 mm, or rotation > 2.5° were excluded (no toddlers met these criteria); (3) spatial normalization to the Montreal Neurological Institute space template using the DARTEL algorithm and resampling to 3 mm isotropic voxels; (4) smoothing with a 5 mm full width at half maximum Gaussian kernel; (5) nuisance regression, including the Friston-24 head motion parameters, mean white matter and cerebrospinal fluid signals, and linear and quadratic trends (global signal was not regressed). No significant group difference was found in mean FD (ASD: 0.07 ± 0.05 mm; NC: 0.08 ± 0.08 mm), and mean FD was included as a covariate in all subsequent analyses.

Group ICA

Following data preprocessing, functional data from the NC and ASD groups were decomposed into spatially independent components (ICs) via two-stage group information-guided ICA (GIGICA) (initial standard group ICA for group-level independent component (IC) templates, followed by subject-level back-reconstruction), implemented using the GIFT toolbox (http://mialab.mrn.org/software/gift) [27]. For each individual, the data were initially condensed into 150 primary components through Principal Component Analysis (PCA). While reducing data at the group level, the dataset created by amalgamating the dimensionally reduced data of individual subjects was broken down into 100 components using PCA. The Infomax algorithm was repeated 20 times in ICASSO [28–30], while the most central run was chosen to ensure the stability of the estimation. The reverse reconstruction method was applied to estimate the spatial map (SM) and time courses (TC) of specific subjects. Of the 100 components analyzed, 55 were identified as physiologically relevant ICs. We have compiled detailed information for 55 ICs, including the component numbers, peak coordinates, and the functional networks to which they belong (Supplementary Table S1).These ICs were characterized by a peak spectral amplitude fluctuation (AF) within the low-frequency power spectrum, while their primary activation was localized to the gray matter regions, showing little spatial overlap with the white matter, ventricles, and peripheral regions, ensuring distinct delineation [31]. The 55 ICs were categorized into seven functional networks. The subcortical network (SUB), auditory network (AUD), somatomotor network (SMN), visual information system (VIS), cognitive control network (COC), DMN, and cerebellar network (CB) (Fig. 1A). Prior to calculating FC, the subject-specific TC for each IC underwent standard post-processing within the GIFT toolbox. This included removing linear, quadratic, and cubic trends and temporal band-pass filtering (0.01–0.1 Hz) to isolate low-frequency resting-state fluctuations. Pairwise Pearson’s correlations were then calculated using these cleaned TCs throughout the entire scan to create a static FNC matrix (Fig. 1B).

Fig. 1.

Spatial maps from the 55 ICs and static functional connectivity network. (A) Fifty-five ICs divided in seven functional networks. (B) The static FNC between ICs in the whole sample. Index numbers of ICs are written on the left side of the matrix, along with a color-coded legend, which corresponds to the overlaid colors of the spatial maps in (A). The color bar represents the value of the correlations; red represents positive correlations and blue represents negative correlations

Dynamic functional network connectivity

The dFNC was determined using a sliding window approach through the dFNC toolbox in GIFT. The window was set to 22 TRs (44 s) with a step size of 1 TR. Specifically, a tapered window was applied, created by convolving a rectangular window with a Gaussian function (σ = 3 TRs), to minimize spectral leakage. Covariance was then calculated for each window across all pairwise ICs for every participant.

To identify the optimal number of unique connectivity states, we evaluated cluster solutions with K values ranging from 2 to 10 using the elbow criterion as implemented in the GIFT toolbox. The elbow point was clearly visible at K = 4 (Supplementary Fig S1). Subsequently, K-means clustering with K = 4 was applied to the individual regularized covariance matrices, using Manhattan distance (L1 norm) and 500 replicates with random initializations, following default parameters in the GIFT dFNC toolbox. Covariance matrices were regularized using graphical LASSO (regularization parameter = 0.05) prior to being converted to correlations and Fisher z-transformed to improve normality and stabilize variance across windows.For each participant, several dFNC indices were calculated: OR: occurrence rate, percentage of time each dFNC state was observed; MDT: mean dwell time, average duration spent in each state before transitioning to the next; NT: total number of times a participant switched states.

