Method for evaluating neurodisease fusion image transcriptomics and spatial topology constraints

Abstract

This invention discloses a method and device for assessing neurological diseases by integrating imaging transcriptomics and spatial topological constraints, belonging to the field of neurological disease diagnostic technology. Addressing the shortcomings of traditional methods, such as reliance on a single imaging modality, susceptibility to spatial autocorrelation artifacts, and lack of cross-scale validation, this method constructs a 3036-dimensional structure-function heterogeneous tensor library based on BNA-246, quantifies the decoupling degree of SC-FC nonlinear coupling at multiple scales, and corrects for 10,000 spatial rotation permutations. Through LASSO-GBM modeling with binomial bias constraints, an AUC of 0.814 for ALS diagnosis is achieved. Furthermore, combining PLS and single-cell gene activity scores enables cross-scale closed-loop correction, improving the accuracy and interpretability of early diagnosis of neurological diseases, especially ALS, and demonstrating significant clinical application value.

Description

Technical Field

[0001] This invention belongs to the field of neurological disease diagnosis technology. It relates to a method for assessing neurological diseases by integrating imaging transcriptomics and spatial topological constraints. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Among neurological diseases, amyotrophic lateral sclerosis (ALS) exhibits significant insidious early pathological changes, posing a core challenge to its clinical diagnosis and intervention. Traditional diagnostic and assessment methods largely rely on a framework combining medical imaging examinations with neuropsychological assessments. However, limited by the technical limitations of single-imaging modalities, these methods struggle to accurately capture the dynamic mismatch between brain structural integrity and functional activity, failing to reveal the intrinsic link between early microscopic pathology and macroscopic functional changes.

[0004] At the same time, the inherent spatial autocorrelation characteristics of brain network data are another major statistical bottleneck in traditional analysis methods. This can easily lead to spatial artifacts in statistical inference, interfering with the identification and quantitative analysis of abnormal features, and further reducing the reliability of diagnostic conclusions.

[0005] Current clinical diagnostic strategies for ALS generally rely on single types of brain scan data: for example, analyzing morphological atrophy changes in specific brain regions using structural magnetic resonance imaging (SMRI), or calculating changes in functional connectivity between brain regions using functional magnetic resonance imaging (fMRI). While these methods can identify some overt abnormal features, they are limited by their single modality and one-sided dimension, making it difficult to comprehensively characterize the complex pathological mechanisms of ALS, such as brain network remodeling and structure-function decoupling, and also failing to meet the clinical need for early and accurate diagnosis.

[0006] The cross-scale assessment system based on spatial topological constraints breaks through the above limitations and can realize cross-modal fusion of macroscopic imaging phenotypes and microscopic single-cell omics information. It quantifies the coupling relationship between brain structure and function at multiple scales, including regions, sub-networks, and the whole brain, effectively avoids statistical artifacts caused by spatial autocorrelation, and ultimately significantly improves the accuracy and specificity of early ALS diagnosis.

[0007] Therefore, cross-scale assessment methods based on spatial topological constraints have become a core solution to break through the bottlenecks of traditional diagnostic techniques and solve the problem of identifying early occult pathological features of neurological diseases such as ALS. Through cross-modal fusion and multi-scale quantitative analysis, it can not only improve the accuracy and specificity of early diagnosis, but also provide a new approach for the study of the pathological mechanisms of neurological diseases and their clinical translation. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for evaluating neurological diseases by fusing imaging transcriptomics and spatial topological constraints, comprising the following steps:

[0009] (1) Constructing a high-dimensional heterogeneous tensor feature mapping space: Multimodal image data of the subjects were acquired, including high-resolution T1, rs-fMRI and DTI data; Brainnetome Atlas (BNA-246) brain map was used as the anatomical anchor point; the network sparsity threshold sequence S∈[0.05,0.50] was set with a step size of 0.01; by calculating the area under the cubic spline interpolation evolution curve of the topological metric value under each step threshold, integrated feature indicators with topological robustness were extracted. Then, using the refined anatomical atlas as the three-dimensional spatial anchor point, a 3036-dimensional structure-function heterogeneous feature tensor library was constructed. This feature space covers the physical connectivity integrity of the structural network (SC), the dynamic synchronization stability of the functional network (FC), and the nonlinear deviation index of the cross-modal coupling of structure-function (SC-FC), providing full-dimensional candidate variables for subsequent feature screening based on binomial deviation constraints;

