Parkinson disease recognition system based on brain network control energy features, medium and product

Abstract

This invention belongs to the fields of medical image processing, neural engineering, and artificial intelligence-assisted diagnosis. It provides a Parkinson's disease identification system, medium, and product based on brain network control energy characteristics. Based on the time series of brain region functional activities, it extracts brain states at rest and calculates the temporal dynamic characteristics of each brain state. Using the whole-brain structural connectivity matrix as the network topology, it constructs a linear network control model to calculate the optimal control energy characteristics for brain state transitions and the energy regulation capacity characteristics of whole-brain regions. Based on brain dynamics feature sets, brain network energy feature sets, and brain network graph theory feature sets, it extracts a feature set significantly related to Parkinson's disease and constructs a multi-dimensional fusion feature matrix. Using this multi-dimensional fusion feature matrix as input, it trains the Parkinson's disease identification model. Based on the multimodal brain imaging data of the subject and the trained Parkinson's disease identification model, it outputs the Parkinson's disease identification result. This achieves non-invasive, accurate, and early identification of Parkinson's disease.

Description

Technical Field

[0001] This invention belongs to the fields of medical image processing, neural engineering and artificial intelligence-assisted diagnosis technology, and in particular relates to a Parkinson's disease identification system, medium and product based on brain network control energy characteristics. 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] Currently, the clinical diagnosis of Parkinson's disease mainly relies on doctors' assessment of motor symptoms. However, by the time patients develop typical motor symptoms, the disease has already progressed to the middle or late stages, missing the optimal window for early intervention. Among existing imaging-assisted diagnostic methods, traditional structural MRI can only observe morphological changes such as brain atrophy in patients with middle or late stages, and its sensitivity for identifying early Parkinson's disease is extremely low. Positron emission tomography (PET) imaging can detect the function of dopaminergic neurons, but it is radioactive, expensive, and poorly accessible, making it unsuitable as a routine screening method. Studies of resting-state functional MRI-related brain networks mostly focus on inter-group differences in static functional connectivity patterns, failing to characterize the dynamic temporal characteristics of brain activity in Parkinson's disease, nor revealing the energy regulation mechanisms behind abnormal brain networks, and lacking objectively quantifiable biomarkers for early identification.

[0004] Network Control Theory (NCT) is a core framework connecting brain structural networks and functional dynamics. It quantifies the optimal control energy required for brain state transitions and the energy regulation capacity of brain regions, revealing a core link between brain energy inefficiency and disease pathology in mental illnesses such as depression and schizophrenia. Parkinson's disease is characterized by significant mitochondrial dysfunction, abnormal brain energy metabolism, and disordered brain network dynamics. However, current technologies have not incorporated the brain energy characteristics of NCT into the identification of Parkinson's disease, nor have they developed a non-invasive and accurate identification scheme based on brain network energy characteristics. Furthermore, most existing Parkinson's disease identification models are based on single-modality features, resulting in poor interpretability, inability to simultaneously achieve disease identification and severity assessment, and failure to provide target references for subsequent neuromodulation therapies. Summary of the Invention

[0005] To address at least one of the technical problems mentioned above, this invention provides a Parkinson's disease identification system, medium, and product based on brain network control energy characteristics, which enables non-invasive, accurate, and early identification of Parkinson's disease.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A first aspect of the present invention provides a Parkinson's disease identification system based on brain network control energy characteristics, comprising: The multimodal brain imaging data preprocessing module is used to preprocess the acquired multimodal brain magnetic resonance imaging data to construct the whole brain structural connectivity matrix and the brain region functional activity time series, respectively. The brain dynamics feature calculation module is used to extract the brain state at rest based on the time series of brain region functional activities, calculate the time dynamics features of each brain state, and jointly constitute the brain dynamics feature set. The brain energy feature extraction module is used to construct a linear network control model based on the whole brain structure connection matrix as the network topology, calculate the optimal control energy features of brain state transitions and the energy regulation ability features of whole brain regions, and form a brain network energy feature set. The multi-dimensional feature fusion module is used to extract core feature sets that are significantly related to Parkinson's disease based on brain dynamics feature sets, brain network energy feature sets, and brain network graph theory feature sets, and to construct a multi-dimensional fusion feature matrix; the brain network graph theory feature set includes graph theory indicators calculated based on structural connectivity matrices and functional connectivity matrices; The Parkinson's disease identification model construction module is used to train the constructed Parkinson's disease identification model by taking a multi-dimensional fusion feature matrix as input and the subject's Parkinson's disease clinical diagnosis label and disease severity scale score as output. The Parkinson's disease identification result output module is used to output Parkinson's disease identification results based on the multimodal brain imaging data of the subject to be tested and the trained Parkinson's disease identification model.

