Vertebral artery dissection risk assessment method and system

Abstract

The invention discloses a vertebral artery dissection risk assessment method and system, and relates to the technical field of risk assessment, and the method comprises the steps: collecting ultrasonic dynamic blood parameter data of a vertebral artery of a patient; based on the ultrasonic dynamic blood parameter data, constructing a local blood flow kinetic energy disturbance time sequence diagram by adopting a hemodynamics time sequence analysis method, and performing time sequence prediction to generate a blood flow function disturbance prediction result; dynamically adjusting a risk score threshold by adopting a reinforcement learning algorithm, combining the adjusted risk score threshold with a blood flow function disturbance prediction result, and generating a personalized vertebral artery dissection risk score according to a risk score calculation model; and generating a vertebral artery dissection risk assessment result by using a multi-modal fusion mechanism in combination with the personalized vertebral artery dissection risk score and the multi-source heterogeneous medical feature set. According to the method, the patient static characteristics and the future disturbance trend serve as input, the optimal risk score threshold value is dynamically selected, and the personalized score is generated according to the disturbance threshold exceeding proportion and time weight information.

Description

Technical Field

[0001] The present invention relates to the technical field of risk assessment, and in particular to a vertebral artery dissection risk assessment method and system. Background Art

[0002] Vertebral artery dissection is an acute cerebrovascular event with a high disability rate and high misdiagnosis rate. Early identification and risk assessment are of great value for clinical prevention. Currently, vertebral artery dissection risk assessment technology mainly relies on static images, biomarkers or structured clinical data, combined with traditional statistical methods or machine learning models for discrimination. However, the discrimination methods often lack the ability to continuously perceive hemodynamic changes and find it difficult to characterize the functional disturbance process of individuals under different physiological states. In addition, the patient's ultrasound dynamic blood parameters and medical characteristics such as historical clinical information and genetic data are often in a heterogeneous distribution state, and a unified and efficient fusion mechanism has not yet been established to improve the personalization and timeliness of the assessment model.

[0003] There are two limitations in the field of vertebral artery dissection risk assessment: first, fixed rules or empirical thresholds are often used to set risk levels, which cannot adapt to the individual dynamic changes of patients' blood flow function disturbances and limit the accuracy of the scoring system; second, as a non-invasive, real-time, high-resolution blood flow monitoring method, ultrasound Doppler imaging technology has been widely used in capturing vertebral artery blood flow status. However, ultrasound Doppler imaging technology has not yet been integrated into the dynamic time series analysis and intelligent scoring framework in the current risk assessment process, and urgently needs further integration to enhance its clinical value. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a vertebral artery dissection risk assessment method to solve the problems of the existing methods in that the risk score is insufficiently adjusted in a personalized manner and ultrasound Doppler blood flow information is not effectively integrated.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a method for assessing the risk of vertebral artery dissection, comprising: Collect ultrasonic dynamic blood parameter data of the patient's vertebral artery; Based on the ultrasound dynamic blood parameter data, the hemodynamic time series analysis method is used to construct the local blood flow energy disturbance time series diagram, and the time series prediction is performed to generate the blood flow function disturbance prediction results; A reinforcement learning algorithm is used to dynamically adjust the risk score threshold. This adjusted risk score threshold is then combined with the blood flow function disturbance prediction results to generate a personalized vertebral artery dissection risk score based on the risk score calculation model. Combining personalized vertebral artery dissection risk score and multi-source heterogeneous medical feature set, a multimodal fusion mechanism is used to generate vertebral artery dissection risk assessment results.

[0007] As a preferred embodiment of the vertebral artery dissection risk assessment method of the present invention, the ultrasonic dynamic blood parameter data of the patient's vertebral artery includes the blood flow velocity, blood flow pressure and blood flow shear rate change values ​​of the patient's vertebral artery in multiple cardiac cycles.

