Cerebral infarction dyskinesia quantitative evaluation method and evaluation system based on hierarchical Bayesian network

Abstract

The invention relates to a hierarchical Bayesian network-based cerebral infarction dyskinesia quantitative evaluation method and evaluation system, and the method comprises the steps: constructing an evaluation framework based on a motion function hierarchical control theory, screening and parameterizing a key influence factor set from multi-source data, and carrying out the quantitative evaluation of cerebral infarction dyskinesia based on the key influence factor set and the evaluation framework. Constructing a multi-path Bayesian network model; configuring prior probability distribution and conditional probability parameters for the model, and generating an expert prior model; obtaining multi-modal data of a to-be-evaluated object, mapping the multi-modal data to an input node of the model, and outputting the quantization level of the dyskinesia and the probability of each cause path by executing bidirectional probabilistic reasoning; the system comprises a data input unit, a reasoning output unit and a verification and optimization feedback unit. According to the method, the dependence of a traditional method on large-scale labeled samples is broken through, accurate quantitative evaluation and pathological mechanism analysis of the motor function of the cerebral infarction patient are achieved through hierarchical causal modeling, and individualized rehabilitation intervention can be effectively guided.

Description

Technical Field

[0001] This invention relates to healthcare informatics, specifically the field of information and communication technologies for handling or processing medical or health data, and particularly to a method and system for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks. Background Technology

[0002] Cerebral infarction is the main cause of motor dysfunction. The subacute phase is a critical window of greatest neuroplasticity, and accurate quantitative assessment of it is of great significance for rehabilitation decisions.

[0003] Currently, clinical assessments mainly rely on manual scales such as Fugl-Meyer, which suffer from strong subjectivity and difficulty in reflecting real-time dynamic neurophysiological changes. Although existing technologies have begun to utilize EEG and EMG data combined with deep learning algorithms for automated assessment, the following limitations still exist: First, the models have poor interpretability. Deep learning models are mostly "black box" architectures, unable to simulate the hierarchical transmission mechanisms in the motor control system from "cortical planning" to "brain region integration" and then to "peripheral execution," making it difficult to reveal the causal path of functional impairment. Second, the models are highly dependent on large-scale labeled samples. In the specific field of subacute ischemic stroke, the cost of acquiring multimodal correlation data is high and samples are scarce, which limits the accuracy and robustness of traditional data-driven models under small sample conditions.

[0004] Regarding related patents, Chinese patent CN116665895A discloses a quantitative analysis system for Parkinson's disease movement disorders, which extracts indicators through video analysis, but its evaluation mechanism is a static scoring system that lacks dynamic causal logic; Chinese patent CN117594245A discloses a method and system for tracking the rehabilitation process of orthopedic patients, which introduces Bayesian networks, but it is designed for orthopedic bone modeling and lacks deep coupling of the neuromuscular link; Chinese patent CN119541137A discloses a real-time fall risk assessment system and method based on big data, which focuses on general early warning and privacy protection, but cannot achieve quantitative tracing of specific nerve damage links.

[0005] In summary, existing technologies lack a quantitative assessment method that can reduce reliance on large-scale labeled data, deeply integrate motion control theory, and possess hierarchical causal analysis capabilities, thus failing to provide precise target strategies for clinical rehabilitation of cerebral infarction. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a quantitative assessment method and system for motor dysfunction in cerebral infarction based on hierarchical Bayesian networks.

[0007] To achieve the above objectives, the technical solution adopted by this invention is a method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks, the method comprising the following steps:

[0008] S1 constructs an evaluation framework based on the hierarchical control theory of motion function, and selects and parameterizes a set of key influencing factors from multi-source data;

[0009] S2 Based on the set of key influencing factors and the evaluation framework, a multi-path Bayesian network model is constructed; prior probability distribution and conditional probability parameters are configured for the model to generate an expert prior model.

[0010] S3 acquires the multimodal data of the object to be evaluated, maps it to the input nodes of the model, and outputs the quantitative level of motor dysfunction and the probability of each causal path by performing bidirectional probabilistic inference.

[0011] Preferably, screening the set of key influencing factors includes the following steps:

[0012] S1.1 Based on the motion function hierarchical control theory, a causal hierarchical structure is constructed, and each level is abstracted as an intermediate node;

[0013] S1.2 Collect expert experience and construct a multimodal initial influencing factor pool corresponding to each level;

[0014] S1.3 Establish a multi-dimensional evaluation matrix and score the importance of initial influencing factors by integrating expert experience with subjective and objective weighting methods;

[0015] S1.4 Based on the scoring and ranking results, key influencing factors are determined by the quantile screening method and used as input nodes for the multi-path Bayesian network model.