[0007] Furthermore, the acquired multimodal brain magnetic resonance imaging data included T1-weighted structural MRI data, diffusion tensor imaging (DTI) data, and resting-state functional magnetic resonance imaging (rs-fMRI) data.

[0008] Furthermore, in the brain dynamics feature calculation module, based on the time series of brain region functional activities, the resting state of the brain is extracted, including: Based on the Brainnetome brain atlas, brain regions were divided into classic resting-state brain networks. The BOLD signal of brain regions at each time point was Z-score normalized, and the average activity level of each brain network was calculated. The brain network with the highest activity level at each time point was defined as the dominant brain network at that time point. Time points where no network activity exceeded the mean were removed. The brain region activity patterns of the same dominant network at all time points were averaged to obtain the core brain state of the corresponding number of brain networks.

[0009] Furthermore, in the brain dynamics feature calculation module, the temporal dynamics features of each brain state specifically include: Fractional occupancy (FO) is the proportion of time a single brain state occurs within the total scan time. Dwell time (DT) is the average duration of a single brain state that is maintained continuously. Occurrence rate (AR) is the number of times a single brain state occurs per minute; state transition probability (TP) is the probability of switching from one brain state to another.

[0010] Furthermore, in the brain energy feature extraction module, the calculation process of the optimal control energy feature for brain state transitions includes: Based on optimal control theory, design starting from the initial brain state x 0 to target brain state The optimal control trajectory, the energy consumption of equilibrium state transition, and the target state approximation error are used to construct the objective function. The constructed brain network dynamics equations are used as constraints to solve the objective function. The optimal control signal is obtained by simultaneously minimizing the state approximation error and control energy consumption. The conversion energy between different brain states and the sustaining energy of maintaining a single brain state are calculated by combining the optimal control signal. The average conversion energy of the whole brain is obtained by averaging the conversion energy of all brain state pairs. The whole brain control stability is calculated based on the sustaining energy of each brain state.

[0011] Furthermore, in the brain energy feature extraction module, the extraction process of energy regulation capacity features of the whole brain region includes: The baseline whole-brain average conversion energy is obtained by using uniformly controlled input weights throughout the brain. For a single brain region Increase its control weight by 1, while keeping the weights of other brain regions unchanged, and calculate the corresponding average conversion energy of the whole brain. , combined and Get brain regions The regional energy regulation ability is obtained by traversing all brain regions to obtain the rERC value of each brain region, thus forming regional energy regulation characteristics.

[0012] Furthermore, in the multi-dimensional feature fusion module, a core feature set significantly related to Parkinson's disease is extracted through a feature selection algorithm. The feature selection algorithm is either LASSO regression algorithm with L1 regularization or partial least squares (PLS) algorithm. During the selection process, age, gender, and years of education are used as covariates. Spatial autocorrelation and multiple comparison errors are corrected through permutation tests and spin tests. Features with p_FDR < a set value are retained as the core feature set, where p_FDR represents the p value obtained after error detection rate correction.

[0013] Furthermore, the output of the Parkinson's disease identification model also includes the probability of disease, disease severity classification, and abnormal brain region feature reports. The abnormal brain region feature reports include the location of brain regions with significant rERC abnormalities in Parkinson's disease patients compared to healthy controls, the statistical value of the degree of abnormality, and the brain network to which the abnormal brain regions belong.

[0014] A second aspect of the present invention provides a computer-readable storage medium.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the identification method steps corresponding to the Parkinson's disease identification system based on brain network control energy characteristics as described above.

[0016] A third aspect of the present invention provides a program product.

[0017] A program product, which is a computer program product, includes a computer program that, when executed by a processor, implements the identification method steps corresponding to the Parkinson's disease identification system based on brain network control energy characteristics as described above.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention is the first to introduce network control theory into the identification of Parkinson's disease, breaking through the limitations of traditional brain network analysis that only focuses on static connection patterns. It uncovers Parkinson's disease-specific biomarkers from the perspective of energy regulation mechanisms in brain state dynamics. The core pathology of Parkinson's disease includes mitochondrial dysfunction and abnormal brain energy metabolism. The control energy and regional energy regulation capabilities extracted in this invention directly align with the pathophysiological mechanisms of the disease, exhibiting stronger disease specificity and biological interpretability compared to traditional features.

[0019] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] 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.