[0010] (2) Perform connection decoupling analysis based on nonlinear dynamic operators: Using a nonlinear mapping algorithm, calculate the coupling weight index between SC and FC at three scales: whole brain, region, and sub-network, to quantify the degree of dynamic disconnection of macroscopic neural connections during phenotypic evolution. The preferred formula is as follows:

[0011] ①Regional Scale (Ci): Adaptive Rank Transform Nonlinear Mapping Operator

[0012] For each node i in the 246 brain regions, its coupling weight exponent is defined as the nonlinear rank correlation integral of the structural topological manifold and the functional communication manifold in the sparsity evolution space:

[0013]

[0014] Parameter description:

[0015] Brain region nodes The coupling weight index. The network sparsity threshold follows a cubic spline interpolation evolution sequence. Upper and lower limits of the sparsity threshold sequence (according to the specification). ). / : Represent the structural connectivity matrix and functional connectivity matrix at a specific sparsity, respectively. The node connection rank vector. / : The average value of the corresponding rank vector. Total number of brain regions (BNA-246 atlas used in the text, therefore...) ).

[0016] Physical meaning: This operator captures the nonlinear constraint effectiveness of the structural skeleton of node i on functional communication by performing Riemann integration in the full threshold space.

[0017] ② Sub-network scale (Δ) net Cross-modal information consistency deviation index:

[0018] For the Yeo-7 functional subnetwork, the preferred one is the somatic motor network (SMN). Eigenvalue decomposition bias quantifies the structure-function disconnect of the local system.

[0019]

[0020] Parameter description:

[0021] : These represent the first principal eigenvalues ​​of the sub-network structure / functional adjacency matrices, respectively, characterizing the strength of the network synchronization core. Kullback-Leibler divergence (relative entropy) measures the difference in importance distribution between two modal nodes. : No. The importance distribution of each node under the structural / functional modes. The number of nodes in this subnetwork.

[0022] Physical meaning: This formula measures the information entropy increment of the topological pattern of functional signal flow deviating from the anatomical physical path within a specific subnetwork, and is used to identify asymmetric decoupling in the early motor cortex of ALS.

[0023] ③ Global scale (G score ): Integrated Coupling Index under Anatomical Manifold Constraints

[0024] The degree of macroscopic disconnect across the entire brain is globally integrated by introducing a geodesic distance-weighted operator:

[0025]

[0026] Parameter description:

[0027] Brain regions Geodesic centrality on cortical manifolds. Spatial autocorrelation correction factor generated based on the Spin Permutation algorithm. Spatial decay constant.

[0028] Physical significance: This index achieves global risk quantification under the geometric constraints of the anatomical manifold, and the Spin correction ensures that the coupled score eliminates the statistical interference of artifacts caused by physical spatial proximity.

[0029] (3) Establish a spatial rotation permutation correction model based on anatomical manifold constraints: project the three-dimensional brain region coordinates onto a two-dimensional closed manifold sphere, perform 10,000 rotation operators that preserve spatial autocorrelation characteristics to generate a statistical null model distribution, and eliminate spatial autocorrelation interference. The specific formula is as follows:

[0030] ①Spherical manifold transformation mapping of brain region coordinates

[0031] First, project the centroid coordinates Vi=(xi,yi,zi) of each brain region in the BNA-246 atlas onto a standard unit spherical manifold:

[0032]

[0033] Parameter description:

[0034] Brain regions The original centroid coordinates . : The geometric center of the hemisphere to which it belongs. : The coordinate vector of the standard unit spherical manifold after projection.

[0035] Physical significance: This step transforms discrete anatomical locations into continuous manifold coordinates, laying the foundation for rotation that preserves the topological structure.

[0036] ② Generation of rotation operators that preserve spatial autocorrelation

[0037] Generate a random rotation matrix R(k) 10,000 times. For the k-th permutation, the new coordinates of each brain region are:

[0038]

[0039] Parameter description:

[0040] : Represents the first After several random rotational permutations, brain regions The new coordinate vector obtained.

[0041] : Represents the first The randomly generated rotation matrix.

[0042] : Represents brain regions The original projected coordinate vector on the standard unit spherical manifold.

[0043] Technical details: By randomly selecting rotation angles, feature values ​​are redistributed without disrupting the geodesic distances between brain regions.