[0008] As a preferred embodiment of the vertebral artery dissection risk assessment method of the present invention, the steps of using a hemodynamic time series analysis method to construct a local blood flow energy disturbance time series diagram, and performing time series prediction to generate a blood flow function disturbance prediction result are as follows: Based on ultrasound dynamic blood parameter data, a multi-parameter time series sampling and feature construction method is used to extract a multi-dimensional dynamic hemodynamic parameter set at continuous time points; The local blood kinetic energy value at each time point is calculated by the local blood kinetic energy density calculation method, and the blood kinetic energy perturbation sequence is constructed using the sliding mean perturbation method. The blood kinetic energy perturbation sequence is then organized into a local blood kinetic energy perturbation time series diagram in chronological order. Normalizing the local blood flow energy disturbance time series diagram, and substituting the processed local blood flow energy disturbance time series diagram into the time series prediction model to obtain a standardized disturbance prediction sequence; The standardized disturbance prediction sequence is restored to the actual kinetic energy disturbance value through the anti-standard method, and the time series prediction model is used to infer and generate the blood flow function disturbance prediction results in the future period.

[0009] As a preferred embodiment of the vertebral artery dissection risk assessment method of the present invention, the risk score threshold is dynamically adjusted using a reinforcement learning algorithm, and the adjusted risk score threshold is combined with the blood flow function disturbance prediction result to generate a personalized vertebral artery dissection risk score based on the risk score calculation model. The steps are as follows: Based on the prediction results of blood flow function disturbance, statistical features that characterize future disturbance trends are extracted and combined with the patient's static features to construct a risk state vector; Substitute the risk state vector into the reinforcement learning model and select the optimal risk score threshold under the current risk state from the set of discrete actions of risk score threshold; Based on the perturbation exceeding threshold ratio and time weight information in the risk score threshold, a risk score calculation model is used to generate a personalized vertebral artery dissection risk score.

[0010] As a preferred embodiment of the vertebral artery dissection risk assessment method of the present invention, the patient's static characteristics refer to medical characteristic information that does not change with time during the vertebral artery dissection risk assessment process and remains unchanged within the prediction period.

[0011] As a preferred embodiment of the vertebral artery dissection risk assessment method of the present invention, the steps of combining the personalized vertebral artery dissection risk score and the multi-source heterogeneous medical feature set and generating the vertebral artery dissection risk assessment result using the multimodal fusion mechanism are as follows: Based on the multi-source heterogeneous medical feature set, heterogeneous medical features are extracted and combined with the personalized vertebral artery dissection risk score to perform unified indexing and time alignment to construct a multimodal input vector. A multimodal fusion network with a dual-tower architecture projects the multimodal input vector into a shared feature space and uses an attention-weighted fusion strategy to generate a fused representation vector. A multi-layer perceptron risk assessment model was used to convert the fusion representation vector into a risk level probability distribution, and the probability of the vertebral artery dissection risk level was obtained through the Softmax function.

[0012] As a preferred solution of the vertebral artery dissection risk assessment method described in the present invention, the multimodal fusion network with a dual-tower architecture is a neural network structure for processing multimodal input, and feature extraction of data of different modalities is performed through two independent but structurally symmetrical network paths.

[0013] In a second aspect, the present invention provides a vertebral artery dissection risk assessment system, comprising: Data acquisition module, collecting ultrasonic dynamic blood parameter data of the patient's vertebral artery; The time series analysis module uses the hemodynamic time series analysis method to construct a local blood flow energy disturbance time series diagram based on ultrasonic dynamic blood parameter data, and performs time series prediction to generate blood flow function disturbance prediction results; The enhanced control module uses a reinforcement learning algorithm to dynamically adjust the risk score threshold, combines the adjusted risk score threshold with the blood flow function disturbance prediction results, and generates a personalized vertebral artery dissection risk score based on the risk score calculation model; The risk assessment module combines personalized vertebral artery dissection risk scores and multi-source heterogeneous medical feature sets, and uses a multimodal fusion mechanism to generate vertebral artery dissection risk assessment results.

[0014] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the vertebral artery dissection risk assessment method as described in the first aspect of the present invention is implemented.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the vertebral artery dissection risk assessment method as described in the first aspect of the present invention is implemented.

[0016] The beneficial effects of the present invention are: by adopting a reinforcement learning algorithm to dynamically adjust the risk score threshold, and combining the blood flow function disturbance prediction results to generate a personalized vertebral artery dissection risk score, it is possible to build a reinforcement learning model based on the expression of risk status, use the patient's static characteristics and future disturbance trends as input, dynamically select the optimal risk score threshold, and generate a personalized score based on the disturbance exceeding the threshold ratio and time weight information. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 Flowchart of the vertebral artery dissection risk assessment method.