[0021] Figure 1 This is a flowchart of a Parkinson's disease identification system based on brain network control energy characteristics provided in an embodiment of the present invention; Figure 2 This is a flowchart of a Parkinson's disease identification method based on brain network control energy characteristics provided in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] See Figure 1 This embodiment provides a Parkinson's disease identification system based on brain network control energy characteristics, including: The multimodal brain imaging data preprocessing module is used to preprocess the acquired multimodal brain magnetic resonance imaging data to construct the whole brain structural connectivity matrix and the brain region functional activity time series, respectively. In this embodiment, the acquired multimodal brain magnetic resonance imaging data includes T1-weighted structural MRI data, diffusion tensor imaging (DTI) data, and resting-state functional magnetic resonance imaging (rs-fMRI) data. Preprocessing of multimodal brain magnetic resonance imaging data specifically includes: T1-weighted structural MRI data were subjected to skull dissection, tissue segmentation, spatial standardization, and brain region segmentation. The Brainnetome brain atlas was used to divide the whole brain into a set number of brain regions, including cortical brain regions and subcortical brain regions. The specific number was set according to the actual situation, such as 246 brain regions, including 210 cortical brain regions and 36 subcortical brain regions. Eddy current correction, head motion correction, and diffusion tensor fitting were performed on the DTI data. The whole brain structural connectivity matrix was constructed by probabilistic fiber tract tracking, and the matrix elements were the average anisotropy fractions of white matter fiber tracts in brain regions. The rs-fMRI data were subjected to the following steps: first 10 time points were discarded, slice time correction, head motion correction, spatial normalization, smoothing, delinear drift correction, bandpass filtering, and physiological noise regression. The average BOLD signal time series of each brain region was extracted to form a functional time series matrix.

[0026] The brain dynamics feature calculation module is used to extract the brain state at rest based on the time series of brain region functional activities, calculate the time dynamics features of each brain state, and jointly constitute the brain dynamics feature set. In this embodiment, the brain dynamics feature calculation module extracts the brain state at rest based on the time series of brain region functional activities, including: Based on the Brainnetome brain atlas, the brain regions are divided into classic resting-state brain networks. The BOLD signal of the brain region at each time point is Z-score normalized, and the average activity level of each brain network is calculated. The brain network with the highest activity level at each time point is defined as the dominant brain network at that time point. Time points where no network activity exceeds the mean are removed. The brain region activity patterns of the same dominant network at all time points are averaged to obtain the core brain state of the corresponding number of brain networks. As a further implementation method, based on the classic Yeo 7 brain network division, seven corresponding resting-state brain states are extracted, and the temporal distribution and transition characteristics of brain states are quantified to characterize the dynamic disorder pattern of brain activity in Parkinson's disease. Specifically, based on the Brainnetome brain atlas, 210 cortical brain regions are divided into seven classic resting-state brain networks, including the visual network (VIS), sensorimotor network (SOM), dorsal attention network (DAT), ventral attention network (VAT), limbic network (LIM), frontoparietal network (FPN), and default mode network (DMN). In this embodiment, the brain dynamics feature calculation module specifically includes calculating the time dynamics features of each brain state, including: Fractional Occupancy (FO) is the proportion of time a single brain state occurs within the total scan time. Dwell Time (DT) is the average duration of a single brain state that is maintained continuously. Appearance Rate (AR) is the number of times a single brain state occurs per minute; State Transition Probability (TP) is the probability of switching from one brain state to another. The above-mentioned time-dynamic characteristics of all brain states are averaged to obtain the average dwell time and average occurrence rate of the whole brain, which together constitute the brain dynamic characteristic set.

[0027] The brain energy feature extraction module is used to construct a linear network control model based on the whole brain structure connection matrix as the network topology, calculate the optimal control energy features of brain state transitions and the energy regulation ability features of whole brain regions, and form a brain network energy feature set. In the brain energy feature extraction module, a linear network control model is constructed based on the whole-brain structural connectivity matrix as the network topology, specifically including: Using the whole-brain structural connectivity matrix A as the system state matrix, and the brain region functional activity state vectors... To determine the system state, construct the brain network dynamics equations: , in, ,in, N This represents the total number of brain regions; in this embodiment, it is taken as 246. express The derivative of represents t The rate of change in brain activity over time Represents the whole-brain structural connectivity matrix, as follows: N × N The matrix, elements Indicates from brain region j to brain region i The structural connectivity strength is represented by B, which is the control input weight matrix, describing how external inputs drive various brain regions, and the sensitivity or access pattern of each brain region to external signals. N × M The matrix, where M It is the number of external input sources. For time-dependent control input signals; indicating at time... t External driving signals that act on the system.