Statistical analyses

All analyses were performed using statistical software (SPSS 22.0; SPSS Inc., Chicago, IL). A two-sample t-test was employed to compare the age and clinical characteristics of toddlers with ASD and those in the NC group. The Pearson’s chi-square test was used to assess sex differences. Additionally, a two-sample t-test was utilized to compare the dynamic metrics between dFNC state transitions, controlling for age, sex, and mean FD as covariates. DQ was not included as a covariate given its strong collinearity with ASD diagnosis, which could lead to overcorrection of disease-related neural effects. Data normality was assessed via the Shapiro–Wilk test, and all t‑test data were normally distributed.

The NBS approach was utilized to identify the group differences in FNC between children with ASD and NC [32]. NBS controls the family-wise error (FWE) rate at the level of interconnected network components. The application of NBS to our dynamic states proceeded as follows:

First, to reflect the connectivity pattern of each participant in a particular state, a median connectivity matrix was calculated based on the specific participant windows allocated to that respective K-means cluster. Nodes were established based on the peak coordinates of the ICs, while edges were determined by the correlation values from these state-specific FNC matrices. Next, two-tailed t-tests were conducted for every edge between the ASD and NC groups, and an initial primary t-statistic threshold of t > 2.8 was applied to define a set of suprathreshold connections. Finally, to determine the statistical significance of any observed subnetworks (components), permutation testing (5,000 permutations) was performed. During each permutation, the size of the largest connected component was noted to create an empirical estimate of the null distribution, allowing us to calculate an FWE-corrected P-value for every component based on its size. Crucially, this NBS permutation correction (contrasts: ASD > NC and ASD < NC) was applied independently for each of the four dFNC states. FWE correction was applied for each state separately, with corrected P < 0.05 considered statistically significant, but no correction was applied across the four separate state comparisons.”

For dFNC metrics (OR and MDT), two-sample t-tests were performed across the four states, with the False Discovery Rate (FDR) correction applied to adjust for multiple comparisons across the four states. Statistical significance was defined as an FDR-adjusted P < 0.05, while an FDR-adjusted P-value between 0.05 and 0.10 was considered to indicate a nominal trend. Finally, Pearson correlation analyses were conducted to examine the associations between dFNC metrics and total scores of the ABC and CARS, as well as five subscale scores of the ABC. Uncorrected P-values are reported, with a P < 0.05 considered to indicate an exploratory trend.

Results

Demographics characteristics and behavior

There were no significant variations in sex or age between the ASD and NC groups (Table 1). Additionally, conventional MRI did not reveal any significant brain lesions in either group. The ASD group exhibited considerably lower scores on the DQ assessments compared to the NC group (P < 0.001).

Table 1.

Characteristics of participants in ASD and normal development control groups

Characteristic	ASD (n = 41)	NC (n = 23)	P-value
Sex (Male/Female)	27/14	15/8	0.959
Age (months, Mean ± SD)	32 ± 7	32 ± 9	0.998
DQ score	65 ± 8	88 ± 4	< 0.001*
ABC score	83.4 ± 7	-	-
CARS score	35 ± 3	-	-

ASD, autism spectrum disorder; NC, normal development control; DQ, development quotient; ABC, Autism Behavior Checklist; CARS, Childhood Autism Rating Scale

Dynamic FNC differences

K-means clustering with k = 4 identified four distinct connectivity states (States 1–4). (Fig. 2A). Connections with an absolute value > 0.2 for each state were retained for better visualization to distinguish different patterns of the four states (Fig. 2B).

Fig. 2.

Dynamic FNC states in the whole sample. (A) Cluster centroids for each state. The total number of occurrences and percentage of total occurrences are listed for each state. The color bar represents the z value of the correlations, wherein red represents positive correlations and blue represents negative correlations. (B) Connections with an absolute value exceeding 0.2 of each state. Each rectangle on the circumference of the big circle represents each IC, and each rectangle color represents one of the seven networks.Colored lines indicate connections within the same functional network, while gray lines indicate connections between different functional networks

Compared to NCs, children with ASD exhibited a significant increase in the OR (t = 2.76; FDR-adjusted P = 0.031) (Fig. 3B) and strong trend toward an increase in the MDT (t = 2.37; FDR-adjusted P = 0.057) in State 1 (Fig. 3A), characterized by globally weak functional connectivity across the whole brain, with a marked deficit in between-network connections (Fig. 2A). Conversely, in State 4, children with ASD showed a nominal reduction in the MDT (t = − 2.24; uncorrected P = 0.026, FDR-adjusted P = 0.057) compared to NCs (Fig. 3A). State 4 was distinguished by stronger connectivity, demonstrating modularization within various networks. Further, there was no notable variation in the frequency of transitions between states (Fig. 3C).