[0019] Figure 2 Flowchart generated for the prediction of blood flow perturbations.

[0020] Figure 3 Flowchart generated for personalized vertebral artery dissection risk score.

[0021] Figure 4 Flowchart generated for the risk assessment of vertebral artery dissection. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0025] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a method for assessing the risk of vertebral artery dissection, comprising the following steps: S1: Collect ultrasonic dynamic blood parameter data of the patient's vertebral artery.

[0026] Specifically, ultrasonic dynamic blood parameter data of the patient's vertebral artery are obtained through hemodynamic monitoring equipment. As for the acquisition method, a color Doppler ultrasound device that penetrates the neck window is preferably used (including but not limited to: a portable color Doppler ultrasound instrument, a high-resolution carotid artery Doppler ultrasound diagnostic instrument, and a color Doppler ultrasound device with a cranial Doppler function, etc.), so as to capture continuous blood flow signals in a non-invasive manner. During the acquisition, the patient is selected to be in a resting state to ensure that a time-series blood flow parameter sequence of no less than one complete respiratory cycle is continuously recorded.

[0027] For example, the operating frequency of the vascular ultrasound instrument can be set to 10 MHz to meet the imaging requirements of the superficial vascular structure in the vertebral artery through the neck penetration window and improve the spatial resolution of color Doppler imaging. On this basis, the blood flow signals on the left and right sides of the vertebral artery are synchronously collected through the neck color Doppler ultrasound imaging technology. Combined with the high time sampling rate, high-precision characterization of the blood flow velocity fluctuation process is achieved. In addition, the left and right sides of the patient's vertebral artery are synchronously measured through the neck color Doppler ultrasound imaging technology to obtain high-time-resolution ultrasound dynamic blood parameter data.

[0028] Among them, ultrasonic dynamic blood parameter data specifically include real-time collection of blood flow velocity, blood flow pressure, blood flow shear rate change value of the patient's vertebral artery during multiple cardiac cycles, pulsation index reflecting the characteristics of hemodynamic changes, resistance index and cycle average flow and other parameters.

[0029] S2: Based on the ultrasonic dynamic blood parameter data, the hemodynamic time series analysis method is used to construct a local blood flow energy disturbance time series diagram, and time series prediction is performed to generate the blood flow function disturbance prediction results.

[0030] Specifically, the following steps are included: S2.1: Based on ultrasonic dynamic blood parameter data, a multi-parameter time series sampling and feature construction method is used to extract a set of multidimensional dynamic hemodynamic parameters at continuous time points. The local hemodynamic energy value at each time point is calculated by the local hemodynamic energy density calculation method, and the sliding mean perturbation method is used to construct a hemodynamic energy perturbation sequence. The hemodynamic energy perturbation sequence is then organized into a local hemodynamic energy perturbation time series diagram in chronological order.

[0031] Specifically, after collecting the ultrasonic dynamic blood parameter data of the patient's vertebral artery, a multidimensional dynamic hemodynamic parameter set of continuous time points is extracted based on the ultrasonic dynamic blood parameter data, and arranged in chronological order to form a dynamic hemodynamic parameter sequence. Subsequently, a local blood kinetic energy density calculation method is used to calculate the local blood kinetic energy value at the corresponding time point based on the dynamic hemodynamic parameter sequence corresponding to each time point. Finally, the blood kinetic energy disturbance value is calculated by the deviation between the local blood kinetic energy mean value within a fixed-length time window and the kinetic energy value at the current time point using a local mean deviation method based on a sliding window. The blood kinetic energy disturbance value of each time point is arranged in chronological order to construct a blood kinetic energy disturbance sequence, and the constructed blood kinetic energy disturbance sequence is organized in chronological order to generate a local blood kinetic energy disturbance time series diagram.

[0032] It should be explained that the multidimensional dynamic hemodynamic parameter set includes parameters such as local blood flow velocity, shear stress, blood viscosity, and pressure changes sampled once per second.