[0016] Preferably, in S1.1, the causal hierarchical structure includes a cortical motor planning layer, a brain functional integration layer, and a neuromuscular output layer.

[0017] Preferably, in S1.3, a multi-dimensional evaluation matrix is ​​established; by integrating expert experience through subjective and objective weighting methods, the initial weights of each dimension are calculated and obtained, thereby updating the multi-dimensional evaluation matrix and its corresponding weights.

[0018] Preferably, in S2, the multi-path Bayesian network model includes an input layer, an intermediate node layer, and an output evaluation layer;

[0019] The input layer corresponds to key influencing factors, the intermediate node layer performs specific functional fusion of multi-source information from the input layer, and the output evaluation layer integrates the states of each intermediate node to generate the final evaluation result.

[0020] Preferably, the information transmission paths in the multi-path Bayesian network model include a vertical transmission path from the input layer through the intermediate node layer to the output layer, a horizontal coupling path across functional intermediate nodes, and a comprehensive convergence path integrating intermediate nodes and background default nodes.

[0021] Preferably, the node parameters of the multipath Bayesian network model include the marginal probabilities of input nodes and the conditional probabilities of intermediate and output nodes.

[0022] The weighted strength of causal relationships between nodes is determined by fusing expert scores with the fuzzy Hausdorff distance algorithm.

[0023] Preferably, a conditional probability table is generated for nodes in the non-input layer based on whether the multimodal data to be evaluated of the object to be evaluated is a linear normal state or a nonlinear abnormal state.

[0024] Preferably, in S3, bidirectional probabilistic reasoning includes calculating the probability distribution of movement disorder levels through forward propagation, tracing key causal pathways through backward diagnosis, and evaluating the contribution of different intervention targets based on the change in the ratio of the posterior probability distribution to the prior probability distribution.

[0025] An assessment system employing the aforementioned hierarchical Bayesian network-based quantitative assessment method for motor dysfunction in cerebral infarction includes:

[0026] A data input unit is used to receive multimodal data to be evaluated that matches key influencing factors;

[0027] An inference output unit, which incorporates the multi-path Bayesian network model, is used to perform bidirectional probabilistic inference on the multimodal data to be evaluated, and outputs the probability distribution of obstacle level, the causal path analysis results, and the matching priority strategy.

[0028] A verification and optimization feedback unit is used to correct the prior probability and conditional probability parameters in the model based on feedback.

[0029] This invention relates to a method and system for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks. The assessment framework is constructed based on the hierarchical control theory of motor function. A set of key influencing factors is screened and parameterized from multi-source data. Based on the set of key influencing factors and the assessment framework, a multi-path Bayesian network model is constructed. Prior probability distributions and conditional probability parameters are configured for the model to generate an expert prior model. Multimodal data of the object to be assessed is acquired and mapped to the input nodes of the model. Through bidirectional probabilistic inference, the quantitative level of motor dysfunction and the probability of each causal path are output. The system includes a data input unit, an inference output unit, and a verification and optimization feedback unit.

[0030] The beneficial effects of this invention are as follows:

[0031] (1) Construct a hierarchical causal framework that is in deep fit with motor control theory; break through the limitations of traditional assessment methods in cross-level integration and causal mechanism analysis, realize full-pathway quantitative analysis from "cortical planning and brain region integration" to "neuromuscular execution", and accurately locate the hierarchical damaged links of functional disorders through multi-path Bayesian network modeling, providing target guidance with pathological mechanism explanation for clinical rehabilitation.

[0032] (2) A multi-layered mechanism for the scientific screening and knowledge integration of influencing factors was established. Through the four-layered mechanism of "theory guidance - data screening - expert calibration", the scientific transformation of multimodal data and clinical experience was realized by using the AHP-entropy weight method and the fuzzy Hausdorff distance algorithm. This effectively solved the problem of scarce labeled samples in the field of subacute cerebral infarction, enabling the model to obtain prior parameter configurations with clinical consistency without the need for large-scale clinical data training.

[0033] (3) The innovative introduction of nonlinear evolution and logical compensation algorithms enhances the robustness and accuracy of the assessment. Addressing the complexity of clinical decision-making, an exponential decay mechanism simulates the nonlinear accumulation of risk, an overflow correction operator eliminates computational dead zones, and strong constraint rules ensure the certainty of severe injury assessment. This customized probabilistic fusion algorithm effectively solves the logical "passivation" and "discontinuity" problems exhibited by traditional Bayesian models under complex evidence combinations, significantly enhancing the model's diagnostic robustness under extreme pathological conditions.