[0044] ③ Significance evaluation operator based on manifold constraints

[0045] Calculate the spatial correlation significance P between image feature vector SC-FC coupling degree X and gene expression vector Y. spin :

[0046]

[0047] Parameter description:

[0048] Number of rotational permutations (set to 10,000 in the text). : Correlation measurement function. : Feature vector of the original observed image. : No. The "fake" image atlas generated after the rotational permutation. Gene expression vector. : Indicator function, returns 1 if the condition in parentheses is true, otherwise returns 0.

[0049] Physical meaning: P is defined by this formula spin This reflects the probability of a true pathological association between imaging phenotypes and transcriptomic features after excluding spatial proximity interference. Only when P is satisfied... spin Only features with a value <0.05 are considered to have cross-scale biological significance.

[0050] (4) Construct an integrated learning evaluation engine based on binomial bias constraint: Use a two-stage recursive feature dimensionality reduction algorithm to lock the prediction factor set, construct a classification model based on gradient boosting machine (GBM), and combine Shapley value (SHAP) to perform nonlinear attribution of feature contribution.

[0051] (5) Perform cross-scale transcriptional imprinting mapping and algorithm closed-loop verification: Use partial least squares regression algorithm to extract gene feature weight vectors, and combine nonlinear gene activity scores at single-cell resolution to dynamically correct macroscopic discrimination logic.

[0052] The beneficial effects of this invention are as follows: It provides a method and device for evaluating neurological diseases by integrating imaging transcriptomics and spatial topological constraints. This addresses the shortcomings of traditional neurological disease diagnosis methods, such as reliance on a single imaging modality, susceptibility to spatial autocorrelation artifacts, lack of cross-scale validation from macro to micro levels, and poor model interpretability. Through the integration of multimodal feature construction, multi-scale decoupling analysis, spatial topological correction, ensemble learning modeling, and cross-scale closed-loop validation, it achieves more precise and clinically applicable early assessment of neurological diseases. In specific applications, the constructed binomial bias-constrained ensemble learning model achieves an AUC of 0.814 in diagnosing ALS. SHAP attribution ensures model interpretability, and the combination of transcriptomics completes cross-scale closed-loop validation of macroscopic imaging and microscopic gene mechanisms, further improving the accuracy of ALS diagnosis by correcting the score. Attached Figure Description

[0053] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0054] Figure 1 This is a flowchart of steps S1 to S5 of the amyotrophic lateral sclerosis (ALS) assessment method that integrates imaging transcriptomics and spatial topological constraints as described in Example 1.

[0055] Figure 2 This is a schematic diagram of the amyotrophic lateral sclerosis (ALS) assessment device that integrates imaging transcriptomics and spatial topological constraints as described in Example 2. Detailed Implementation

[0056] To enable those skilled in the art to better understand this application, this application will now be further described in conjunction with specific embodiments.

[0057] Example 1

[0058] A Method for Assessing Amyotrophic Lateral Sclerosis by Integrating Imaging Transcriptomics and Spatial Topological Constraints

[0059] This embodiment discloses a method for assessing neurological diseases that integrates imaging transcriptomics and spatial topological constraints, using amyotrophic lateral sclerosis (ALS) as a specific application. The implementation process, experimental conditions, operational details, and validation results of this method are described in detail. Technical parameters not explicitly defined in this embodiment are all conventional parameters in the fields of neuroimaging and bioinformatics analysis; all experimental operations comply with medical ethics review requirements, and all research subjects signed informed consent forms.

[0060] I. Experimental Materials

[0061] 1. Study subjects: 73 clinically diagnosed ALS patients were selected as the case group, and 74 healthy volunteers matched for gender, age and education level were selected as the healthy control group (HC). Study subjects with other neurological diseases, organic brain lesions and contraindications to magnetic resonance imaging were excluded.

[0062] 2. Experimental equipment: A 3.0T superconducting magnetic resonance imaging system was used, equipped with a standard head coil, to acquire multimodal brain imaging data; a bioinformatics analysis workstation was used, configured with an Intel(R) Xeon(R) Gold6248R CPU, 128GB DDR4 memory, NVIDIA A100 graphics card, and running Python 3.9, Matlab 2023b, SPM12, and FSL 6.0 analysis environment.

[0063] 3. Data and Analysis Tools: The Allen Human Brain Atlas (AHBA) publicly available transcriptome dataset yielded 15,633 high-quality candidate genes; the JuSpace toolbox was used for geodesic distance calculation and spatial association analysis; FSL and PANDA tools were used for constructing brain network connectivity matrices; scikit-learn and XGBoost toolkits were used for machine learning modeling; PLS, ssGSEA, and AUCell analysis tools were used for cross-scale association in transcriptomics; and the SHAP feature attribution module was used for model interpretability analysis.