[0033] Among them, organizing the blood flow energy disturbance sequence in chronological order is to linearly arrange and structure the blood flow energy disturbance values ​​corresponding to the blood flow energy disturbance sequence at each time point in chronological order to form an orderly local blood flow energy disturbance time series diagram.

[0034] Among them, the sliding mean perturbation method is a time series fluctuation analysis method based on a local time window. The principle is based on the degree of deviation between the local blood flow energy value corresponding to the current time point and the mean value within a time window. It effectively reflects the non-stationary fluctuations in dynamic blood flow parameters that occur in a short period of time. For example, when the disturbance value at a certain time point is significantly higher than that at other time points, it may indicate that there is a mutation or potential risk change in the local blood flow corresponding to that point.

[0035] Among them, the calculation formula of the local mean deviation method is as follows: ; in, Representation and time point The corresponding blood flow energy disturbance value, represents the blood density, Indicates at a point in time The corresponding local blood flow velocity.

[0036] S2.2: normalize the local blood flow energy disturbance time series diagram, and substitute the processed local blood flow energy disturbance time series diagram into the time series prediction model to obtain a standardized disturbance prediction sequence.

[0037] Specifically, the minimum-maximum normalization method is used to normalize the local blood flow energy disturbance time series diagram, and the disturbance value at each time point is converted into the interval [0, 1]. Then, the normalized local blood flow energy disturbance time series diagram is substituted into the time series prediction model to predict the disturbance trend at future moments and obtain the standardized disturbance prediction sequence.

[0038] It should be explained that the time series prediction model is constructed based on the temporal dependence of the local blood flow energy perturbation time series graph. It is modeled using a long-short-term memory network with a memory mechanism. During the modeling process, the normalized local blood flow energy perturbation time series graph is substituted in chronological order, and a fixed-length sliding window is set. The local blood flow energy perturbation time series graph is then used as supervised training data, and each blood flow energy perturbation data in the normalized local blood flow energy perturbation time series graph within the previous time window is used as input. The model's internal parameters are iteratively updated by minimizing the error between the predicted result and the true perturbation value. An early stopping mechanism is used during training to monitor the error changes on the validation set. Once training is complete, the time series prediction model can predict the new normalized local blood flow energy perturbation time series graph, obtaining the corresponding standardized perturbation prediction sequence.

[0039] Among them, the early stopping mechanism is a training strategy used to prevent overfitting of time series prediction models during training. The basic principle is to continuously monitor the changes in prediction error on the validation set during training. When the validation error no longer decreases in several consecutive rounds of training, it is considered that the model has reached the optimal learning state and the training process is terminated early.

[0040] S2.3: The standardized disturbance prediction sequence is restored to the actual kinetic energy disturbance value through the anti-standard method, and the time series prediction model is used to infer and generate the blood flow function disturbance prediction results in the future period.

[0041] Specifically, the anti-standard method is used to restore the standardized disturbance prediction sequence to the actual blood flow energy disturbance value. Then, a time series prediction model is adopted, with the normalized local blood flow energy disturbance time series diagram as input data, to infer the blood flow function disturbance prediction results in the future period based on the time change characteristics.

[0042] It's important to explain that the denormalization method involves numerically restoring the hemodynamic energy perturbation values ​​after minimum-maximum normalization. Essentially, the normalized perturbation prediction sequence is converted back to the original perturbation value according to the corresponding proportions, given the known minimum and maximum values ​​used during normalization. For example, given a known minimum of 0.2 and a known maximum of 1.8, each predicted normalized perturbation value is proportionally mapped back to the interval [0.2, 1.8] based on its position within the interval [0, 1], yielding the corresponding actual hemodynamic energy perturbation value.

[0043] The blood flow function disturbance prediction result refers to the blood flow energy disturbance value corresponding to each future time point within the prediction time window.

[0044] S3: A reinforcement learning algorithm is used to dynamically adjust the risk score threshold, and the adjusted risk score threshold is combined with the blood flow function disturbance prediction results to generate a personalized vertebral artery dissection risk score based on the risk score calculation model.

