Cerebral apoplexy onset risk assessment and reminding method and cerebral apoplexy onset risk assessment and reminding system

Abstract

The invention relates to the technical field of intelligent medical systems, and discloses a cerebral apoplexy onset risk assessment and reminding method and system.The method comprises the steps that continuous medical structured detection data are collected, and the data comprise carotid artery blood flow parameters, brain oxygen saturation, heart rate variability and metabolic indexes; inputting a bidirectional LSTM, a differential convolutional network, a wavelet residual network and a multi-layer perceptron to extract nonlinear features; constructing a neural function coupling structure diagram of four nodes of cerebral blood supply, oxygen supply, autonomous regulation and metabolic steady state; calculating inter-node time sequence offset correlation and a stable factor to obtain a coupling anomaly coefficient; and driving the embedded network by using a graph structure and a node feature input mechanism, and outputting a risk state assessment result. According to the method, the neural function coupling structure diagram is constructed and the mechanism is introduced to drive the embedded network, so that high-precision identification of the multi-system collaborative abnormal state and dynamic evaluation of the stroke risk level are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical system technology, and more specifically, to a method and system for assessing and alerting the risk of stroke. Background Technology

[0002] Clinical studies have shown that stroke is usually caused by the synergistic abnormalities of multiple physiological indicators, especially in the hours to day before onset, often accompanied by abnormal cerebral blood supply, autonomic dysfunction, decreased cerebral oxygen supply, and disruption of metabolic homeostasis. However, traditional risk assessment methods often rely on single indicators (such as blood pressure, heart rate, or blood glucose) or static medical images for judgment, which cannot fully capture the dynamic synergistic relationships between different functional systems and their evolutionary processes. This results in problems such as low assessment accuracy, poor early warning timeliness, and inability to adapt to individual differences.

[0003] Furthermore, while some studies have attempted to introduce neural network models to analyze medical test data, they often neglect the functional coupling mechanisms between different physiological systems, modeling each indicator independently. This makes it difficult to form a highly interpretable and clearly structured risk identification path. Moreover, the lack of quantitative analysis of abnormal correlations between indicators makes it difficult to determine whether synergistic changes reach clinical intervention thresholds, affecting the reliability of early warnings. Simultaneously, most existing methods do not incorporate graph structure modeling mechanisms for risk state identification, lacking the ability to map multi-source time-series data into structured neural functional atlases, making it difficult to support multi-level comprehensive judgment and grade assessment of stroke risk.

[0004] Therefore, there is an urgent need to develop a stroke risk assessment method that integrates multiple sets of physiological indicators, has the ability to model neural functional systems, can capture abnormal coupling between systems, and classify risk status levels, so as to improve the early identification and intervention efficiency of high-risk stroke populations. Summary of the Invention

[0005] In view of this, the present invention proposes a method and system for assessing and alerting the risk of stroke, in order to solve the above problems.

[0006] On the one hand, the present invention proposes a method for assessing and alerting the risk of stroke, comprising:

[0007] Collect continuous medical structured test data of stroke risk prediction subjects. The medical structured test data includes carotid peak blood flow velocity, pulsatility index, resistance index, cerebral oxygen saturation, heart rate variability parameters, serum lactate concentration, arterial carbon dioxide partial pressure and blood pH value.

[0008] Feature extraction is performed on the medical structured detection data, including: inputting carotid artery peak blood flow velocity, pulsatility index, and resistance index into a first bidirectional long short-term memory network; inputting cerebral oxygen saturation into a convolutional neural network structure containing differential operations; inputting heart rate variability parameters into a residual neural network structure based on wavelet transform; inputting serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH into a multilayer perceptron structure; outputting corresponding nonlinear mapping features respectively; and constructing a neural functional coupling structure diagram of the stroke risk prediction object based on the nonlinear mapping features. The neural functional coupling structure diagram includes multiple nodes and multiple edges. Nodes represent the cerebral blood supply system, cerebral oxygen supply system, autonomic nervous regulation system, and metabolic homeostasis system, and edges represent the intensity of synergistic abnormalities between different nodes.