[0028] In the brain energy feature extraction module, the calculation process of the optimal control energy feature for brain state transitions includes: Based on optimal control theory, design starting from the initial brain state x 0 to target brain state The optimal control trajectory, energy consumption during equilibrium state transition, and target state approximation error are considered. The objective function is: , in, This represents the optimal control signal, which, under certain constraints, is the brain network control input signal that minimizes the objective function. The target brain state vector refers to the N×1 dimensional vector corresponding to the final brain functional state that needs to be reached during the brain network state transition process. To control the time domain, The penalty parameter represents the weighting coefficient used to balance the approximation error and control energy consumption; in this embodiment, it is set to 1. Under the constraint of the brain network dynamics equations, the objective function is solved, and the optimal control signal is obtained by simultaneously minimizing the state approximation error and control energy consumption. ; The optimal control energy is the integral of the optimal control signal in the time domain. : , Where N is the total number of brain regions. This represents the optimal control signal for the k-th brain region; The transition energy (TE) between different brain states and the sustaining energy (PE) for maintaining a single brain state are calculated separately. The average transition energy (TE) for all brain state pairs is then taken to obtain the whole-brain average transition energy. Whole-brain control stability is calculated based on the sustaining energy. The formula is: , in, Indicates the stability of whole-brain control. The continuous energy source for maintaining the brain's self-state; The calculation process for the transition energy TE between different brain states is as follows: Select a set of different initial brain states. and target brain state The optimal control signal is obtained by solving the linear dynamic equation of the brain network. The optimal control signal is then substituted into the overall formula of the optimal control energy to calculate the transformation energy value of the state transition. All pairwise different combinations of all brain states are traversed to obtain the total transformation energy, which forms the transformation energy matrix. The calculation process for the sustained energy PE of maintaining a single brain state is as follows: set the initial state and the target state to be equal, substitute the same constraints and objective function, solve for the optimal control signal to maintain the state, substitute the optimal control signal to maintain the state into the overall formula for optimal control energy, and calculate the sustained energy value of the brain state. Calculate the PE value for each brain state separately.

[0029] In the brain energy feature extraction module, the extraction process of whole-brain region energy regulation ability features includes: obtaining the baseline whole-brain average conversion energy by uniformly controlling the input weights throughout the whole brain. For a single brain region Increase its control weight by 1, while keeping the weights of other brain regions unchanged, and calculate the corresponding average conversion energy of the whole brain. brain regions Regional energy regulation capability for: , In this embodiment, the control input weight matrix is ​​set as the baseline matrix B0, and the control weights of all brain regions are set to be the same. The baseline whole-brain average conversion energy is then calculated. ; For the k-th brain region, based on B0, only the control weight of the k-th brain region is increased to obtain B. k Based on B k The whole-brain average conversion energy after recalculating the control weights of the k-th brain region was increased. ; in, Indicates brain regions Regional energy regulation capabilities The larger the value, the more significant the decrease in the average conversion energy of the whole brain after strengthening the control weight of the k-th brain region, indicating that the brain region has a stronger energy regulation ability for the whole brain state; By traversing all brain regions, the rERC values ​​of each brain region are obtained, forming regional energy regulation characteristics; The brain network energy feature set is composed of the average conversion energy of the whole brain, the stability of whole brain control, the conversion energy and sustained energy of each brain state, and the rERC value of each brain region.

[0030] The regional energy regulation capacity (rERC) features extracted in this embodiment can not only achieve binary classification of Parkinson's disease, but also show significant correlation with the Hoehn-Yahr disease classification and UPDRS scale score, enabling quantitative assessment of disease severity. At the same time, it can accurately locate the core brain regions with abnormal energy regulation in Parkinson's disease (such as the sensorimotor cortex, dorsolateral prefrontal cortex, anterior cingulate cortex, insula, etc.), providing objective imaging reference for target screening of neuromodulation therapies such as transcranial magnetic stimulation and deep brain stimulation.

[0031] By using the brain energy feature extraction module, based on the topological constraints of brain structural connections, we can quantify the energy consumption of the brain when switching between different functional states, as well as the ability of a single brain region to regulate the energy of the whole brain, and explore the core pathological features of Parkinson's disease from the perspective of energy mechanisms.