[0034] This invention breaks through the reliance of traditional methods on large-scale labeled samples. Through hierarchical causal modeling, it achieves accurate quantitative assessment of motor function and pathological mechanism analysis in patients with cerebral infarction, which can effectively guide individualized rehabilitation intervention. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method of the present invention;

[0036] Figure 2 This is a structural diagram of the multipath Bayesian network model in this invention;

[0037] Figure 3 This is a schematic diagram of the system structure of the present invention;

[0038] Figure 4 A visualization of the contribution of influencing factors in this invention;

[0039] Figure 5 This is a schematic diagram of the model reasoning and verification process in this invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0041] This invention relates to a method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks, such as... Figure 1 As shown, the method includes the following steps:

[0042] (1) An evaluation framework was constructed based on the Hierarchical Motor Control Theory, and a set of key influencing factors was screened and parameterized from multi-source data;

[0043] (2) Based on the set of key influencing factors and the evaluation framework, construct a multi-path Bayesian network model; configure the prior probability distribution and conditional probability parameters for the model to generate an expert prior model;

[0044] (3) Obtain the multimodal data of the object to be evaluated, map it to the input node of the model, and output the quantitative level of motor dysfunction and the probability of each causal path by performing bidirectional probabilistic reasoning.

[0045] The above method achieves quantitative assessment and mechanism analysis of motor dysfunction in cerebral infarction through knowledge-driven structural modeling and algorithm integration.

[0046] The method will be explained below with specific steps.

[0047] (1) An evaluation framework was constructed based on the hierarchical control theory of motion function, and a set of key influencing factors was screened and parameterized from multi-source data;

[0048] (1-1) Hierarchical Abstraction

[0049] Based on the hierarchical control theory of motion function, a causal hierarchical structure is constructed, and each level is abstracted as an intermediate node.

[0050] The causal hierarchical structure includes a cortical motor planning layer, a brain functional integration layer, and a neuromuscular output layer.

[0051] (1-2) Indicator Construction

[0052] Collect expert experience and construct a multimodal initial influencing factor pool corresponding to each level;

[0053] (1-3) Overall Score

[0054] Establish a multi-dimensional evaluation matrix and score the importance of initial influencing factors by integrating subjective and objective weighting methods with expert experience;

[0055] A multi-dimensional evaluation matrix is ​​established; by integrating expert experience through subjective and objective weighting methods, the initial weights of each dimension are calculated and obtained, thereby updating the multi-dimensional evaluation matrix and its corresponding weights.

[0056] (1-4) Quantile Filtering

[0057] Based on the scoring and ranking results, key influencing factors are determined using the quantile screening method and used as input nodes for the multi-path Bayesian network model.

[0058] In this invention, the process of determining the set of key influencing factors employs a multi-layered safeguard mechanism combining theoretical guidance and empirical calibration to ensure the scientific rigor of factor identification and the stability of the system; specifically as follows:

[0059] First, based on Hierarchical Motor Control Theory, a causal hierarchical framework comprising three functional levels is constructed:

[0060] The motor planning layer of the cortex involves the regulatory state of the cerebral cortex in the generation of motor intentions, strategy formulation and action plan selection, and reflects the cortical regulatory capacity of motor preparation.

[0061] The brain region functional integration layer describes the coordination mechanism between various functional brain regions during motor tasks, including cross-regional neural synchronization, network connectivity efficiency, and functional lateralization, reflecting the integrative and adaptive nature of the brain's motor control network.

[0062] The neuromuscular output layer corresponds to the final execution stage of the descending commands, including the final action execution characteristics such as muscle recruitment sequence, joint range of motion, and postural stability.

[0063] Within the above framework, technical information is sequentially transmitted between the cortical motor planning layer, the brain functional integration layer, and the neuromuscular output layer. The information from each level eventually converges to the subsequent output evaluation layer for quantitative output.

[0064] Based on this theoretical structure, an initial pool of influencing factors was constructed, and by analyzing the physiological nature and functional attributes of each influencing factor, its correspondence with the aforementioned specific levels was clarified. For example, cortical excitability indicators were assigned to the cortical motor planning layer, cross-brain region connectivity efficiency indicators were assigned to the brain region functional integration layer, and electromyographic synergy parameters and kinematic characteristics were assigned to the neuromuscular output layer.

[0065] Expert experience A refers to the introduction of an expert experience calibration mechanism to verify and optimize the clinical rationality of the above hierarchical division. Specifically, each classification item is corrected by setting a preset consistency threshold, and indicators with a degree of disagreement higher than the preset threshold are eliminated, thereby forming a stable factor classification scheme under the consensus of clinical experts.