[0064] II. Assessment Steps

[0065] S1 Constructing a High-Dimensional Heterogeneous Tensor Feature Mapping Space: Through the acquisition, preprocessing, and feature extraction of multimodal image data, a full-dimensional structure-function heterogeneous feature tensor library is constructed, specifically including the following sub-steps:

[0066] 1. Multimodal imaging data acquisition: Brain scans were performed on all study subjects, and three types of core imaging data were acquired: ① High-resolution T1-weighted images, using 3D-T1WI sequence with parameters set as TR=2500ms, TE=3.5ms, slice thickness=1mm, and no slice spacing; ② Resting-state functional magnetic resonance imaging (rs-fMRI), using EPI sequence with parameters set as TR=2000ms, TE=30ms, slice thickness=3mm, number of slices=43, and scan duration=8min; ③ Diffusion tensor imaging (DTI), using SE-EPI sequence with parameters set as TR=6000ms, TE=80ms, b-value=1000s / mm², and 30 diffusion directions.

[0067] 2. Image data preprocessing: ① Head motion correction, temporal correction, spatial normalization, 6mm Gaussian kernel smoothing, white matter removal, cerebrospinal fluid removal, head motion parameter-related noise removal, and low-frequency drift removal were performed on rs-fMRI data; ② Head motion correction and eddy current correction were performed on DTI data, a diffusion tensor model was fitted, and anisotropy fractions were calculated; ③ Skull dissection and brain region segmentation were performed on high-resolution T1-weighted images, and spatial registration was performed with rs-fMRI and DTI data to unify them to the MNI152 standard space.

[0068] 3. Brain network connectivity matrix construction: Using the BNA-246 brain atlas as a refined anatomical anchor, the whole brain was divided into 246 brain regions, with the centroid of each brain region as a node; ① Based on DTI data, a deterministic fiber tract tracking algorithm was used to construct a structural connectivity (SC) matrix, and the structural connectivity strength was quantified by the number of fiber tracts between brain regions; ② Based on rs-fMRI data, the Pearson correlation coefficient of blood oxygenation level-dependent signals between brain regions was calculated, and an FC matrix was constructed. After Fisher-Z transformation, the standardized functional connectivity strength was obtained.

[0069] 4. Topology Feature Extraction: Set the network sparsity threshold sequence S∈[0.05,0.50] with a step size of 0.01. Calculate the node degree, clustering coefficient, and feature path length, among other topology metrics, for the SC and FC matrices at each sparsity level. Fit the evolution curve of the topology metrics with sparsity using cubic spline interpolation, calculate the area under the curve (AUC), and extract the integrated feature index with topological robustness.

[0070] 5. Construction of Heterogeneous Feature Tensor Library: Integrating the physical connectivity integrity index of the SC matrix, the dynamic synchronization stability index of the FC matrix, and the nonlinear deviation index of SC-FC cross-modal coupling, the tensor dimensions are divided according to brain region, scale, and index type. Finally, a 3036-dimensional structure-function heterogeneous feature tensor library is constructed to provide full-dimensional candidate variables for subsequent feature selection.

[0071] S2 performs connection decoupling analysis based on nonlinear dynamic operators.

[0072] This step uses a multi-scale nonlinear mapping algorithm to quantify the SC-FC coupling weights and dynamic decoupling degree, and identify ALS-specific network decoupling patterns. Specifically, it includes the following sub-steps:

[0073] 1. Preparation for multi-scale analysis: Based on the Yeo-7 functional brain network classification criteria, 246 brain regions were classified into functional subnetworks of the somatic motor network (SMN), default mode network (DMN), dorsal attention network (DAN), and ventral attention network (VAN). The geodesic distance of each brain region on the cortical manifold was calculated using the JuSpace toolbox, providing a basis for whole-brain scale weighted analysis.

[0074] 2. Calculation of region-scale coupling weights: For each node in 246 brain regions, an adaptive rank transform nonlinear mapping operator is introduced to calculate the nonlinear rank correlation integral of the structural topological manifold and the functional communication manifold in the sparsity evolution space, and obtain the SC-FC coupling weight index of each brain region; the Spearman rank correlation coefficient is used to characterize the nonlinear coupling strength between the physical integrity of white matter fiber bundles and the resting-state blood oxygenation level-dependent signal synchronicity between brain regions.