[0045] Specifically, the following steps are included: S3.1: Define the risk score threshold. Based on the blood flow function perturbation prediction results, use the perturbation trend statistical coding method to extract the statistical features that characterize the future perturbation trend, and splice them with the patient's static features to construct the risk state vector.

[0046] Specifically, based on the clinical risk judgment criteria and the statistical distribution of historical cases, an initial risk score threshold interval is set, and the risk score threshold interval is discretized as the risk score threshold. Based on the blood flow function disturbance prediction results, the disturbance data sequence is extracted within a predefined time window. Subsequently, the disturbance data sequence is analyzed using the disturbance trend statistical coding method to obtain a set of statistical features that quantitatively represent future disturbance trends. The data feature sequence is arranged in a fixed order to obtain a data feature sequence. Finally, the data feature sequence is spliced ​​and integrated with the patient's static feature vector to form a risk state vector with a unified structure and fixed dimension.

[0047] It should be explained that the risk score threshold is set based on previous clinical risk judgment standards and historical case statistical distribution, and the risk level boundary value is formed through discretization. Subsequently, based on the real-time blood flow function disturbance prediction results and individual patient characteristics, the risk score threshold is dynamically adjusted through a dynamic adjustment method of strategy optimization.

[0048] The real-time blood flow function disturbance prediction results are numerical indicators that reflect the changing characteristics of the blood flow energy disturbance trend within the prediction time, including mean, maximum, minimum, or standard deviation. For example, if the blood flow function disturbance prediction results collected in real time at five moments are 1.2, 1.3, 1.1, 1.5, and 1.4, the mean is 1.3, the maximum is 1.5, the minimum is 1.1, and the standard deviation is approximately 0.15.

[0049] Static patient characteristics refer to medical characteristics that do not change over time during the vertebral artery dissection risk assessment process and remain constant over the prediction period. These include basic demographic characteristics (e.g., patient age and gender), basic clinical measurements (e.g., systolic blood pressure, diastolic blood pressure, and resting heart rate), chronic diseases, and medical history information.

[0050] S3.2: Substitute the risk state vector into the reinforcement learning model and select the optimal risk score threshold under the current risk state from the set risk score threshold discrete action set.

[0051] Specifically, the current risk state vector is substituted into the reinforcement learning model, and the pre-set risk score threshold discrete action set is used as the optional action space. The risk score threshold in the risk score threshold discrete action set is strategically evaluated to obtain the strategy score corresponding to the pre-set risk score threshold discrete action set, and the threshold with the highest strategy score is selected as the risk score threshold under the current risk state.

[0052] It should be explained that the reinforcement learning model is constructed and trained based on a risk scoring simulation environment. This environment uses a discrete set of actions with risk score thresholds as the optional set of actions, and defines a reward function based on the consistency between the risk score results and historically accurate annotations. The reinforcement learning model is then trained using a value-optimization-based reinforcement learning method. In each round of interaction, the risk state vector is sampled, a risk score threshold is selected as the action, and the reinforcement learning model's policy function is continuously updated based on the sampled data. Ultimately, after multiple rounds of training, when the cumulative rewards and policy converge, the trained reinforcement learning model is obtained.

[0053] The risk score threshold discrete action set is based on a reasonable division of the risk score threshold value range and precision. Specifically, determine the risk score threshold value range (for example, set it to [0.1, 0.9]), and divide the range into discrete value points at equal intervals based on the threshold precision to obtain the risk score threshold discrete action set. For example, if the risk score threshold value range is [0.1, 0.9], and the steps are 0.1, the risk score threshold discrete action set is {0.1, 0.2, 0.3, ..., 0.9}.

[0054] S3.3: Based on the perturbation exceeding threshold ratio and time weight information in the risk score threshold, a personalized vertebral artery dissection risk score is generated using the risk score calculation model.

[0055] Specifically, based on the perturbation exceeding threshold ratio and time weight information in the risk score threshold, the number of times the blood flow function perturbation prediction results exceeded the risk score threshold within the future prediction time window was counted, and the perturbation exceeding threshold ratio was calculated. For example, if the blood flow energy perturbation value exceeded the risk score threshold at 6 out of a total of 20 time points, the perturbation exceeding threshold ratio would be 0.3. Combined with the time weight information corresponding to each time point, the perturbation exceeding threshold ratio and time weight information were substituted into the risk score calculation model to obtain a personalized vertebral artery dissection risk score.

