Infant intellectual disorder early warning method based on white matter network propagation dynamics abnormity

Abstract

The invention discloses an infant dyspepsia early warning method based on cerebral white matter network propagation dynamics abnormity, and relates to the technical field of brain image analysis and intelligent risk prediction. According to the method, firstly, standard preprocessing is carried out on infant brain MRI or DWI image data, and a structural connection network is constructed based on a newborn template; then, a linear threshold model (LTM) is introduced to simulate the diffusion process of information in the brain network, and multiple propagation dynamic characteristics including propagation time, diffusivity, cooperative speed-up ratio, competitive indexes and the like are extracted; based on the above characteristics, a random forest regression model is used to predict cognition, language and movement development scales of 18 months old, and a key propagation brain region is identified through characteristic importance analysis. According to the method, non-invasive, automatic and structure-driven early development risk assessment can be realized, and the method has relatively high generalizability and clinical application potential.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of medical image analysis, brain network modeling, and intelligent development risk prediction, and in particular to an early warning method for cognitive impairment in infants and young children based on the dynamic characteristics of white matter propagation. Background Technology

[0002] Currently, early identification of intellectual disability in children relies heavily on traditional imaging anatomical indicators or clinical behavioral scores, failing to effectively capture the dynamic abnormalities in the white matter network during information flow and functional integration. Existing methods primarily focus on the morphology, volume, or static network parameters (such as connectivity and network efficiency) of white matter lesions, making it difficult to reflect information transmission barriers in the real brain network. This results in the difficulty in timely detection and intervention of intellectual disability risks in some high-risk infants (such as premature infants and low birth weight infants). Numerous studies have shown that abnormalities in the microstructural development and network integration of the white matter are key foundations for intellectual developmental delays and various neurodevelopmental disorders, especially manifesting as latent transmission dysfunction in early life. Currently, there is a lack of an engineerable and automatically quantifiable method for early identification of white matter network transmission abnormalities. Clinically, there is an urgent need for an integrated new technology that combines advanced network transmission dynamics modeling, intelligent image analysis, and risk prediction to achieve early warning and risk stratification of intellectual disability. Summary of the Invention

[0003] This invention aims to address the shortcomings of existing methods for early identification of intellectual disabilities, such as difficulty in accurately capturing abnormalities in the propagation of white matter networks, lack of sensitive dynamic indicators, and poor engineering feasibility. It proposes a method for early warning of intellectual disabilities in infants and young children based on abnormalities in the propagation dynamics of white matter networks. This method can automatically analyze brain structural images of infants and young children to quantitatively identify information flow obstacles, enabling early warning and intervention guidance for high-risk individuals.

[0004] The specific implementation scheme of this technology is as follows: A method for early warning of intellectual disability in infants and young children based on abnormal propagation dynamics of white matter networks, the method includes the following steps:

[0005] Step 1: Acquire raw data of T2-weighted structural images (T2W) and diffusion-weighted images (DWI) of the infant's brain; extract brain tissue and remove artifacts from the T2W images and register them to the standard space; register the DWI images to the T2W space.

[0006] Step 2: Using partition templates and high-throughput fiber tracing algorithms, construct individual-level brain white matter structural connectivity networks to form a unified weighted adjacency matrix, ensuring the consistency and repeatability of region segmentation and fiber tracing.

[0007] Step 3: Propagation dynamics simulation and index extraction;

[0008] Based on the weighted adjacency matrix, a threshold propagation model is established, and single-point and two-point perturbation simulations are performed on the whole brain network to calculate the dynamic indicators of each node, including propagation time, propagation capacity, cooperative speedup ratio, and competitiveness.

[0009] Step 4: Feature extraction and developmental score prediction;

[0010] Based on the aforementioned single-point and two-point propagation dynamics simulations, multidimensional propagation indicators for each brain region are extracted. These indicators are used as features, and a random forest regression machine learning method is employed to predict the cognitive, language, and motor development scores of infants at 18 months of age. A five-fold cross-validation strategy is used to evaluate the model performance, and statistical indicators such as the correlation between predicted and actual values ​​are output.