[0075] 3. Sub-network scale decoupling index calculation: For functional sub-networks, SC and FC sub-matrices are extracted from the sub-networks respectively, and the first principal eigenvalue of the sub-network adjacency matrix is ​​obtained through eigenvalue decomposition; the Kullback-Leibler divergence (KL divergence) of node importance distribution under both structural and functional modes is calculated, and the cross-modal information consistency deviation index is obtained by combining the principal eigenvalue deviation, which quantifies the degree of SC-FC decoupling within the sub-network.

[0076] 4. Whole-brain scale integrated coupling index calculation: A geodesic distance weighted operator is introduced, using the geodesic centrality of each brain region as the weight, to globally integrate the regional scale coupling weight index; combined with the spatial autocorrelation correction factor P. spin We constructed a whole-brain integrated coupling index under anatomical manifold constraints to quantify the degree of dynamic disconnection of macroscopic neural connections during phenotypic evolution.

[0077] 5. Specific decoupling pattern identification: Compare coupling / decoupling indices at various scales between ALS patients and healthy controls, screen for features with significant inter-group differences, and focus on identifying specific decoupling patterns in which the internal structural integrity of the SMN is preserved but functional connectivity is weakened, as well as abnormal decoupling features between networks.

[0078] S3 establishes a spatial rotational permutation correction model based on anatomical manifold constraints.

[0079] This step constructs a statistical null model through spherical manifold projection and multiple rotational permutations to eliminate statistical artifacts caused by spatial autocorrelation and verify the significant pathological correlation of image features. Specifically, it includes the following sub-steps:

[0080] 1. Brain region coordinate extraction: Extract the three-dimensional centroid coordinates (V) of 246 brain regions from the BNA-246 brain atlas. i =(x i ,y i ,z i Differentiate between the left and right hemispheres of the brain and calculate the geometric center C of each hemisphere. hem .

[0081] 2. Spherical manifold projection: The centroid coordinates of all brain regions are projected onto a standard unit spherical manifold $S^2$ to obtain the projected spherical coordinate vectors, realizing the transformation from discrete anatomical locations to continuous manifold coordinates, while preserving the geodesic distances and topological structure between brain regions.

[0082] 3. Spatial rotation operator generation: Using the Spherical Permutation operator, generate 10,000 random rotation matrices R(k) that satisfy orthogonality constraints (R... T R=I and det(R)=1); by randomly selecting rotation angles θ, Ф, and ψ, the spherical coordinates are rotated and transformed without destroying the geodesic distance between brain regions, to obtain the new coordinates of the brain region after the kth permutation.

[0083] 4. Zero-model distribution construction: Based on the new coordinates after each rotation and permutation, the SC-FC coupling / decoupling index is recalculated to generate 10,000 sets of false image atlases without real pathological associations, and a statistical zero-model distribution conforming to the spatial Laplacian eigenvalue is constructed.

[0084] 5. Spatial Significance Verification: The spatial similarity between the original observed image decoupling map and the neurotransmitter receptor mapping map is calculated and compared with the null model distribution to obtain the spatial correlation significance P. spin When P spin When the value is less than 0.05, the image feature is considered to have significant pathological relevance, excluding artifact interference caused by spatial proximity.

[0085] S4 constructs an ensemble learning evaluation engine based on binomial bias constraints.

[0086] This step involves two-stage feature dimensionality reduction to select the optimal predictor variables, constructing a GBM classification model, and performing interpretability analysis of the model. Specifically, it includes the following sub-steps:

[0087] 1. Feature preprocessing: Z-score normalization was performed on the 3036-dimensional features that were validated as significant in S3 to remove missing values ​​and outliers; the dataset of the research subjects was divided into a training set (51 ALS patients and 52 HC patients) and a test set (22 ALS patients and 22 HC patients) in a 7:3 ratio.

[0088] 2. Two-stage recursive feature dimensionality reduction: ① First stage: Use independent samples t-test to perform inter-group difference analysis on the features of the training set, set the significance level P<0.05, and filter out indifferent features with low information gain; ② Second stage: Introduce the LASSO penalty term operator based on the binomial bias minimization constraint, take the binomial bias of the classification model as the objective function, optimize the penalty coefficient through cross-validation, realize the sparsity selection of features, and adaptively lock 29 optimal predictor variables, including the left anterior central gyrus (PrG) coupling degree, from the heterogeneous feature library.