[0066] Through a multi-layered safeguard mechanism combining theory and experience, this invention has initially established 25 candidate factors covering multi-source data such as EEG signal characteristics, structural imaging parameters, and electromyography indices, forming an initial influencing factor pool, including:

[0067] Cortical motion planning layers: CP1 - slow wave (alpha wave ratio), CP2 - absolute slow wave power, CP3 - DWI infarct volume, CP4 - gray-white matter boundary integrity, CP5 - CT lesion location distribution (PMC), CP6 - CT lesion location distribution (M1), CP7 - MCA mean flow velocity (Vm), CP8 - CST structural integrity (FA value).

[0068] Brain region functional integration layers: NI1 - Laterality Coefficient, NI2 - PMC & M1 connectivity (β segment mean), NI3 - SMA & M1 connectivity (β segment mean), NI4 - Bilateral transmission coherence, NI5 - Delta & Alpha Ratio, NI6 - Microstate C time coverage, NI7 - Microstate C→D transition probability, NI8 - α peak frequency, NI9 - Microstate dynamic characteristics;

[0069] Neuromuscular output layer: MS1 - H / M maximum amplitude ratio, MS2 - CMAP amplitude, MS3 - motor nerve conduction velocity, MS4 - F wave latency, MS5 - MUAP amplitude, MS6 - FMA (upper limb), MS7 - FMA+ADL (upper limb), MS8 - WFMT+ADL (upper limb).

[0070] After obtaining the initial pool of influencing factors, expert experience was introduced to establish a multi-dimensional evaluation matrix. The multi-dimensional dimensions include motion relevance, functional impact, and intervention responsiveness. Each candidate influencing factor was scored, with each dimension divided into 1-5 levels. A subjective and objective weighting fusion method combining AHP (Analytic Hierarchy Process) and entropy weighting (each accounting for 50% of the weight) was used, and the matrix was tested for consistency (CR < 0.1) to ensure its rationality. The weight values ​​of the three dimensions were calculated, and the importance of the initial influencing factors was comprehensively scored accordingly.

[0071] Based on the comprehensive scoring ranking results of the initial factors, the top 80% of the key influencing factors are retained. At the same time, in order to avoid redundancy between scales, a strategy of selective retention is implemented for assessment tools with similar functions.

[0072] After the above quantile screening and optimization process, 20 core influencing factors were finally extracted from the initial candidate set, forming a key influencing factor combination for constructing a multi-path Bayesian network model, such as... Figure 4 As shown.

[0073] The factors that were eliminated mainly included indicators that ranked low in the overall weight score or had a high degree of functional overlap with the factors that were retained.

[0074] (2) Based on the set of key influencing factors and the evaluation framework, construct a multi-path Bayesian network model; configure the prior probability distribution and conditional probability parameters for the model to generate an expert prior model;

[0075] To construct the model, it is necessary to determine the causal strength relationship between nodes. This relationship is obtained by collecting and processing expert experience B. The expert experience B is different from the expert experience A used for factor screening in (1), and specifically refers to the qualitative assessment data on the strength of the causal relationship between nodes in the combination of influencing factors provided by clinical experts.

[0076] To overcome the inherent ambiguity and uncertainty in expert subjective judgment and improve the rationality of weight allocation and the accuracy of conditional probability table generation, this invention employs the following steps for consistency fusion and weight calculation of expert questionnaire B:

[0077] (2-1) Expert experience integration and weight calculation

[0078] The node parameters of the multipath Bayesian network model include the marginal probabilities of the input nodes and the conditional probabilities of the intermediate and output nodes.

[0079] The weighted strength of causal relationships between nodes is determined by fusing expert scores with the fuzzy Hausdorff distance algorithm.

[0080] A custom triangular fuzzy number scoring scale is defined to map the 1-5 point integer scores of experts on the strength of causal relationships to the corresponding set of triangular fuzzy numbers. The mapping relationship is shown in Table 1 below.

[0081] Table 1 Mapping Relationship Table

[0082]

[0083] To quantify the differences in scores from multiple experts and calculate their weights, fuzzy Hausdorff distance is used, and the calculation formula is as follows:

[0084]

[0085] Where A=(a1,a2,a3) and B=(b1,b2,b3) are triangular fuzzy numbers;

[0086] Based on this, calculate the sum of the differences between each expert's score and the scores of all other experts. ,

[0087]

[0088] in, For the sum of experts, For experts With experts Fuzzy Hausdorff distance between scores;

[0089] Subsequently, the weight of each expert was determined using the reciprocal normalization method. :

[0090]

[0091] This weighting mechanism ensures that experts with high scoring consistency receive greater weight, effectively suppressing the impact of extreme scores.

[0092] (2-2) Network Topology Construction

[0093] The logical association of topological paths is based on the weighted strength calculated above. In this embodiment, the logical association of multiple paths is constructed in a Bayesian network.