[0011] Furthermore, the method for constructing the weighted adjacency matrix in step 2 is as follows:

[0012] Step 2.1: For each infant subject, based on the preprocessed diffusion-weighted imaging data, the high-throughput fiber tracing algorithm was used to automatically reconstruct the white matter fiber bundles in the brain;

[0013] Step 2.2: The whole brain was segmented into 90 anatomical functional areas using a standard partitioning template suitable for the structural characteristics of the neonatal brain. Based on this, the number of fiber bundles connecting each pair of brain areas by fiber tracing results was counted, and the following normalization strategy was used to construct a 90×90 structural connectivity matrix at the individual level:

[0014]

[0015] in, This represents the normalized structural connectivity strength between brain regions i and j in the k-th subject; The number of fiber bundles connecting the two regions. The average length of the connecting fibers. This represents the average surface area of ​​the cortical regions of the two brain areas.

[0016] Step 2.3: Average the individual structural connectivity matrices of all subjects to construct a group-level structural connectivity network, forming a uniform 90×90 weighted adjacency matrix.

[0017] Furthermore, the single-point disturbance model in step 3 is as follows:

[0018] All nodes are initially set to inactive (represented by 0); the selected seed node is set to active (represented by 1); and the state of each node is updated according to the following dynamic rules during the iteration process:

[0019]

[0020] Where: x i (t) represents the state of node i at time t; w ij K represents the connection weight between nodes i and j; i =∑ j w ij θ represents the total input received by node i; θ is the normalized propagation threshold, used to unify and compare the propagation dynamics process.

[0021] During the propagation simulation, the minimum time step t required for any node j in the network to be activated from seed point i is recorded. ij Construct the propagation time matrix T = [t ij ];

[0022] After obtaining the propagation time matrix, the following key metrics are defined:

[0023] external propagation time out i for:

[0024]

[0025] Where N represents the total number of nodes, out i The speed at which node i, as a seed point, influences the propagation throughout the brain;

[0026] Internal propagation time i for:

[0027]

[0028] in i It reflects the sensitivity of node i to external disturbances;

[0029] Transmission capability SA i for:

[0030] SA i =max{θ i |f(i,θ i )≥f c}

[0031] Where, θ i f(i,θ) represents the threshold at which node i can trigger a global cascade. i ) indicates that at the threshold θ i The following is the proportion of activated nodes in the network when node i is used as the seed, f. c This indicates the proportion of nodes that are set to be active. When f c When SA = 1, a global cascading mechanism is triggered, meaning that all nodes in the network must be activated. iThis represents the upper limit of the minimum threshold required to trigger a global cascade when node i becomes a seed point. A higher threshold indicates that a global cascade can still be triggered under more difficult conditions.

[0032] Furthermore, in step 3, the two-point perturbation is a two-point cooperative perturbation, and the two-point cooperative perturbation model is as follows:

[0033] For each pair of nodes (i,j), simultaneously set them to active, simulate the two-point propagation process, and record the propagation time A required for their cooperative propagation. ij Define the speedup ratio S for collaboration as:

[0034]

[0035] Wherein, min(A) i A j ) indicates that option A is selected. i A j The minimum value in, A i A j The outer diffusion time that triggers global cascading when nodes i and j are used as seed points alone; the larger the value, the more obvious the acceleration effect.

[0036] Furthermore, in step 3, the two-point perturbation is a two-point competition perturbation, and the two-point competition perturbation model is as follows:

[0037] Let nodes i and j be two competing seed nodes in different states. During propagation, each node is allowed to be occupied by only one propagation cluster, ultimately forming a two-cluster propagation structure; define the competitiveness index C. i for:

[0038]

[0039] Where S ij The size or number of seed points that spread during the competition between seed point i and seed point j.

[0040] The method of this invention can sensitively capture subtle abnormalities in the propagation dynamics of the brain's white matter network and quantitatively predict the long-term cognitive, language, and motor development outcomes of infants and young children using multiple indicators, greatly improving the efficiency of early screening and precise intervention for high-risk individuals in clinical practice. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0042] Figure 2 A visual diagram illustrating the extraction of structural network and propagation dynamics indicators of the white matter in the brain of infants and young children.