[0089] 3. GBM Model Training and Validation: Based on 29 optimal predictor variables, a GBM classification model was constructed in the training set, with 1000 decision trees, a learning rate of 0.01, and a maximum depth of 5. Five-fold cross-validation was used to optimize the model, and the model performance was validated in the test set. Evaluation metrics such as Area Under the Receiver Operating Characteristic (AUC), accuracy, sensitivity, and specificity were calculated.

[0090] 4. SHAP Feature Attribution Analysis: The SHAP algorithm is introduced to calculate the contribution and direction of each optimal predictor variable to the classification results of the GBM model, generate SHAP summary plot and dependency plot, analyze the role mechanism of left anterior central gyrus coupling in ALS diagnosis, and realize the nonlinear attribution and interpretability analysis of the model.

[0091] S5 performs cross-scale transcriptional imprinting mapping and algorithmic closed-loop verification.

[0092] This step integrates macroscopic image features with microscopic transcriptome data to achieve cross-scale correlation analysis and dynamically corrects macroscopic risk scores, completing the algorithm loop. Specifically, it includes the following sub-steps:

[0093] 1. Transcriptome data preprocessing: Human cortical gene expression data were extracted from AHBA, and low-expression genes, pseudogenes, and mitochondrial genes were removed to obtain 15,633 high-quality candidate gene expression profiles. The gene expression data were standardized and spatially registered with brain atlases to match the BNA-246 brain region division.

[0094] 2. PLS1 Feature Mapping: Using the PLS regression algorithm, a correlation model is constructed between macroscopic image features, namely the 29 optimal predictor variables in S4 and microscopic gene expression features, and the PLS1 weight vector operator is extracted. Through this operator, the macroscopic image decoupled features are mapped to the expression space of 15633 candidate genes, and the gene set that is significantly related to the image features is selected.

[0095] 3. ssGSEA gene set enrichment analysis: The single-sample gene set enrichment analysis algorithm was used to perform enrichment analysis on the mapped gene set. The enrichment threshold FDR < 0.05 was set to identify significantly enriched genomes in the cortical Layer III and Layer V levels and analyze their corresponding biological functional pathways.

[0096] 4. AUCell Single-Cell Activity Score Quantification: Introducing the AUCell ranking probability density function, based on single-cell transcriptome data, the immune pathway activity score of the FMN1 key gene in microglia is quantified; combined with the expression level of enriched genomes, a disease risk score at the micromolecular level is obtained.

[0097] 5. Macroscopic scoring nonlinear correction: Based on the single-cell immune pathway activity score, a nonlinear feedback weight is generated to dynamically correct the macroscopic image risk score output by the GBM model in S4; the diagnostic performance of the corrected model is verified, the cross-scale consistency verification of macroscopic image phenotype and microscopic gene mechanism is realized, and the closed loop of the entire evaluation algorithm is completed.

[0098] III. Experimental Results

[0099] 1. SC-FC decoupling feature identification results: S2 and S3 analyses showed that, compared to healthy controls, ALS patients had significantly higher levels of somatosensory network (SMN, β=-0.023, P0.023) characteristics. FDR <0.001) and the default mode network (DMN, β=-0.015, P FDR The SC-FC coupling was significantly reduced within the range of 0.03, and there was extensive inter-network decoupling. The decoupling features between the SMN and the dorsal attention network (DAN, P=0.010) and the ventral attention network (VAN, P=0.005) were the most significant. The changes in SC-FC coupling in ALS patients were negatively correlated with serotonin receptor expression (5-HT2A: r=-0.35, p=0.007; 5-HT1B: r=-0.32, p=0.021).

[0100] 2. GBM model diagnostic results: The GBM classification model constructed by S4 achieved an area under the discriminant curve (AUC) of 0.814 for early ALS patients in the test set, with an accuracy of 0.786, a sensitivity of 0.773, and a specificity of 0.800. SHAP attribution analysis showed that the coupling degree of the left anterior central gyrus and the decoupling index of the SMN subnetwork were the core predictors for ALS diagnosis.

[0101] 3. Cross-scale correlation and closed-loop verification results: S5 analysis revealed 985 genes significantly associated with the SC-FC coupling pattern. Among them, PLS1 negatively correlated genes were significantly enriched in synaptic processes (FDR<0.001) and excitatory neurons in cortical layers III / V (NES=-2.34, p<0.05). The macroscopic risk score, corrected by single-cell gene activity score, showed a significantly higher correlation with clinical disease severity (r=0.62, p<0.001) than the original score (r=0.45, p<0.001), verifying the effectiveness of the algorithm's closed-loop mechanism.