[0094] The multipath Bayesian network model includes an input layer, an intermediate node layer, and an output evaluation layer;

[0095] The input layer corresponds to key influencing factors, the intermediate node layer performs specific functional fusion of multi-source information from the input layer, and the output evaluation layer integrates the states of each intermediate node to generate the final evaluation result.

[0096] Furthermore, the information transmission paths in the multi-path Bayesian network model include a longitudinal transmission path from the input layer through intermediate node layers to the output layer, a lateral coupling path across functional intermediate nodes, and a comprehensive convergence path integrating intermediate nodes and background default nodes. The longitudinal transmission path describes the cascading transmission from "input layer index → ​​intermediate functional level node → output layer obstacle level," reflecting the hierarchical command flow of motor control. The lateral coupling path describes the mutual constraints between intermediate nodes such as the "cortical motor planning layer" and the "brain region functional integration layer," reflecting lateral inhibition or synergy within the system. The convergence path directs the outputs of all intermediate layer nodes to the final evaluation node and reserves a default node interface, such as... Figure 2 As shown.

[0097] In this invention, a network structure is constructed based on the causal strength relationship influence values ​​defined by expert questionnaire B. The input layer of the network contains three input units. In this embodiment, the encoding CP1~CP6, NI1~NI8, and MS1~MS6 are three input units respectively. The intermediate node layer includes the corresponding CP9, NI10, MS9 and the default layer as intermediate nodes. The output layer corresponding to Out1 represents the overall degree of motor dysfunction. The Leaky Noisy-OR model is used to parametrically model the causal mechanism of the model.

[0098] The probability synthesis logic of the output evaluation layer integrates a Leak Probability parameter, which is constructed based on an improved Leaky Noisy-OR model to quantify the causal effect of potential background factors not explicitly modeled in the input layer on the final motor function impairment assessment results.

[0099] Unlike discrete nodes that require external observation input, this invention sets the omission probability parameter as a built-in constant factor L to characterize the baseline background risk level in a clinical setting. During probabilistic inference, if all explicit intermediate nodes are in a normal state, the system automatically provides a noise floor probability distribution from this omission probability factor L, ensuring that the assessment results still conform to the baseline risk distribution in clinical statistics even without any explicit evidence of injury. Through this mechanism, the model achieves automated closed-loop processing of "invisible risks" without requiring external provision of active status data for default nodes.

[0100] The output layer includes three discrete nodes. The outputs of all discrete nodes are normalized and the calculation deviation is calibrated to ensure the probabilistic rationality of the evaluation structure.

[0101] (2-3) Conditional probability table generation mechanism

[0102] In this invention, the node parameters of the multi-path Bayesian network model are driven by the prior probability of the root node and the conditional probability table of the non-root nodes. To ensure the interpretability of the model, a parameter configuration path from expert knowledge to nonlinear evolution algorithm is proposed.

[0103] Based on the multimodal evaluation data of the object to be evaluated, which is either a linear normal state or a nonlinear abnormal state, a conditional probability table is generated for nodes in the non-input layer; the exponential decay logic is different in the two states.

[0104] Specifically, in implementation, most functional nodes are set to three states—normal state, mild impairment state, and severe impairment state, with the latter two being "abnormal states." However, for cases involving key anatomical locations (such as M1 and PMC), a binary state is set, transforming continuous multimodal signals into logical evidence with statistical discriminative power. The classification criteria and discretization standards for these factors can be based on, but are not limited to, guidelines, literature, and expert experience. For example, for the slow wave-alpha wave ratio (CP1), a value less than 1 indicates a normal state, a value greater than 2 indicates a severe impairment state, and the rest indicate a mild impairment state (cited from People's Medical Publishing House. Neuroelectrophysiology (EEG) Technology (Technician / Intermediate) Examination Guide [M]. Beijing: People's Medical Publishing House, 2025.). Those skilled in the art can set classification criteria according to their needs and adopt standards with equal or higher requirements.

[0105] Define the probability that a node is in a normal state. With the number of abnormal parent nodes The increase in decreases exponentially, satisfying

[0106]

[0107] Wherein, γ is the evolution decay coefficient, which is taken as 0.4 in this embodiment. As a benchmark health weight, To correct for residuals; this formula reflects the biological characteristic that the more damaged modules there are, the lower the probability of the system maintaining steady state becomes.

[0108] When multiple parent nodes are damaged simultaneously, the state enters the category of abnormal state. A synergy factor is introduced to nonlinearly enhance the obstacle probability and simulate the resonance effect of multiple weak damages.

[0109]

[0110] in, It is based on the initial weights after expert knowledge fusion, i.e., the aforementioned ; It is the collaborative gain parameter;

[0111] To prevent inference hardening caused by zero-probability dead zones during computation, an overflow correction operator is introduced to perform a minor redistribution of the probability distribution.