[0032] The multi-dimensional feature fusion module is used to extract core feature sets that are significantly related to Parkinson's disease based on brain dynamics feature sets, brain network energy feature sets, and brain network graph theory feature sets, and to construct a multi-dimensional fusion feature matrix. In the multi-dimensional feature fusion module, the brain network graph theory feature set includes graph theory metrics calculated based on the structural connectivity matrix and the functional connectivity matrix, specifically node degree, clustering coefficient, shortest path length, global efficiency, local efficiency, betweenness centrality, and feature path length. In the multi-dimensional feature fusion module, the brain dynamics feature set, brain network energy feature set, and brain network graph theory feature set are standardized and redundancy-removed. A feature selection algorithm is used to extract the core feature set significantly related to Parkinson's disease. In this implementation, the feature selection algorithm is either LASSO regression with L1 regularization or partial least squares (PLS) algorithm. Age, gender, and years of education are used as covariates during the selection process. Spatial autocorrelation and multiple comparison errors are corrected using permutation tests and spin tests. Features with p_FDR < a set value, such as 0.05, are retained as the core feature set. This reduces the risk of model overfitting and improves the model's generalization ability. Here, p_FDR represents the p-value obtained after correction for the false discovery rate (FDR). This embodiment integrates structural-functional network features, brain state dynamics features, and brain network energy features from multimodal MRI to characterize the whole-brain network abnormalities in Parkinson's disease from multiple dimensions, solving the problems of limited information from single-modal features and low sensitivity in early identification. Validated on clinical datasets, this invention achieves an accuracy of ≥92% in identifying early-stage Parkinson's disease with an AUC ≥0.94, significantly outperforming traditional functional connectivity-based identification methods.

[0033] The Parkinson's disease identification model training module is used to train the constructed Parkinson's disease identification model by taking a multi-dimensional fusion feature matrix as input and the subject's Parkinson's disease clinical diagnosis label and disease severity scale score as output. In this embodiment, the Parkinson's disease identification model can employ Support Vector Machine (SVM), Random Forest (RF), XGBoost, or Lightweight Convolutional Neural Network. The model output labels include binary labels (Parkinson's disease patients / healthy controls), disease severity grading labels (Hoehn-Yahr classification), and continuous value labels (Unified Parkinson's Disease Rating Scale (UPDRS) score). During model training, the final Parkinson's disease recognition model is obtained through cross-validation and hyperparameter optimization. Specifically, the dataset is divided into training, validation and test sets in a 7:1:2 ratio. Five-fold cross-validation is used for model training, and grid search is used for hyperparameter optimization. The area under the receiver operating characteristic curve (AUC), recognition accuracy, sensitivity and specificity are used as model evaluation indicators.

[0034] The Parkinson's disease identification result output module is used to output Parkinson's disease identification results based on the multimodal brain imaging data of the subject to be tested and the trained Parkinson's disease identification model.

[0035] The output of the Parkinson's disease identification model also includes the probability of disease, disease severity classification, and abnormal brain region feature reports. The abnormal brain region feature reports include the location of brain regions with significant rERC abnormalities in Parkinson's disease patients compared to healthy controls, the statistical value of the degree of abnormality, and the brain network to which the abnormal brain regions belong, providing a reference for the screening of neuromodulation therapy targets.

[0036] The above approach, based on non-invasive magnetic resonance imaging data, is non-radioactive, easy to operate, highly reproducible, and compatible with routine 3.0T MRI equipment in clinical settings. It requires no additional hardware investment and can serve as an early screening tool for high-risk groups of Parkinson's disease, as well as an auxiliary decision-making tool for clinical diagnosis, thus possessing extremely high clinical translational value.

[0037] As a specific embodiment, this embodiment implements the Parkinson's disease identification system of the present invention on a dataset containing 120 patients with primary Parkinson's disease and 120 healthy controls matched for age, sex, and years of education. All subjects signed informed consent forms, and the research protocol was approved by the ethics committee.

[0038] See Figure 2 The specific steps are as follows: Step 1: Multimodal brain imaging data acquisition and preprocessing; Data Acquisition: Multimodal brain imaging data of all subjects were acquired using a 3.0T Siemens Skyra MRI scanner. Specific sequence parameters are as follows: T1-weighted structural images: 3D MPRAGE sequence, TR=2000ms, TE=2.01ms, flip angle=9°, resolution 1×1×1mm, scan duration 5min; Diffusion tensor imaging (DTI): Spin echo EPI sequence, TR=10000ms, TE=74ms, b-value=1000s / mm², 64 diffusion directions, resolution 2×2×2mm, scan time 7min; Resting-state functional MRI: Gradient echo EPI sequence, TR=2000ms, TE=30ms, flip angle=90°, resolution 3×3×3mm, 35 axial slices, 240 time points, scan duration 8min. During the scan, the subject was required to close their eyes, remain awake, and keep their head still.