[0043] Figure 3 This is a schematic diagram of the results of developmental outcome prediction and feature importance analysis. Detailed Implementation

[0044] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0045] For detailed implementation steps, please refer to... Figure 1

[0046] Step A: Magnetic Resonance Data Acquisition and Preprocessing

[0047] The data used in this study were obtained from the dHCP (Developing Human Connectome Project) database, including neonatal T2 structural images, diffusion-weighted imaging (DWI), and resting-state functional magnetic resonance imaging (rs-fMRI). A total of 310 full-term infants and 87 preterm infants were included. Scans were performed at the Evelina Imaging Centre at St. Thomas Hospital in London, UK, using a Philips Achieva 3.0T MRI scanner and a 32-channel neonatal head coil. T2 structural images: acquisition resolution 0.8 × 0.8 × 1.6 mm. 3 The reconstructed image is a 0.8 mm isovoxel. DWI parameters: multi-shell design, b-values ​​of 400, 1000, and 2600 s / mm. 2 300 directions, resolution 1.5×1.5×3mm 3 T2 structural image preprocessing: First, brain extraction and artifact removal were performed on the T2W structural images, followed by linear and nonlinear registration to the 40-week-old T2 standard template. Then, tissue segmentation was performed on the T2 images, extracting gray matter, white matter, and cerebrospinal fluid mask. DWI processing workflow: This included denoising, artifact removal, head motion correction, distortion correction (SHARD algorithm), and NDC value filtering, followed by fiber tracking based on GQI reconstruction using DSIStudio. Tracking parameters were: angle threshold 60°, step size 1mm, fiber length 30–300mm, generating 1,000,000 fibers.

[0048] Brain region segmentation and group-level connectivity matrix construction: First, for each infant subject, based on preprocessed diffusion-weighted imaging (DWI) data, a high-throughput fiber tracing algorithm was used to automatically reconstruct white matter fiber tracts. Then, a standard partitioning template suitable for neonatal brain structural characteristics (such as a modified AAL brain region template) was used to segment the whole brain, dividing it into 90 anatomical functional regions. Based on this, the number of fiber tracts connecting each pair of brain regions by fiber tracing results was counted, and the following normalization strategy was used to construct an individual-level structural connectivity matrix (90×90):

[0049]

[0050] in, This represents the normalized structural connectivity strength between brain regions i and j in the k-th subject; The number of fiber bundles connecting the two regions. The average length of the connecting fibers. This represents the average surface area of ​​the cortical regions of the two brain regions.

[0051] After the above normalization is completed, the individual structural connectivity matrices of all subjects are averaged to construct a group-level structural connectivity network, forming a uniform 90×90 weighted adjacency matrix. This group-level structural connectivity network serves as the basic topological foundation for the propagation modeling process of this invention, and is used to subsequently simulate the information diffusion process and extract key propagation dynamics indicators.

[0052] Step B: Construction of the cascade propagation model and extraction of its propagation dynamics indices

[0053] In one embodiment of the present invention, based on the constructed brain structure connectivity matrix, a linear threshold model (LTM) is used to model the diffusion process of information in the structural brain network, and various quantitative indicators are extracted based on different perturbation scenarios.

[0054] 1: Single-point disturbance modeling

[0055] All nodes are initially set to inactive (0), a single node is selected as the seed node and set to active (1), and the state of each node is updated according to the following dynamic rules during the iteration:

[0056]

[0057] Where: x i (t) represents the state of node i at time t; w ij K represents the connection weight between nodes i and j; ij =∑ j w ij θ represents the total input received by node i; θ is the normalized propagation threshold, used to unify and compare the propagation dynamics process.

[0058] During the propagation simulation, the minimum number of time steps required for any node j in the network to be activated from seed point i is recorded, and the propagation time matrix T = [t] is constructed. ij ].