[0009] Calculate the coupling anomaly coefficient between each node in the neural functional coupling structure diagram. The coupling anomaly coefficient is obtained by multiplying the Pearson correlation coefficient between the time series offset vectors of the corresponding index groups of adjacent nodes for a preset duration by a collaborative reference stability factor.

[0010] The adjacency matrix of the neural functional coupling structure graph and the structural feature vector of each node are input into the mechanism-driven embedding network for joint modeling. The mechanism-driven embedding network includes a structure nested graph encoding layer, a side-gated linkage propagation layer and a graph global response convolutional layer.

[0011] Based on the output of the embedded network driven by the aforementioned mechanism, it is determined whether the stroke risk prediction object is currently in a stroke risk state, and the assessed risk state is output.

[0012] Furthermore, when inputting the carotid peak blood flow velocity, pulsatility index, and resistance index into the first bidirectional long short-term memory network, the following parameters are included:

[0013] The peak blood flow velocity, pulsatility index and resistance index of the carotid artery collected at fixed time intervals during a continuous monitoring period for the stroke risk prediction object are respectively constructed into a three-dimensional input vector sequence in chronological order;

[0014] The three-dimensional input vector sequence is input into the first bidirectional long short-term memory network, which is composed of forward LSTM units and backward LSTM units connected in parallel. Information extraction is performed in the forward time flow and the backward time flow respectively to obtain the forward hidden state vector and the backward hidden state vector corresponding to each time point.

[0015] The forward hidden state vector and the backward hidden state vector are concatenated to obtain the bidirectional context joint feature sequence;

[0016] The bidirectional context joint feature sequence is processed based on a sliding window. The mean value of the time period, the maximum variation rate and the local mutation gradient within each window are extracted and jointly encoded with the mean value of the original input index of the corresponding time period to construct the time period feature vector of the cerebral blood supply system.

[0017] Arrange all time-segment feature vectors in chronological order to obtain a sequence of structural feature representations for nodes in the brain blood supply system in the neural functional coupling structure diagram.

[0018] The average value of the original input index is the average value of all sampling points calculated for the peak blood flow velocity, pulsatility index and resistance index of the carotid artery within each sliding window time period.

[0019] Furthermore, brain oxygen saturation is input into a convolutional neural network structure that includes differential operations, including:

[0020] Brain oxygen saturation data are constructed into a one-dimensional time series; a first-order difference operation is performed on the brain oxygen saturation time series to generate a difference series representing the change amplitude of adjacent sampling points;

[0021] The one-dimensional time series and the difference series are concatenated along the channel dimension to obtain a two-dimensional input matrix;

[0022] The two-dimensional input matrix is ​​sequentially input into a first one-dimensional convolutional layer, a second one-dimensional convolutional layer, and a third one-dimensional convolutional layer. The first one-dimensional convolutional layer has multiple convolutional kernels of length 3, the second one-dimensional convolutional layer has multiple convolutional kernels of length 5, and the third one-dimensional convolutional layer has multiple convolutional kernels of length 7. The output of each convolutional layer is processed by batch normalization and ReLU activation function and then passed to the next convolutional layer. After the output of the third one-dimensional convolutional layer, a one-dimensional global average pooling layer is connected to output a feature vector of fixed dimension.

[0023] The aforementioned feature vector is used as the structural feature vector of the brain oxygen supply system in the neural functional coupling structure diagram.

[0024] Furthermore, the heart rate variability parameters are input into a wavelet transform-based residual neural network structure, including:

[0025] Heart rate variability parameters were constructed into time series data, and wavelet multi-scale decomposition was performed to extract two high-frequency sub-band signals, which were used as the first sub-band sequence and the second sub-band sequence, respectively.

[0026] Each sub-band sequence is divided into time sliding windows of a preset length. Within each time sliding window, the absolute value sequence of the difference between two adjacent sampling points is calculated to construct several local fluctuation amplitude sequences.

[0027] For each local fluctuation amplitude sequence, calculate the mean and standard deviation. If the mean is greater than the mean of the corresponding sub-band sequence in the complete monitoring period plus twice the standard deviation, then the current time sliding window is determined to be the heart rate mutation window.