[0112]

[0113] Where λ is the minimum offset, which is 0.01 in this embodiment, and K is the total number of node states; this correction ensures the numerical stability of the Bayesian network under extreme pathological sample conditions.

[0114] The output layer nodes employ a strong constraint Leaky Noisy-OR ensemble. At the final evaluation node, an improved mechanism integrating explicit constraints and background noise absorption is used, following the "critical link failure determinism." If any intermediate layer node triggers the S2 severe obstacle state, the model activates the strong constraint logic, directly locking the posterior probability of the severe obstacle at the output node; otherwise, it uses P... leak By absorbing the influence of non-modeling background factors, the probability distribution of the final grade is calculated.

[0115]

[0116] Where m is the total number of abnormal input nodes in the current node; k is the index of the intermediate node; P leakis the background noise probability; wki is the contribution strength of the k-th intermediate node to state i, i=0,1,2, corresponding to the normal state, the slightly obstructed state, and the severely obstructed state, respectively, and is associated with P(S0|n) and P(S1); the synthesized initial probability is... The distribution is redistributed again using the overflow correction operator to obtain the final quantitative evaluation distribution that eliminates the zero-probability dead zone. .

[0117] Table 2 shows examples of conditional probability distributions for key nodes, including typical scenarios such as all nodes being normal, a single node being abnormal, and multiple nodes being abnormal.

[0118] Table 2 Conditional Probability Distribution of Key Nodes

[0119]

[0120] In Table 2, when all node combinations are "normal", the output layer probability distribution is {0.85, 0.10, 0.05}; this distribution is determined by the built-in leak probability factor P. leak The stable numerical results are obtained by simulating the basic disease background risk and then performing a small redistribution through the overflow correction operator.

[0121] (3) Obtain multimodal data of the object to be evaluated, map it to the input node of the model, and output the quantitative level of motor dysfunction and the probability of each causal path by performing bidirectional probabilistic inference.

[0122] Specifically, the data for CP1~CP2 and NI1~NI8 were collected from EEG detection data and determined based on clinical measured values ​​and standardized delay rates. The standardized delay rate here is a normalization concept in the data processing flow, which is used to convert the degree of deviation of the measured value from the normal range into a relative utility value that can be compared across indicators. For CP4, CP6~CP7, and MS6 data, they are generally collected from clinical experts (expert experience), while CP8 data is processed using 3D slicer software. The software delineates the area of ​​cerebral infarction in the patient's brain and automatically calculates the infarct volume. For MS1, the M-wave amplitude and H-wave amplitude of the left and right tibial nerves are extracted from the H-reflex table, the corresponding H / M-wave amplitude ratio is calculated, and classification is performed based on the threshold. For the MS2-MS3 data, the lower limit of normal percentage (LLN%) was used for normalization (Tomanovic-Vujadinovic et al., 2020), mapping the absolute physical quantity to a relative utility value representing the functional achievement rate; the data of the MS1, MS4 and MS5 input nodes were determined based on clinical measured values ​​and standardized delay rates. Specifically, for MS2, the distal amplitude is extracted from the motor nerve conduction velocity table, the ratio of the distal amplitude to the LLN benchmark of this item is calculated to obtain the achievement rate, and classification is performed based on the threshold. For MS3, segmental velocity is extracted from the motor nerve conduction velocity table, the ratio of segmental velocity to the LLN benchmark of this item is calculated, the motor nerve conduction velocity achievement rate is obtained, and classification is performed based on the threshold. For MS4, the F latency is extracted from the F-wave table. The relationship between the measured F latency and the normal upper limit of the tibial nerve F wave is calculated and expressed as (measured F latency - normal upper limit) / normal upper limit * 100%. The F wave latency delay rate is obtained and classified based on the threshold. For MS5, after obtaining the average amplitude from the EMG summary table, classification is performed based on a preset range threshold. After processing all the data, it is mapped to the input node of the model.

[0123] Two-way probabilistic reasoning includes calculating the probability distribution of movement disorder levels through forward propagation and tracing key causal pathways through backward diagnosis;

[0124] Among them, forward reasoning is used to calculate the posterior probability distribution of each level of motor dysfunction to quantify the degree of damage, while backward reasoning is used to trace the key causal path leading to the dysfunction based on the diagnostic reasoning algorithm and locate the core damaged link.

[0125] The contribution of different intervention targets is assessed based on the change in the ratio of the posterior probability distribution to the prior probability distribution.