[0039] Data preprocessing; T1 structural image preprocessing: FreeSurfer 7.4.1 was used to complete skull dissection, gray and white matter segmentation, and spatial normalization to the MNI152 template. Brainnetome Atlas was used to divide the whole brain into 246 brain regions (210 cortical brain regions and 36 subcortical brain regions). DTI data preprocessing: MP-PCA denoising and Gibbs artifact removal were performed using MRtrix3, and eddy current correction, head motion correction, and diffusion tensor fitting were performed using FSL. 10 million whole-brain white matter fiber tracts were generated through probabilistic fiber tract tracking, and a 246×246 structural connectivity matrix was constructed. The matrix elements are the average anisotropy fractions (FA) of fiber tracts in two brain regions. The structural connectivity matrix was then symmetricized and normalized. rs-fMRI data preprocessing: DPABI 6.0 was used to discard the first 10 time points, perform slice time correction, head motion correction (excluding subjects with translation >2mm and rotation >2°), spatially normalize to the MNI152 template, smooth the 6mm full-width half-height Gaussian kernel, delinearize drift, and apply 0.01-0.1Hz bandpass filtering. Nuisance covariates such as head motion parameters, white matter signal, and cerebrospinal fluid signal were regressed. Finally, the mean BOLD time series of 246 brain regions was extracted to form a 246×230 functional time series matrix.

[0040] Step 2: Brain state extraction and dynamic feature calculation; Brain state extraction: Based on Brainnetome Atlas, 210 cortical brain regions were divided into 7 resting-state brain networks (VIS, SOM, DAT, VAT, LIM, FPN, DMN). The BOLD signal of the brain regions at each time point was Z-score normalized, and the average activity level of each brain network was calculated. The brain network with the highest activity level at each time point was defined as the dominant network at that time point. Time points where no network activity exceeded the mean were removed. The brain region activity patterns of the same dominant network at all time points were averaged to obtain 7 core brain states for each subject.

[0041] Dynamic feature calculation: For each subject's brain state time sequence, the following features were calculated: Fractional occupancy (FO), dwell time (DT), and occurrence rate (AR) of each brain state; The transition probabilities (TPs) between each of the seven brain states are formed into a 7×7 transition probability matrix; Average time spent in the whole brain (mean of DT in 7 brain states), average occurrence rate in the whole brain (mean of AR in 7 brain states). These features together constitute the brain dynamics feature set, which consists of 65 dimensions.

[0042] Step 3: Brain energy feature extraction based on network control theory; Constructing a linear network control model: using the preprocessed 246×246 structure connection matrix Given the system state matrix, construct the linear dynamic equations of the brain network: , In this embodiment, Let M = N = 246, representing the brain region activity state vector of 246 × 1. To ensure that each brain region can act as an independent control node, let M = N = 246. The control input weight matrix is ​​246×246. It is a time-dependent control input signal of 246×1.

[0043] Optimal control energy calculation: Setting the control time domain T=1 and the penalty parameter ρ=1, the solution is obtained based on optimal control theory from the initial brain state. x 0 to target brain state Optimal control signal Calculate the corresponding optimal control energy:

[0044] Calculate separately: The energy transfer (TE) between each pair of the seven brain states forms a 7×7 energy transfer matrix; The sustained energy (PE) for each brain state to maintain itself. Average energy conversion across the whole brain (mean of TE across all states). Whole-brain control stability is based on continuous energy calculation, and the formula is as follows: The global control stability of the whole brain is obtained by averaging the whole brain control stability of the seven brain states.

[0045] Regional Energy Regulation Capacity (rERC) Calculation: First, a whole-brain uniform control input weighting method is adopted. In this embodiment, The baseline whole-brain average conversion energy was calculated using a 246×246 identity matrix. ; Subsequently, for the k-th brain region, the weight value in the k-th row and k-th column of matrix B is increased by 1, while the other elements remain unchanged, and the corresponding whole-brain average conversion energy is calculated. The rERC value of brain region k is: ; Using this method, the rERC values ​​of each of the 246 brain regions were obtained.

[0046] Brain network energy feature set construction: The brain network energy feature set consists of the whole brain average conversion energy, global control stability, PE value of 7 brain states, TE value of 49 state pairs, and rERC value of 246 brain regions, with a total of 304 features.

[0047] Step 4: Multi-dimensional feature fusion and filtering; Basic feature set construction: Graph theory features of the structural connectivity matrix and functional connectivity matrix are additionally calculated, including node degree, clustering coefficient, shortest path length, global efficiency, local efficiency, and betweenness centrality, totaling 1476 features; these are then fused with the brain dynamics feature set and the brain network energy feature set to form the initial feature set, totaling 1845 features.