[0028] The data from all heart rate mutation windows are combined into a two-dimensional tensor in chronological order and then input into a one-dimensional convolutional network containing multiple residual structures for feature extraction to obtain the residual feature tensor.

[0029] The first and second sub-band sequences within the complete monitoring period are respectively input into a one-dimensional convolutional network with multiple kernel sizes to extract periodic variation features, and then concatenated after average pooling to form a global high-frequency sub-band feature vector.

[0030] The residual feature tensor is concatenated with the global high-frequency subband feature vector to form a fused feature tensor. The fused feature tensor is then subjected to max pooling and vector compression to obtain the structural feature vector representing the autonomic nervous system.

[0031] Furthermore, serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH are input into the multilayer sensor structure, including:

[0032] Serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH were constructed into a three-dimensional time series vector; the three-dimensional time series vector was normalized and divided into several time segments using an overlapping sliding window method;

[0033] In each time segment, the covariance amplitude between lactate concentration and arterial carbon dioxide partial pressure, the correlation direction coefficient between serum lactate concentration and blood pH, and the range index of blood pH change rate are calculated to construct a metabolic synergistic state characteristic group.

[0034] The metabolic co-state feature set is input into a multilayer perceptron structure containing three hidden layers. Each hidden layer of the multilayer perceptron structure has 32, 16 and 8 nodes respectively. The ReLU activation function is used to output a mapping feature with consistent dimension.

[0035] The perceptron outputs for all time segments are averaged over time to obtain the full-cycle state expression vector of the metabolic homeostasis system; this state expression vector is then used as the structural feature vector of the metabolic homeostasis system in the neural functional coupling structure diagram.

[0036] Further, the steps for obtaining the covariance amplitude are as follows: within a fixed-length time sliding window, continuous sampling values ​​of lactic acid concentration and carbon dioxide partial pressure are extracted respectively, the mean values ​​of the two sets of sampling values ​​within the current window are calculated respectively, the difference between each sampling point value and the corresponding mean value is calculated one by one, and the average of the corresponding positions of the two sets of differences is calculated to obtain the covariance amplitude within the current time window.

[0037] The relevant direction coefficients are obtained by calculating cosine similarity.

[0038] The range index is the difference between the maximum and minimum values ​​within a time sliding window.

[0039] Furthermore, obtaining the coupling anomaly coefficient includes:

[0040] The structural feature vector sequences collected at fixed time intervals within a preset time period from the coupled node pairs are extracted into index time series groups; the coupled node pairs are pairwise combinations of four nodes.

[0041] Perform first-order differencing on each set of indicator time series to construct the corresponding offset vector;

[0042] Two offset vectors are selected to calculate the Pearson correlation coefficient, which serves as the correlation trend factor for the coupled node pair in the current time period;

[0043] Take the historical Pearson correlation coefficient and obtain the corresponding correlation trend factor. Take the reciprocal of the variance of the corresponding correlation trend factor as the stability factor.

[0044] The coupling anomaly coefficient is obtained by multiplying the relevant trend factor and the stability factor for the current time period.

[0045] If the coupling anomaly coefficient is lower than the mean of the same coupled node pair over consecutive periods minus two standard deviations, then the node pair is marked as having a coupling anomaly.

[0046] Furthermore, determining whether the individual subject to stroke risk prediction is currently at risk of stroke includes:

[0047] The embedding vectors of each functional system node in the mechanism-driven embedded network output are concatenated to construct a comprehensive state representation tensor.

[0048] The discrimination index is calculated based on the comprehensive state representation tensor. The discrimination index includes global similarity offset, local edge weight anomaly density and clustering consistency loss value between nodes.

[0049] The global similarity offset is the cosine distance between the current integrated state tensor and the integrated state representation tensor of the normal sample;

[0050] Local edge weight anomaly density is the percentage of edges with edge weights greater than a set upper limit among all node pairs;

[0051] The clustering consistency loss between nodes is the ratio of the cohesion loss of the current embedded vector in the preset functional cluster to the cross-cluster distance.