[0102] Example 2

[0103] A neurological disease assessment device integrating imaging transcriptomics and spatial topological constraints

[0104] This embodiment discloses a neurological disease assessment device for implementing the method described in Embodiment 1. The device is implemented through a combination of hardware and software, including interconnected image tensor construction, dynamic decoupling analysis, spatial topology correction, integrated gradient assessment, and cross-scale closed-loop feedback modules. Each module can be integrated into a magnetic resonance imaging (MRI) analysis device or a standalone bioinformatics analysis workstation, and is suitable for the early assessment and diagnosis of neurological diseases such as ALS. The specific functional configurations of each module are as follows:

[0105] 1. Image Tensor Construction Module: Configured to receive multimodal MRI data acquired by a 3.0T magnetic resonance imaging system, it incorporates SPM12 and FSL preprocessing algorithms to perform head motion correction, spatial registration, denoising, and other operations. Using the BNA-246 brain atlas as anatomical anchor points, it calls the PANDA tool to construct SC and FC matrices, calculates the topological metric AUC through a cubic spline interpolation evolution algorithm, and finally establishes a 3036-dimensional spatiotemporal feature tensor library covering connectivity topological attributes, and realizes feature standardization and storage.

[0106] 2. Dynamic Decoupling Analysis Module: Configured to incorporate cross-modal information consistency operators and Spearman rank correlation algorithm to calculate SC-FC cross-modal nonlinear coupling weights at three scales: whole brain, region, and sub-network. It has a built-in Yeo-7 functional sub-network partitioning model to automatically extract decoupling indices of core sub-networks such as SMN and DMN, quantifying the degree of functional imbalance between neural output pathways and cortical circuits. It focuses on identifying and marking specific decoupling patterns within the SMN that are "structurally intact but functionally weakened," and outputs multi-scale decoupling feature maps.

[0107] 3. Spatial Topology Correction Module: Configured to call the JuSpace toolbox to extract brain region centroid coordinates, calculate geodesic distances, and project spherical manifolds; it incorporates a Spherical Permutation rotation operator generation algorithm, which can automatically complete 10,000 spatial rotation permutations and construct a statistical null model distribution; it calculates P through spatial correlation analysis. spin The system automatically determines the pathological relevance of image features, removes spatial autocorrelation artifacts, and outputs a corrected set of significant decoupled features.

[0108] 4. Integrated gradient evaluation module: Configured to achieve two-stage recursive feature dimensionality reduction, with built-in t-test and binomial bias constraint LASSO algorithm to automatically select the optimal predictive variable; integrates GBM classification model training and five-fold cross-validation functions, supports automatic training / test set splitting and model performance evaluation, and outputs macroscopic image risk scores and diagnostic results; built-in SHAP feature attribution module, which can generate feature contribution visualization maps to realize nonlinear attribution and interpretability analysis of the model.

[0109] 5. Cross-scale closed-loop feedback module: Configured to interface with public transcriptome datasets such as AHBA, enabling automatic loading and preprocessing of expression profiles for 15,633 candidate genes; incorporates PLS1, ssGSEA, and AUCell algorithms to complete the mapping of macroscopic image features to gene space, cortical-level gene set enrichment analysis, and quantification of single-cell immune pathway activity scores; can generate nonlinear feedback weights based on single-cell scores to dynamically correct macroscopic risk scores, complete cross-scale closed-loop verification, and output the final disease assessment report and related molecular mechanism analysis results.

[0110] Each of the above modules can be implemented by computer program code and stored in a computer-readable storage medium. When the program code is executed by the processor, the neurological disease assessment method described in this invention is completed. At the same time, each module can be equipped with a visual interactive interface to support operators in setting parameters, viewing results, and exporting data, thereby improving the convenience of clinical application.