[0048] Feature preprocessing: All features are Z-score standardized, and residual processing is performed with age, gender, and years of education as covariates to remove the influence of confounding factors.

[0049] Feature filtering: Step 1: Use a two-sample t-test to compare the characteristic differences between the Parkinson's disease group and the healthy control group, retain the characteristics with p_FDR<0.05, and remove the characteristics with no significant difference between the groups; The second step involves using the LASSO regression algorithm with L1 regularization, with the Parkinson's disease diagnosis label as the dependent variable, to further filter the remaining features. The optimal regularization parameter is selected through 10-fold cross-validation, and finally 87 core features are retained to construct a multi-dimensional fusion feature matrix.

[0050] Step 5: Construction and training of the intelligent recognition model for Parkinson's disease; Dataset splitting: The 240 subjects were randomly divided into a training set (168 cases), a validation set (24 cases), and a test set (48 cases) in a ratio of 7:1:2, ensuring that the ratio of Parkinson's disease patients to healthy controls in each group was 1:1.

[0051] Model selection and training: The XGBoost algorithm was used to construct a Parkinson's disease identification model, with the fused feature matrix as input and binary classification labels (Parkinson's disease = 1, healthy control = 0) as output. The model was trained on the training set using 5-fold cross-validation, and the hyperparameters (learning rate, tree depth, subsampling ratio, etc.) were optimized through grid search, with the AUC value of the validation set as the optimization objective, and the optimal model parameters were finally determined.

[0052] Model performance evaluation: The model performance was evaluated on the independent test set. The results showed that the model had a recognition accuracy of 93.75%, an AUC of 0.958, a sensitivity of 91.67%, and a specificity of 95.83%, which was significantly better than the XGBoost model that only used traditional functional connectivity features (accuracy 81.25%, AUC 0.842).

[0053] Disease severity assessment model: Using the UPDRS-III motor score and Hoehn-Yahr classification of Parkinson's disease patients as outputs, a support vector regression model was constructed based on the core feature set. The correlation coefficients r between the model's predicted values ​​and the actual values ​​were 0.78 and 0.72, respectively (p < 0.001), which achieved an effective quantitative assessment of disease severity.

[0054] Step 6: Subject Identification and Result Output For the subjects to be tested, multimodal brain imaging data is collected, processed in steps 1-4 to obtain a fusion feature matrix, input into the trained recognition model, and the following results are output: Classification and identification results: whether the person is a Parkinson's disease patient, and the corresponding probability of developing the disease; Disease severity assessment results: Hoehn-Yahr classification predictive value, UPDRS score predictive value; Abnormal brain region feature report: This report shows the brain regions in the subject that have significant rERC abnormalities compared to healthy controls, and labels the name of the abnormal brain region, the brain network to which it belongs, and the t-value of the degree of abnormality, providing target references for clinical neuromodulation therapy.

[0055] The method of this invention is standardized and highly scalable, and can be further integrated with multi-omics data such as peripheral blood metabolomics and neurotransmitters to continuously optimize model performance; at the same time, it can be extended to clinical scenarios such as Parkinson's disease subtype classification and drug efficacy prediction, with a wide range of applications.

[0056] As one embodiment, this embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the identification method corresponding to the Parkinson's disease identification system based on brain network control energy characteristics as described above.

[0057] As one embodiment, this embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the identification method corresponding to the Parkinson's disease identification system based on brain network control energy characteristics as described above.

[0058] As one embodiment, this embodiment provides a program product, which is a computer program product, including a computer program. When the computer program is executed by a processor, it implements the steps in the identification method corresponding to the Parkinson's disease identification system based on brain network control energy characteristics as described above.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A Parkinson's disease identification system based on brain network control energy characteristics, characterized in that, include: The multimodal brain imaging data preprocessing module is used to preprocess the acquired multimodal brain magnetic resonance imaging data to construct the whole brain structural connectivity matrix and the brain region functional activity time series, respectively. The brain dynamics feature calculation module is used to extract the brain state at rest based on the time series of brain region functional activities, calculate the time dynamics features of each brain state, and jointly constitute the brain dynamics feature set. The brain energy feature extraction module is used to construct a linear network control model based on the whole brain structure connection matrix as the network topology, calculate the optimal control energy features of brain state transitions and the energy regulation ability features of whole brain regions, and form a brain network energy feature set. The multi-dimensional feature fusion module is used to extract core feature sets that are significantly related to Parkinson's disease based on brain dynamics feature sets, brain network energy feature sets, and brain network graph theory feature sets, and to construct a multi-dimensional fusion feature matrix; the brain network graph theory feature set includes graph theory indicators calculated based on structural connectivity matrices and functional connectivity matrices; The Parkinson's disease identification model construction module is used to train the constructed Parkinson's disease identification model by taking a multi-dimensional fusion feature matrix as input and the subject's Parkinson's disease clinical diagnosis label and disease severity scale score as output. The Parkinson's disease identification result output module is used to output Parkinson's disease identification results based on the multimodal brain imaging data of the subject to be tested and the trained Parkinson's disease identification model.