[0059] After obtaining the propagation time matrix, the following key metrics are defined:

[0060] External transmission time:

[0061]

[0062] This reflects the speed of influence of node i as a seed point in the whole-brain propagation.

[0063] Internal propagation time:

[0064]

[0065] It reflects the sensitivity of node i to external disturbances.

[0066] Dissemination capability:

[0067] SA i =max{θ i |f(i,θ i )≥f c}

[0068] This represents the upper limit of the minimum threshold required to trigger a global cascade when node i becomes a seed point; a higher value indicates that it is more difficult to trigger a global cascade.

[0069] 2: Two-point cooperative perturbation modeling

[0070] For each pair of nodes (i,j), simultaneously set them to active, simulate the two-point propagation process, and record the propagation time A required for their cooperative propagation. ij Define the speedup ratio for collaboration as:

[0071]

[0072] Where A i A j The outer diffusion time that triggers the global cascading when nodes i and j are used as seed points alone. A larger value indicates a more significant acceleration effect.

[0073] 3: Two-point competition perturbation modeling

[0074] Let nodes i and j be two competing seed nodes in different states. During propagation, each node is allowed to be occupied by only one propagation cluster, ultimately forming a two-cluster propagation structure. Define the competitiveness index:

[0075]

[0076] Where S ij The size or number of seed points that spread during the competition between seed point i and seed point j.

[0077] Step C:

[0078] In one embodiment of the present invention, a machine learning prediction model is constructed based on the extracted brain region-level propagation dynamics indices to assess the neurodevelopmental outcomes of infants at 18 months of age. The predictor input consists of propagation indices from 90 brain regions in the structural brain network for each subject. Each brain region contains five features, derived from both single-point and two-point perturbation scenarios, including: external propagation time, internal propagation time, propagation capacity, cooperative speedup ratio, and competitive indices. Therefore, a total of 450 features (90 brain regions × 5 indices) are extracted for each subject.

[0079] Subsequently, a random forest regression model was used as the prediction algorithm. The model was trained to predict standardized infant developmental scores, including the three dimensions of the BSID-III scale: cognitive score, language score, and motor score. The model employed a five-fold cross-validation strategy to evaluate its generalization performance, and the prediction results were quantified using the Pearson correlation coefficient between the predictions and the true scores.

[0080] To enhance the interpretability of the results, the feature importance assessment mechanism in the random forest model was further utilized to extract brain regions and propagation features that make key contributions to cognitive, language, and motor predictions. This analysis reveals the potential physiological mechanisms linking propagation patterns with different developmental functions, demonstrating the complementarity of single-point and two-point perturbations in multidimensional neurodevelopment prediction.

[0081] This method not only achieves efficient prediction of infants' future developmental levels, but also enhances the interpretability and biological relevance of the model through feature importance mapping, which is helpful for subsequent personalized clinical early warning and brain region function research.

Claims

1. A method for early warning of intellectual disability in infants and young children based on abnormal propagation dynamics of white matter networks, the method comprising the following steps: Step 1: Obtain raw data of T2-weighted structural images (T2W) and diffusion-weighted images (DWI) of the infant's brain; Brain tissue was extracted and artifacts were removed from T2W images and registered to standard space; DWI images were registered to T2W space. Step 2: Using partition templates and high-throughput fiber tracing algorithms, construct individual-level brain white matter structural connectivity networks to form a unified weighted adjacency matrix, ensuring the consistency and repeatability of region segmentation and fiber tracing. Step 3: Propagation dynamics simulation and index extraction; Based on the weighted adjacency matrix, a threshold propagation model is established, and single-point and two-point perturbation simulations are performed on the whole brain network to calculate the dynamic indicators of each node, including propagation time, propagation capacity, cooperative speedup ratio, and competitiveness. Step 4: Feature extraction and developmental score prediction; Based on the aforementioned single-point and two-point propagation dynamics simulations, multidimensional propagation indicators for each brain region are extracted. These indicators are used as features, and a random forest regression machine learning method is employed to predict the cognitive, language, and motor development scores of infants at 18 months of age. A five-fold cross-validation strategy is used to evaluate the model performance, and statistical indicators such as the correlation between predicted and actual values ​​are output.