[0056] It should be noted that the risk score calculation model is based on a training sample set constructed based on historical annotated vertebral artery dissection case data. A multi-feature regression modeling method is used, and features such as the multidimensional dynamic blood flow disturbance sequence, the proportion of disturbance exceeding the threshold, and the time weighting factor in the vertebral artery dissection case data are used as input features. The historical real risk score label is used as the supervision signal for model training. Among them, the risk score calculation model is constructed based on the multi-layer perceptron-based recurrent neural network architecture. During the training process, the mean square error is used as the loss function, and the generalization ability and stability of the risk score calculation model are evaluated through k-fold cross-validation. The performance of the risk score calculation model is quantitatively compared by combining indicators such as mean square error, determination coefficient, and mean absolute error. The risk score calculation model is optimized by adjusting hyperparameters (such as learning rate and regularization term, etc.), and finally an optimized risk score calculation model is generated.

[0057] Among them, the time weight information refers to the contribution weight of the disturbance value at different time points to the personalized vertebral artery dissection risk score, specifically including: time attenuation factor, cardiac cycle phase weight, and pathological event association weight.

[0058] S4: Combine personalized vertebral artery dissection risk score and multi-source heterogeneous medical feature set to generate vertebral artery dissection risk assessment results using multimodal fusion mechanism.

[0059] Specifically, the steps are as follows: S4.1: Extract heterogeneous medical features based on a multi-source heterogeneous medical feature set, combine the heterogeneous medical features with a personalized vertebral artery dissection risk score, perform unified indexing and time alignment to construct a multimodal input vector.

[0060] Specifically, heterogeneous medical features are extracted based on a multi-source heterogeneous medical feature set, and the heterogeneous medical features and the personalized vertebral artery dissection risk score are aligned according to a unified time index. The features of different modalities are mapped to a unified time scale based on the index timestamp. The heterogeneous medical features and the personalized vertebral artery dissection risk score are spliced ​​at the same time node to construct a time-consistent multimodal input vector.

[0061] It should be noted that heterogeneous medical features include imaging features, physiological monitoring features, test index features, medication record features, and medical history information features, etc.

[0062] S4.2: A multimodal fusion network with a dual-tower architecture projects the multimodal input vector into a shared feature space and uses an attention-weighted fusion strategy to generate a fused representation vector.

[0063] Specifically, a dual-tower multimodal fusion network extracts features from a multimodal input vector containing heterogeneous medical features and a personalized vertebral artery dissection risk score. Two symmetrical deep neural network pathways are constructed. The first pathway is designed as a feedforward neural network consisting of a stack of fully connected layers for the heterogeneous medical features and preprocessed using corresponding feature embedding or normalization methods. The second pathway uses the personalized vertebral artery dissection risk score generated based on blood flow function perturbation as input and processes it using a feedforward network similar to the first pathway to extract heterogeneous medical features representing the dynamic blood flow function risk state. The heterogeneous medical feature representations and the personalized vertebral artery dissection risk score representation are then projected into a shared feature space. An attention-weighted fusion strategy is then employed to form a fused representation vector based on the importance distribution.

[0064] The dual-tower multimodal fusion network is a neural network structure designed to process multimodal inputs. It extracts features from data of different modalities through two independent but structurally symmetrical network pathways. The first tower takes heterogeneous medical features as input, such as age, gender, smoking history, underlying disease status, previous cerebrovascular event records, and genomic variation information. This pathway is used to extract high-level semantic representations that reflect the patient's static individual characteristics and basic medical background. The second tower takes a personalized vertebral artery dissection risk score as input, including a score generated based on the prediction of blood flow functional perturbations. This pathway extracts feature representations related to dynamic functional risk factors.

[0065] S4.3: Using the multi-layer perceptron risk assessment model, the fusion representation vector is converted into a risk level probability distribution, and the probability of the vertebral artery dissection risk level is obtained through the Softmax function.