Claims

1. A method for assessing neurological diseases by integrating imaging transcriptomics and spatial topological constraints, characterized in that, Includes the following steps: S1, Constructing a high-dimensional heterogeneous tensor feature mapping space: Acquire multimodal image data of the subjects, use refined anatomical atlases as three-dimensional spatial anchors, and construct a structure-function heterogeneous feature tensor library; S2, perform connection decoupling analysis based on nonlinear dynamic operators: using a nonlinear mapping algorithm, calculate the coupling weight index of structural and functional connections at three scales: whole brain, region, and sub-network, to quantify the degree of dynamic decoupling of macroscopic neural connections during phenotypic evolution; S3, Establish a spatial rotational permutation correction model based on anatomical manifold constraints: Project the three-dimensional brain region coordinates onto a two-dimensional closed manifold sphere, execute 10,000 rotation operators that preserve spatial autocorrelation characteristics to generate a statistical null model distribution, and eliminate spatial autocorrelation interference; S4. Construct an integrated learning evaluation engine based on binomial bias constraints: Use a two-stage recursive feature dimensionality reduction algorithm to lock the predictor set, construct a classification model based on gradient boosting machine, and combine Shapley value to perform nonlinear attribution of feature contribution. S5, Perform cross-scale transcriptional imprinting mapping and algorithm closed-loop verification: Use partial least squares regression algorithm to extract gene feature weight vectors, and combine nonlinear gene activity scores at single-cell resolution to dynamically correct macroscopic discrimination logic.

2. The method for evaluating the fusion of imaging transcriptomics and spatial topological constraints in neurological diseases as described in claim 1, characterized in that: In S1, the imaging data includes high-resolution T1, rs-fMRI, and DTI data; the Brainnetome Atlas is used as the anatomical anchor point; the network sparsity threshold sequence S∈[0.05,0.50] is set with a step size of 0.01; and the integrated feature index with topological robustness is extracted by calculating the area under the cubic spline interpolation evolution curve of the topological metric value under each step threshold.

3. The method for evaluating the fusion of imaging transcriptomics and spatial topological constraints in neurological diseases as described in claim 1, characterized in that: In S2, a cross-modal information consistency operator is introduced; the Spearman rank correlation coefficient is used to characterize the nonlinear coupling strength between the physical integrity of white matter fiber bundles and the resting-state blood oxygen level-dependent signal synchronicity, and to identify significant decoupling patterns within the subject's somatic motor network.

4. The method for evaluating the fusion of imaging transcriptomics and spatial topological constraints in neurological diseases as described in claim 1, characterized in that: In S3, the JuSpace toolbox is used to perform spatial correlation calculations under geodesic distance constraints; the cortical anatomical structure is rotated to a random angle by the Spherical Permutation operator to generate a zero-distribution mapping that conforms to the spatial Laplacian eigenvalue. When the spatial similarity between the image decoupling map and the neurotransmitter receptor mapping map is <0.05, the image feature is determined to have significant pathological correlation.

5. The method for evaluating the fusion of imaging transcriptomics and spatial topological constraints in neurological diseases as described in claim 1, characterized in that: In S4, the first stage uses a t-test to filter features with low information gain; The second stage introduces a LASSO penalty term operator based on binomial bias minimization constraint to adaptively lock 29 optimal predictor variables, including the coupling degree of the left anterior central gyrus, from a heterogeneous feature library; the Gradient Boosting Machine algorithm is used to perform classification prediction for neurological diseases.

6. The method for evaluating the fusion of imaging transcriptomics and spatial topological constraints in neurological diseases as described in claim 1, characterized in that: In S5, the macroscopic features are mapped to the expression space of 15,633 candidate genes using the PLS1 weight vector operator; the ssGSEA algorithm is used to identify significantly enriched genomes at the Layer III and Layer V cortical levels; the AUCell ranking probability density function is introduced to quantify the immune pathway activity score of the FMN1 key gene in microglia at the single-cell level, and feedback weights are generated to nonlinearly correct the macroscopic risk score.

7. A disease assessment device integrating imaging transcriptomics and spatial topological constraints, characterized in that: It includes: Image Tensor Construction Module: Configured to acquire multimodal MRI data, it uses a cubic spline interpolation evolution algorithm to build a 3036-dimensional spatiotemporal feature tensor library covering connectivity and topological properties; Dynamic decoupling analysis module: configured to calculate the cross-modal nonlinear coupling weights of SC-FC and extract decoupling indices that reflect the functional imbalance between neural output pathways and cortical circuits; Spatial topology correction module: configured to execute a spatial rotation permutation algorithm based on geodesic distance constraints to generate a zero-model verification space on a two-dimensional closed manifold sphere; Integrated gradient evaluation module: configured to perform Gradient Boosting Machine classification modeling based on binomial bias optimization and feature contribution analysis based on SHAP attribution theory; Cross-scale closed-loop feedback module: configured to perform macro-micro consistency weighted feedback correction by using PLS weight distribution and single-cell pathway activity score.

Images