Fig. 3.

Comparisons between ASD and NC groups in dynamic indices in each dFNC state. The left violin figure is a comparison of the mean dwell time for each state of the two groups, the middle violin figure is a comparison of the occurrence rate for each state of the two groups, the right violin figure is a comparison of the number of transitions of the two groups. The P-values shown above represent FDR-adjusted P-values

Additionally, differences in FNC between the groups were calculated for each dFNC state, with disparities emerging in State 1. A total of 41 children with ASD and 21 NCs provided valid data windows for State 1. In contrast, valid data windows for the other states were provided by fewer participants (see Supplementary Table S2). No significant results were observed in states 2 to 4, which might be due to the small sample size in these states. A larger subnetwork with decreased connectivity and two subnetworks with increased connectivity were identified. The reduction in connectivity primarily occurred between the COC and other functional networks (P = 0.0044), between the COC and VIS, and between the COC and SMN (Fig. 4A). Conversely, the two subnetworks with increased connectivity involved enhanced connections between the DMN and other networks (P = 0.0388; P = 0.0466), showing an increase in connections between the DMN and the COC, VIS, and SMN (Fig. 4B-C).

Fig. 4.

Results of dFNC analysis using the NBS approach. (A) FNC reductions in State 1 in ASD compared to NC. (B–C) FNC increase in State 1 in ASD compared to NC. (D–F) Location of increased and decreased connections in the brain using the BrainNet Viewer software (http://www.nitrc.org/projects/bnv/)

Behavioral variables

Exploratory correlation analysis indicated a nominal positive trend between the CARS score and the MDT in State 1 (r = 0.337, uncorrected P = 0.034). Similarly, the MDT in State 4 showed an nominal negative trend with the sensory subsection of the ABC score (r = − 0.48, uncorrected P = 0.044). Moreover, there was a nominal negative trend between the NT across different states and the CARS scores (r = − 0.35, uncorrected P = 0.027) (Fig. 5).

Fig. 5.

Correlation between neuroimaging indices and behavioral scores in ASD group. (A) The CARS score was correlated positively with the mean dwell time in State 1. (B) The sensory portion of the ABC score was negatively correlated with the mean dwell time in State 4. (C) Number of transitions between states was negatively correlated with the CARS score

Discussion

We examined irregularities in resting-state dFNC in toddlers with ASD, and defined four different states. The NC and ASD groups showed marked differences in OR and MDT values in individual states, and the detected irregularities in dFNC were linked to assessments of ASD symptoms and severity. NBS methods further identified intergroup differences in FNC between the ASD and NC groups. ASD toddlers had a larger subnetwork with decreased connectivity, as well as two subnetworks with increased connectivity. The outcomes of the dFNC assessment highlight distinct neuroimaging characteristics associated with ASD in toddlers and support the potential value of these approaches for understanding the neural basis of ASD at early ages.

Compared with NCs, toddlers with ASD exhibited a significant rise in OR and a corresponding trend in MDT in State 1, characterized by globally weak connectivity and a breakdown of integration between networks. This state was defined by a profound deficiency in connections between networks and globally reduced connectivity across both cortical and subcortical regions. Conversely, in State 4, toddlers with ASD demonstrated a nominal reduction in MDT compared with NC. State 4 was distinguished by stronger overall connectivity, demonstrating modularization within various networks. These results indicated that the weak connection state occupied a higher percentage in children with ASD, with a trend toward longer duration, while the strong connection state tended to be shorter. In other words, in the resting state in toddlers with ASD, the connections between the dynamic brain networks were significantly weakened, the synchronization of activities between brain networks was significantly reduced, aligning with previous findings [35–37]. Further, previous studies comparing dFNC across a broad age range found no notable variations in MDT between ASD and NC children [38], which differs from the findings of the present study focusing on ASD toddlers. This discrepancy may be due to the wide age range (from 5 months and older) utilized in previous research.

Individuals exhibiting more autistic traits show weaker static connections [39], whereas children with fewer autistic traits tend to spend more time in a globally modularized state. In the present study, correlation analyses between dynamic indices and CARS scores in the ASD group further revealed exploratory trends where the CARS score nominal positively correlated with MDT in State 1 and negatively correlated with NT across brain states. Given that higher CARS scores represent more severe ASD symptoms [40], these findings support the notion that more severe autistic symptoms are closely linked to more severely desynchronized dFNC. Specifically, children with ASD exhibited less frequent state transitions and less stable dynamic connectivity patterns, which may reflect impaired temporal coordination and aberrant information transmission within large-scale brain networks.