2. The Parkinson's disease identification system based on brain network control energy characteristics as described in claim 1, characterized in that, The acquired multimodal brain magnetic resonance imaging data included T1-weighted structural MRI data, diffusion tensor imaging (DTI) data, and resting-state functional magnetic resonance imaging (rs-fMRI) data.

3. The Parkinson's disease identification system based on brain network control energy characteristics as described in claim 1, characterized in that, In the brain dynamics feature calculation module, based on the time series of brain region functional activities, the resting state of the brain is extracted, including: Based on the Brainnetome brain atlas, brain regions were divided into classic resting-state brain networks. The BOLD signal of brain regions at each time point was Z-score normalized, and the average activity level of each brain network was calculated. The brain network with the highest activity level at each time point was defined as the dominant brain network at that time point. Time points where no network activity exceeded the mean were removed. The brain region activity patterns of the same dominant network at all time points were averaged to obtain the core brain state of the corresponding number of brain networks.

4. The Parkinson's disease identification system based on brain network control energy characteristics as described in claim 1, characterized in that, In the brain dynamics feature calculation module, the temporal dynamics features of each brain state specifically include: Fractional occupancy (FO) is the proportion of time a single brain state occurs within the total scan time. Dwell time (DT) is the average duration of a single brain state that is maintained continuously. Occurrence rate (AR) is the number of times a single brain state occurs per minute; state transition probability (TP) is the probability of switching from one brain state to another.

5. The Parkinson's disease identification system based on brain network control energy characteristics as described in claim 1, characterized in that, In the brain energy feature extraction module, the calculation process of the optimal control energy feature for brain state transitions includes: Based on optimal control theory, a design is created starting from the initial brain state. x 0 to target brain state The optimal control trajectory, the energy consumption of equilibrium state transition, and the target state approximation error are used to construct the objective function; The constructed brain network dynamics equations are used as constraints to solve the objective function. The optimal control signal is obtained by simultaneously minimizing the state approximation error and control energy consumption. The conversion energy between different brain states and the sustaining energy of maintaining a single brain state are calculated by combining the optimal control signal. The average conversion energy of the whole brain is obtained by averaging the conversion energy of all brain state pairs. The whole brain control stability is calculated based on the sustaining energy of each brain state.

6. The Parkinson's disease identification system based on brain network control energy characteristics as described in claim 1, characterized in that, In the brain energy feature extraction module, the extraction process of energy regulation ability features of the whole brain region includes: The baseline whole-brain average conversion energy is obtained by using the whole-brain uniform control input weight. For a single brain region Increase its control weight by 1, while keeping the weights of other brain regions unchanged, and calculate the corresponding average conversion energy of the whole brain. , combined and Get brain regions The regional energy regulation ability is obtained by traversing all brain regions to obtain the rERC value of each brain region, thus forming regional energy regulation characteristics.

7. The Parkinson's disease identification system based on brain network control energy characteristics as described in claim 1, characterized in that, In the multi-dimensional feature fusion module, a core feature set significantly related to Parkinson's disease is extracted through a feature selection algorithm. The feature selection algorithm is either LASSO regression algorithm with L1 regularization or partial least squares (PLS) algorithm. During the selection process, age, gender, and years of education are used as covariates. Spatial autocorrelation and multiple comparison errors are corrected through permutation tests and spin tests. Features with p_FDR < set value are retained as the core feature set, where p_FDR represents the p value obtained after error detection rate correction.

8. The Parkinson's disease identification system based on brain network control energy characteristics as described in claim 1, characterized in that, The output of the Parkinson's disease identification model also includes the probability of disease, disease severity classification, and abnormal brain region feature reports. The abnormal brain region feature reports include the location of brain regions with significant rERC abnormalities in Parkinson's disease patients compared to healthy controls, the statistical value of the degree of abnormality, and the brain network to which the abnormal brain regions belong.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the identification method steps corresponding to the Parkinson's disease identification system based on brain network control energy characteristics as described in any one of claims 1-8.

10. A program product, said program product being a computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the regulation method steps corresponding to the Parkinson's disease identification system based on brain network control energy characteristics as described in any one of claims 1-8.

Images