[0126] In this invention, a chain-like inference mechanism based on a probabilistic graphical model is employed to quantitatively map from underlying dispersed factors to the overall obstacle level during the evaluation of the multi-path Bayesian network model. Specifically, the state information of the bottom-level nodes is initially fused based on their conditional dependencies with intermediate nodes, updating the state probability distributions of intermediate nodes such as CP9, NI10, and MS9. The posterior probabilities of these intermediate nodes then serve as upper-level evidence, undergoing a secondary fusion with the conditional probability tables of the final evaluation layer nodes, ultimately forming a comprehensive assessment of motor function impairment status at the out1 node. Throughout this process, causal interactions, synergistic effects, and default effects between nodes are embedded in the conditional probability structure defined by expert knowledge. The system achieves automatic integration and state decoding of cross-level information through rigorous Bayesian inference, thereby completing the quantitative mapping from dispersed factors to the overall obstacle level.

[0127] This invention also relates to an assessment system employing the aforementioned hierarchical Bayesian network-based quantitative assessment method for motor dysfunction in cerebral infarction, such as... Figure 3 As shown, it includes:

[0128] A data input unit is used to receive multimodal data to be evaluated that matches key influencing factors;

[0129] An inference output unit, which incorporates the multi-path Bayesian network model, is used to perform bidirectional probabilistic inference on the multimodal data to be evaluated, and outputs the probability distribution of obstacle level, the causal path analysis results, and the matching priority strategy.

[0130] A verification and optimization feedback unit is used to correct the prior probability and conditional probability parameters in the model based on feedback. Specifically, it obtains clinical follow-up data and scale results, compares the consistency between the model inference results and clinical results, and then corrects the prior probability and conditional probability parameters in the model.

[0131] To verify the evaluation performance and clinical applicability of the multi-path Bayesian network model constructed in this invention, such as... Figure 5 As shown, this embodiment conducts systematic testing and analysis through the following steps to demonstrate the effectiveness, stability, and interpretability of the model in identifying motor dysfunction at the individual level.

[0132] Data preparation for validation: A sample of patients with cerebral infarction who were clinically diagnosed and covered different degrees of motor dysfunction (e.g., normal, mild, and severe). For each sample, based on the 20 key influencing factors determined in step (1) above, corresponding EEG, imaging, EMG, and scale objective test data were extracted. The continuous variables were state-determined according to the preset discretization rules and transformed into discrete state features that the model could recognize, which served as the set of observational evidence.

[0133] Automated inference execution: Preprocessed subject data is input as observational evidence into the constructed expert prior model. The model invokes its embedded precise inference algorithm; in this embodiment, the Netica inference engine is used to automatically perform cross-level probability propagation and logical coupling. Finally, a posterior probability distribution of the motor dysfunction status (normal, mild, severe) is output for each patient.

[0134] Results Comparison and Analysis: The assessment results output by the model are compared and analyzed with the standard results determined by clinical experts. These standard results should be independently determined by at least two rehabilitation medicine experts based on the patient's functional performance. If there is disagreement, it is determined through arbitration by a senior third-party expert to ensure the objectivity of the benchmark. Based on the comparison results, the following multidimensional indicators are used for quantitative analysis:

[0135] Step 1: Calculate the posterior probability;

[0136] For each patient, the multipath Bayesian network model outputs the posterior probability distribution of their impairment state (normal, mild impairment, severe impairment) based on the input key factor states, as follows:

[0137]

[0138] Take the maximum posterior probability value As a confidence index for model predictions;

[0139] Step 2: Consistency assessment;

[0140] The model's predicted results are compared with the standard clinical assessment results; if they are consistent, the consistency index is... ,otherwise Alternatively, consistency coefficients such as the K-value can be used for quantification.

[0141] Step 3: Constructing a comprehensive score;

[0142] For each patient, a comprehensive scoring function is constructed, expressed as follows:

[0143]

[0144] Here, α and β are linear weighting coefficients used to balance the importance of the model's own inference confidence and clinical consistency;

[0145] Step 4: Sort and output;

[0146] Based on the overall score The results of all patients were sorted in descending order.

[0147]

[0148] Obtain a ranking table of the matching results between the model and real cases.

[0149] The case verification results and sorting examples are shown in Table 3;

[0150] Table 3. Case Validation Results and Sorting Examples

[0151]

[0152] The posterior probability distribution is output by the Bayesian network, and the distributions are the probabilities of "normal, slight impairment, and severe impairment". The ranking results are arranged from high to low according to the comprehensive score, reflecting the matching effect and reliability of the model on the case set.

[0153] The results showed that 7 out of 9 patients (approximately 77.8%) had results consistent with clinical expert assessments, validating the effectiveness of the model in identifying motor dysfunction states. Further comprehensive ranking indicated that the predicted results were not only consistent with clinical findings but also exhibited extremely high model confidence, representing the most stable cases. Pt-06 and Pt-09 ranked lowest; although they had high posterior probabilities, they were downgraded due to inconsistencies with expert assessments, suggesting their results were in a potential misjudgment zone and requiring further verification. Pt-03 and Pt-05 had relatively low maximum posterior probabilities, but still received moderate scores due to consistency with expert conclusions, reflecting the model's robustness under boundary conditions.