2. The method for early warning of intellectual disability in infants and young children based on abnormal propagation dynamics of white matter networks as described in claim 1, characterized in that, The method for constructing the weighted adjacency matrix in step 2 is as follows: Step 2.1: For each infant subject, based on the preprocessed diffusion-weighted imaging data, the high-throughput fiber tracing algorithm was used to automatically reconstruct the white matter fiber bundles in the brain; Step 2.2: The whole brain was segmented into 90 anatomical functional areas using a standard partitioning template suitable for the structural characteristics of the neonatal brain. Based on this, the number of fiber bundles connecting each pair of brain areas by fiber tracing results was counted, and the following normalization strategy was used to construct a 90×90 structural connectivity matrix at the individual level: in, This represents the normalized structural connectivity strength between brain regions i and j in the k-th subject; The number of fiber bundles connecting the two regions. The average length of the connecting fibers. This represents the average surface area of ​​the cortical regions of the two brain areas. Step 2.3: Average the individual structural connectivity matrices of all subjects to construct a group-level structural connectivity network, forming a uniform 90×90 weighted adjacency matrix.

3. The method for early warning of intellectual disability in infants and young children based on abnormal propagation dynamics of white matter networks as described in claim 1, characterized in that, The single-point disturbance model in step 3 is as follows: All nodes are initially set to inactive (represented by 0); the selected seed node is set to active (represented by 1); and the state of each node is updated according to the following dynamic rules during the iteration process: Where: x i (t) represents the state of node i at time t; w ij K represents the connection weight between nodes i and j; i =∑ j w ij θ represents the total input received by node i; θ is the normalized propagation threshold, used to unify and compare the propagation dynamics process. During the propagation simulation, the minimum time step t required for any node j in the network to be activated from seed point i is recorded. ij Construct the propagation time matrix T = [t ij ]; After obtaining the propagation time matrix, the following key metrics are defined: external propagation time out i for: Where N represents the total number of nodes, out i The speed at which node i, as a seed point, influences the propagation throughout the brain. Internal propagation time i for: in i It reflects the sensitivity of node i to external disturbances; Transmission capability SA i for: SA i =max{θ i |f(i,θ i )≥f c } Where, θ i f(i,θ) represents the threshold at which node i can trigger a global cascade. i ) indicates that at the threshold θ i The following is the proportion of activated nodes in the network when node i is used as the seed, f. c This indicates the proportion of nodes that are set to be active. When f c When SA = 1, a global cascading mechanism is triggered, meaning that all nodes in the network must be activated. i This represents the upper limit of the minimum threshold required to trigger a global cascade when node i becomes a seed point. A higher threshold indicates that a global cascade can still be triggered under more difficult conditions.

4. The method for early warning of intellectual disability in infants and young children based on abnormal propagation dynamics of white matter networks as described in claim 1, characterized in that, In step 3, the two-point perturbation is a two-point cooperative perturbation, and the two-point cooperative perturbation model is as follows: For each pair of nodes (i,j), simultaneously set them to active, simulate the two-point propagation process, and record the propagation time A required for their cooperative propagation. ij Define the speedup ratio S for collaboration as: Wherein, min(A) i A j ) indicates that option A is selected. i A j The minimum value in, A i A j The outer diffusion time that triggers global cascading when nodes i and j are used as seed points alone; the larger the value, the more obvious the acceleration effect.

5. The method for early warning of intellectual disability in infants and young children based on abnormal propagation dynamics of white matter networks as described in claim 1, characterized in that, In step 3, the two-point disturbance is a two-point competition disturbance, and the two-point competition disturbance model is as follows: Let nodes i and j be two competing seed nodes in different states. During propagation, each node is allowed to be occupied by only one propagation cluster, ultimately forming a two-cluster propagation structure; define the competitiveness index C. i for: Where S ij The size or number of seed points that spread during the competition between seed point i and seed point j.

Images