[0066] Specifically, a multi-layer perceptron risk assessment model was adopted, and the fusion representation vector was used as the input feature vector. Risk score results consistent with the set number of vertebral artery dissection risk levels were obtained through multiple nonlinear mapping layers. Each value of the risk score result was converted into a probability distribution of vertebral artery dissection risk level between 0 and 1 through the Softmax function.

[0067] It should be explained that the multi-layer perceptron risk assessment model is based on a training sample set constructed from a fused representation vector and the corresponding labeled vertebral artery dissection risk level data. The fused representation vector is used as the input feature vector, and the vertebral artery dissection risk level label is used as the supervisory signal. A multi-layer fully connected neural network with a ReLU activation function is used to perform a layer-by-layer nonlinear transformation on the input feature vector to extract high-order semantic features. Subsequently, a Softmax function is used in the output layer to map the output of the last layer of neurons to probability values ​​corresponding to multiple risk levels, forming a risk level probability distribution. The cross-entropy loss function is used as the optimization target, and the model parameters are continuously updated through backpropagation and gradient descent methods to ultimately obtain the multi-layer perceptron risk assessment model.

[0068] Among them, the backpropagation method calculates the error between the risk score result of the multi-layer perceptron risk assessment model and the actual risk score result, and uses the chain rule to transfer the error layer by layer to obtain the gradient information of the parameters of each layer; the gradient descent method iteratively updates the parameters of the multi-layer perceptron risk assessment model according to the calculated objective function gradient and the set learning rate.

[0069] The formula for the gradient descent method is as follows, ; in, Indicates the The objective function gradient at the iteration, Indicates the The objective function gradient at the iteration, represents the learning rate, Indicates the parameters The operation of finding the gradient, Indicates in At the iteration, the parameter value is The loss function value under .

[0070] This embodiment further provides a vertebral artery dissection risk assessment system, comprising: Data acquisition module, collecting ultrasonic dynamic blood parameter data of the patient's vertebral artery; The time series analysis module uses the hemodynamic time series analysis method to construct a local blood flow energy disturbance time series diagram based on ultrasonic dynamic blood parameter data, and performs time series prediction to generate blood flow function disturbance prediction results; The enhanced control module uses a reinforcement learning algorithm to dynamically adjust the risk score threshold, combines the adjusted risk score threshold with the blood flow function disturbance prediction results, and generates a personalized vertebral artery dissection risk score based on the risk score calculation model; The risk assessment module combines personalized vertebral artery dissection risk scores and multi-source heterogeneous medical feature sets, and uses a multimodal fusion mechanism to generate vertebral artery dissection risk assessment results.

[0071] This embodiment also provides a computer device suitable for the vertebral artery dissection risk assessment method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the vertebral artery dissection risk assessment method proposed in the above embodiment.

[0072] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0073] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the vertebral artery dissection risk assessment method proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0074] In summary, the present invention achieves the following steps: dynamically adjusting the risk score threshold using a reinforcement learning algorithm, and generating a personalized vertebral artery dissection risk score based on the blood flow function disturbance prediction results. This achieves the construction of a reinforcement learning model based on the expression of risk status, using the patient's static characteristics and future disturbance trends as input, dynamically selecting the optimal risk score threshold, and generating a personalized score based on the disturbance exceeding the threshold ratio and time weight information.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for assessing the risk of vertebral artery dissection, characterized by: include, Collect ultrasonic dynamic blood parameter data of the patient's vertebral artery; Based on the ultrasound dynamic blood parameter data, the hemodynamic time series analysis method is used to construct the local blood flow energy disturbance time series diagram, and the time series prediction is performed to generate the blood flow function disturbance prediction results; A reinforcement learning algorithm is used to dynamically adjust the risk score threshold. This adjusted risk score threshold is then combined with the blood flow function disturbance prediction results to generate a personalized vertebral artery dissection risk score based on the risk score calculation model. Combining personalized vertebral artery dissection risk score and multi-source heterogeneous medical feature set, a multimodal fusion mechanism is used to generate vertebral artery dissection risk assessment results.

2. The vertebral artery dissection risk assessment method according to claim 1, wherein: The ultrasonic dynamic blood parameter data of the patient's vertebral artery includes the blood flow velocity, blood flow pressure and blood flow shear rate change values ​​of the patient's vertebral artery during multiple cardiac cycles.