[0154] We selected representative cases and conducted an in-depth analysis of their evidence propagation paths within the model. By tracing the probability changes of key nodes in the model, we revealed the core factors and causal chains that dominate the final assessment results. We compared the model's reasoning logic with clinical diagnostic logic to examine the rationality and clinical interpretability of the model's decision-making process.

[0155] Through the above systematic verification process, it can be comprehensively confirmed that the method and system described in this invention have good accuracy, robustness and clinical application value in the quantitative assessment of motor dysfunction.

[0156] In specific implementations, the present invention also relates to a computer-readable storage medium storing a program for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks. When the program is executed by a processor, it implements the aforementioned method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks. The invention also relates to 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 aforementioned method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks.

[0157] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0158] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks, characterized in that: The method includes the following steps: S1 constructs an evaluation framework based on the hierarchical control theory of motion function, and selects and parameterizes a set of key influencing factors from multi-source data; S2 constructs a multi-path Bayesian network model based on the set of key influencing factors and the evaluation framework; configures the prior probability distribution and conditional probability parameters for the model to generate an expert prior model. S3 acquires the multimodal data of the object to be evaluated, maps it to the input nodes of the model, and outputs the quantitative level of motor dysfunction and the probability of each causal path by performing bidirectional probabilistic inference.

2. The method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks according to claim 1, characterized in that: Screening the set of key influencing factors involves the following steps: S1.1 Based on the motion function hierarchical control theory, a causal hierarchical structure is constructed, and each level is abstracted as an intermediate node; S1.2 Collect expert experience and construct a multimodal initial influencing factor pool corresponding to each level; S1.3 Establish a multi-dimensional evaluation matrix and score the importance of initial influencing factors by integrating expert experience with subjective and objective weighting methods; S1.4 Based on the scoring and ranking results, key influencing factors are determined by the quantile screening method and used as input nodes for the multi-path Bayesian network model.

3. The method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks according to claim 2, characterized in that: In S1.1, the causal hierarchical structure includes a cortical motor planning layer, a brain functional integration layer, and a neuromuscular output layer.

4. The method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks according to claim 2, characterized in that: In S1.3, a multi-dimensional evaluation matrix is ​​established; by integrating expert experience through subjective and objective weighting methods, the initial weights of each dimension are calculated and obtained, thereby updating the multi-dimensional evaluation matrix and its corresponding weights.

5. The method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks according to claim 1, characterized in that: In S2, the multipath Bayesian network model includes an input layer, an intermediate node layer, and an output evaluation layer; The input layer corresponds to key influencing factors, the intermediate node layer performs specific functional fusion of multi-source information from the input layer, and the output evaluation layer integrates the states of each intermediate node to generate the final evaluation result.

6. The method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks according to claim 5, characterized in that: The information transmission paths in the multi-path Bayesian network model include the vertical transmission path from the input layer through the intermediate node layer to the output layer, the horizontal coupling path across functional intermediate nodes, and the comprehensive convergence path integrating intermediate nodes and background default nodes.

7. The method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks according to claim 5, characterized in that: The node parameters of the multipath Bayesian network model include the marginal probabilities of the input nodes and the conditional probabilities of the intermediate and output nodes. The weighted strength of causal relationships between nodes is determined by fusing expert scores with the fuzzy Hausdorff distance algorithm.

8. The method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks according to claim 7, characterized in that: Based on the multimodal evaluation data of the object to be evaluated, which are either linear normal states or nonlinear abnormal states, a conditional probability table is generated for nodes in the non-input layer.

9. The method for quantitative assessment of motor dysfunction in cerebral infarction based on hierarchical Bayesian networks according to claim 1, characterized in that: In S3, bidirectional probabilistic reasoning includes calculating the probability distribution of movement disorder levels through forward propagation and tracing key causal pathways through backward diagnosis; and evaluating the contribution of different intervention targets based on the change in the ratio of the posterior probability distribution to the prior probability distribution.

10. An assessment system employing the hierarchical Bayesian network-based quantitative assessment method for motor dysfunction in cerebral infarction as described in any one of claims 1 to 9, characterized in that: include: A data input unit is used to receive multimodal data to be evaluated that matches key influencing factors; An inference output unit, which incorporates the multi-path Bayesian network model, is used to perform bidirectional probabilistic inference on the multimodal data to be evaluated, and outputs the obstacle level probability distribution, causal path analysis results, and matching priority strategy. A verification and optimization feedback unit is used to correct the prior probability and conditional probability parameters in the model based on feedback.

Images