Furthermore, within State 1, children with ASD showed a larger subnetwork with decreased connectivity along with two subnetworks with increased connectivity. Importantly, the NBS approach identifies specific, localized connections rather than only global internetwork or intranetwork connectivity, and the identified subnetworks were predominantly centered on connections involving the middle frontal gyrus (MiFG), a core region within the COC network [41, 42].

First, a reduction in connectivity primarily occurred between the COC and other functional networks, specifically between the COC and VIS and the COC and SMN. The main regions of the COC include the MiFG, dorsolateral prefrontal cortex, and posterior parietal cortex, which are critical for cognitive control, attentional regulation, and top-down modulation of sensory and motor processing [43]. Accumulating evidence has demonstrated that MiFG and COC dysfunction is central to the neuropathological basis of ASD, with individuals with ASD exhibiting weakened COC functional connectivity during both resting state and goal-oriented tasks compared with typical controls [36, 44, 41]. The VIS includes several neurological regions, including the fusiform gyrus, inferior temporal gyrus, middle occipital gyrus, and inferior parietal lobule, which work collaboratively to support visual signal processing and information transmission across the brain, thereby shaping human behavior and cognition [45]. The SMN is involved in processing both external stimuli and internally generated movements [46]. Individuals with ASD exhibit aberrant intrinsic FC within visual and motor regions [47]. Our control group demonstrated higher FNC between the COC (centered on MiFG) and VIS, as well as between the COC and SMN, compared with the ASD group. These findings indicate that children with ASD show weakened communication between cognitive control regions (particularly MiFG), visual perceptual systems, and motor networks, which may contribute to core sensory-motor and cognitive impairments in ASD [42, 48].

The two subnetworks with increased connectivity involved enhanced connections between the DMN and the COC, VIS, and SMN. The DMN shows strong spontaneous activity at rest [49] and is prominently activated during internally focused tasks [50], while its activity is reduced during externally directed cognitive processing [51]. A close overlap between the DMN and the “social brain” network has been well documented [52], and whole-brain FC studies in children with ASD have consistently identified dysconnectivity within social brain circuits [53]. Furthermore, the FC strength of the anterior-posterior DMN subnetwork in ASD is negatively associated with social-emotional deficits and executive dysfunction [54–56]. Normally, the DMN exhibits functional segregation from networks including the COC, VIS, and SMN [57]. However, children with ASD in the present study showed enhanced connectivity between the DMN and these networks, suggesting reduced functional segregation and less efficient switching between internal and external attention [58]. Notably, these hyper-connected pathways were also anchored by the MiFG, highlighting this region as a key hub underlying disrupted network dynamics in young children with ASD [41, 42, 48].

Approximately 69–90% of children with ASD exhibit paresthesia [59], a phenotype now recognized in DSM-V, and this sensory dysfunction correlates with language development delays [60]. Sensation seeking relates to thalamic activity [61] and medial orbitofrontal-anterior cingulate connectivity [62]. We found that ABC sensory scores nominal negatively correlated with MDT in State 4-a highly modular connectivity state-suggesting that aberrant sensory perception in ASD may stem from insufficient inter-network integration. However, substantial inter-subject variability was observed (as illustrated in Fig. 5B), highlighting the need for caution when interpreting such associations at the individual level.

Notably, these results align with and extend the large‑scale consortium study by Holiga et al. (2019) [15], which identified reproducible static FC alterations across major ASD cohorts including EU‑AIMS, ABIDE, and InFoR. Holiga et al. reported reduced intra‑network connectivity and increased inter‑network coupling-particularly between the default mode and sensorimotor networks-in older individuals with ASD, and attributed sensory symptoms to disrupted network specialization. In contrast, our study focused on dynamic FC in toddlers aged 1–3 years and revealed abnormal state transitions, altered dwell times, and unstable network modularity during the earliest symptomatic period. While Holiga et al. described a stable, reproducible static connectivity signature in older populations, we demonstrate that the foundational network dysregulation emerges in infancy as atypical temporal dynamics rather than only fixed static differences.