3. The vertebral artery dissection risk assessment method according to claim 1, wherein: The hemodynamic time series analysis method is used to construct a local blood flow energy disturbance time series diagram, and time series prediction is performed to generate a blood flow function disturbance prediction result. The steps are as follows: Based on ultrasound dynamic blood parameter data, a multi-parameter time series sampling and feature construction method is used to extract a multi-dimensional dynamic hemodynamic parameter set at continuous time points; The local blood kinetic energy value at each time point is calculated by the local blood kinetic energy density calculation method, and the blood kinetic energy perturbation sequence is constructed using the sliding mean perturbation method. The blood kinetic energy perturbation sequence is then organized into a local blood kinetic energy perturbation time series diagram in chronological order. Normalizing the local blood flow energy disturbance time series diagram, and substituting the processed local blood flow energy disturbance time series diagram into the time series prediction model to obtain a standardized disturbance prediction sequence; The standardized disturbance prediction sequence is restored to the actual kinetic energy disturbance value through the anti-standard method, and the time series prediction model is used to infer and generate the blood flow function disturbance prediction results in the future period.

4. The vertebral artery dissection risk assessment method according to claim 1, wherein: The reinforcement learning algorithm is used to dynamically adjust the risk score threshold, and the adjusted risk score threshold is combined with the blood flow function disturbance prediction result. According to the risk score calculation model, a personalized vertebral artery dissection risk score is generated. The steps are as follows: Based on the prediction results of blood flow function disturbance, statistical features that characterize future disturbance trends are extracted and combined with the patient's static features to construct a risk state vector; Substitute the risk state vector into the reinforcement learning model and select the optimal risk score threshold under the current risk state from the set of discrete actions of risk score threshold; Based on the perturbation exceeding threshold ratio and time weight information in the risk score threshold, a risk score calculation model is used to generate a personalized vertebral artery dissection risk score.

5. The vertebral artery dissection risk assessment method according to claim 4, wherein: The patient static characteristics refer to medical characteristic information that does not change over time during the vertebral artery dissection risk assessment process and remains unchanged within the prediction period.

6. The vertebral artery dissection risk assessment method according to claim 1, wherein: The personalized vertebral artery dissection risk score and the multi-source heterogeneous medical feature set are combined to generate the vertebral artery dissection risk assessment result using a multimodal fusion mechanism. The steps are as follows: Based on the multi-source heterogeneous medical feature set, heterogeneous medical features are extracted and combined with the personalized vertebral artery dissection risk score to perform unified indexing and time alignment to construct a multimodal input vector. A multimodal fusion network with a dual-tower architecture projects the multimodal input vector into a shared feature space and uses an attention-weighted fusion strategy to generate a fused representation vector. A multi-layer perceptron risk assessment model was used to convert the fusion representation vector into a risk level probability distribution, and the probability of the vertebral artery dissection risk level was obtained through the Softmax function.

7. The vertebral artery dissection risk assessment method according to claim 6, wherein: The multimodal fusion network with a dual-tower architecture is a neural network structure for processing multimodal inputs, and features of data of different modalities are extracted respectively through two independent but structurally symmetrical network paths.

8. A vertebral artery dissection risk assessment system, based on the vertebral artery dissection risk assessment method according to any one of claims 1 to 7, characterized in that: include, Data acquisition module, collecting ultrasonic dynamic blood parameter data of the patient's vertebral artery; The time series analysis module uses the hemodynamic time series analysis method to construct a local blood flow energy disturbance time series diagram based on ultrasonic dynamic blood parameter data, and performs time series prediction to generate blood flow function disturbance prediction results; The enhanced control module uses a reinforcement learning algorithm to dynamically adjust the risk score threshold, combines the adjusted risk score threshold with the blood flow function disturbance prediction results, and generates a personalized vertebral artery dissection risk score based on the risk score calculation model; The risk assessment module combines personalized vertebral artery dissection risk scores and multi-source heterogeneous medical feature sets, and uses a multimodal fusion mechanism to generate vertebral artery dissection risk assessment results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the vertebral artery dissection risk assessment method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the vertebral artery dissection risk assessment method according to any one of claims 1 to 7 are implemented.

Images