Together, our findings support disrupted network modularity as a core developmental mechanism underlying early sensory features in ASD, and extend the model of Holiga et al. to the earliest stages of ASD pathogenesis.

Several limitations warrant consideration. First, our study included children aged 1–3 years, an age at which ASD is often under‑recognized or undiagnosed clinically. Although all ASD diagnoses were verified per DSM-V criteria at enrollment, we cannot fully rule out undiagnosed ASD or subthreshold autistic traits in control children that may manifest at older ages. Second, early-identified toddlers with ASD may constitute a subgroup with more severe symptoms, potentially limiting generalizability to individuals with milder or later‑onset presentations. Third, the relatively large voxel size and the use of an adult-based template for spatial normalization may introduce spatial distortions and amplify partial volume effects, reducing the overall spatial precision of functional connectivity measurements in toddler brains. Fourth, although the acquisition protocol was standardized across groups, resting-state fMRI data were collected under sedation or natural sleep, which may have influenced FC measures. No sensitivity analyses were performed on window length or the number of clusters (k); future studies should therefore assess the robustness of these parameters. Additionally, DQ was not statistically controlled, given its high collinearity with ASD diagnosis, which could lead to overcorrection of neurobiological signals associated with the disorder. These diagnostic, sampling, and imaging constraints should be considered when interpreting the results. Future longitudinal studies with extended follow-up or prospective high-risk cohorts are needed to better characterize the full early ASD phenotype.

Conclusions

ASD has an early onset and a complex, unclear pathogenesis. Currently, early diagnosis primarily depends on clinical evaluation and symptom observation, which are subjective, commonly resulting in missed diagnosis, thereby delaying treatment beyond the optimal window for children with ASD. We found that the whole-brain network connectivity of toddlers with ASD was disrupted and dysregulated. These results enhance our comprehension of the neural processes associated with early-onset ASD in children under 3 years of age, providing preliminary neuroimaging insights that may inform future research toward early identification. Future longitudinal studies are needed to determine whether such dynamic connectivity measures may contribute to early identification when combined with behavioral assessments.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ASD

Autism spectrum disorder

dFNC

Dynamic functional network connectivity

FC

Functional connectivity

MRI

Magnetic resonance imaging

NC

Normal development control

MDT

Mean dwell time

OR

Occurrence rate

CARS

The Childhood Autism Rating Scale

NT

Number of state transitions

ABC

Autism Behavior Checklist

ICA

Independent component analysis

DMN

Default mode network

NBS

Network-based statistical

DSM-V

Diagnostic and Statistical Manual of Mental Disorders

DQ

Development quotient

fMRI

Functional MRI

DPABI

Data Processing and Analysis of Brain Imaging

ICs

Independent components

IC

Independent component

PCA

Principal Component Analysis

SM

Spatial map

TC

Time courses

AF

Amplitude fluctuation

SUB

The subcortical network

AUD

Auditory network

SMN

Somatomotor network

VIS

Visual information system

COC

Cognitive control network

CB

Cerebellar network

FWE

Family-wise error

MiFG

Middle frontal gyrus

Author contributions

Shengjuan Wu was in charge of the design of the experiment, data collection, data analysis and article writing; Xiuxian Li participated in data processing and statistical analysis; Fuchun Lin participated in functional magnetic resonance data processing and analysis; Wenzhong Yang and Chengcheng Jiang were responsible for the acquisition of functional magnetic resonance data; Zhi Wen participated in the setting of functional magnetic resonance parameters; Henry Xiang participated in article writing; Zhiwei Zhao, Aiqin Zhou, Xinglian Liu, Xiaoyan Wang, Meirong Wu and Yuechi Shen participated in the collection of clinical data; Hong Wang was responsible for the design of the experiment, financial support and article writing.

Funding

This research received funding from the Hubei Provincial Top Medical Young Talents Project (Grant No. 10200112027), Hubei Province Association of Psychological Counselors and Intra-hospital Project of Hubei Provincial Maternal, and Child Health Care Hospital (2023SFYM013).

Data availability

The datasets used and analyzed in the current study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of Hubei Maternal and Child Health Hospital. The research was conducted in accordance with the Declaration of Helsinki. Informed consent was obtained from the parents of all participating children prior to participation. Clinical trial number: not applicable.

Consent for publication

We obtained informed consent of the individuals’ parents to publish this article.

Competing interests

The authors declare no competing interests.