[0052] A risk status score is generated based on the discrimination index. If the score is greater than the high-risk threshold, the patient is determined to be in a high-risk state for stroke. If the risk status score is less than or equal to the high-risk threshold but greater than the medium-risk threshold, the patient is determined to be in a medium-risk state. If the risk status score is less than or equal to the medium-risk threshold, the patient is determined to be in a low-risk state.

[0053] Furthermore, it also includes: once the risk status is determined, mapping the risk assessment level to corresponding prompts and color levels to remind the user.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] By constructing a neural functional coupling structure diagram, this study effectively addresses the problems of isolated analysis of physiological indicators and neglect of inter-system synergistic effects in existing stroke risk assessment methods. This structure diagram unifies and integrates four core physiological subsystems: the cerebral blood supply system, the cerebral oxygen supply system, the autonomic nervous regulation system, and the metabolic homeostasis system. Nodes represent system states, and edges represent the strength of synergistic relationships, realizing a cross-system linkage modeling mechanism. It establishes functional connections between physiological indicators at the structural level, which helps identify hidden or progressive risk characteristics.

[0056] This method introduces a multi-path structure in the feature extraction stage: it utilizes a bidirectional long short-term memory network to process hemodynamic parameters and capture the temporal context features of blood supply; it employs a differential convolutional network to perceive changes in cerebral oxygen saturation, highlighting subtle dynamic responses; it combines wavelet multi-scale decomposition and residual convolutional structures to identify local fluctuations and high-frequency anomalies in heart rate variability; and it introduces the dynamic characterization of metabolic regulatory states by constructing feature groups of covariance amplitude, correlation direction, and range indicators among metabolic parameters. Deep modeling of multiple pathways not only enhances the structural representation of various data types but also strengthens the sensitivity to identifying nonlinear mutations and periodic anomalies.

[0057] To address the functional coupling characteristics, this invention further proposes a method for calculating the coupling anomaly coefficient. By weighting the correlation coefficient of the offset vectors of adjacent functional nodes with the stability factor, it accurately measures the cooperative state changes between different functional systems, thereby improving the ability to identify cross-system anomaly propagation paths. Subsequently, a mechanism-driven embedded network is used to jointly model the coupled structure. High-order graph features are constructed using modules such as nested graph encoding, edge-gated propagation, and global response convolution, significantly enhancing the system's ability to express responses to structural changes.

[0058] In the risk status determination process, this method comprehensively evaluates multiple indicators such as the similarity shift between the comprehensive state tensor and normal samples, the abnormal density of coupling edge weights, and the consistency of node clustering, forming a multi-dimensional risk status scoring mechanism. This mechanism can finely distinguish between high-risk, medium-risk, and low-risk levels, avoiding false alarms caused by single parameters or static thresholds in traditional methods. Furthermore, the final evaluation results can be mapped to multi-level warning information with color prompts and text descriptions, realizing a visual closed-loop feedback mechanism for risk identification and user interaction.

[0059] On the other hand, the present invention proposes a system for assessing and alerting the risk of stroke, comprising:

[0060] The acquisition module is configured to collect continuous medical structured test data from stroke risk prediction subjects;

[0061] The feature module is configured to extract features from the medical structured detection data, including: inputting carotid peak blood flow velocity, pulsatility index, and resistance index into a first bidirectional long short-term memory network; inputting brain oxygen saturation into a convolutional neural network structure containing differential operations; inputting heart rate variability parameters into a residual neural network structure based on wavelet transform; inputting serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH into a multilayer perceptron structure; outputting corresponding nonlinear mapping features respectively; and constructing a neural functional coupling structure diagram of the stroke risk prediction object based on the nonlinear mapping features.

[0062] The state assessment module is configured to calculate the coupling anomaly coefficient between nodes in the neural functional coupling structure graph; input the adjacency matrix of the neural functional coupling structure graph and the structural feature vector of each node into the mechanism-driven embedding network for joint modeling; and, based on the output of the mechanism-driven embedding network, determine whether the stroke risk prediction object is currently in a stroke risk state and output the assessed risk state.

[0063] It should be noted that the method for assessing and reminding about the risk of stroke in this invention has the same beneficial effects as its system, and will not be described in detail here. Attached Figure Description

[0064] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0065] Figure 1 This is a flowchart of a method for assessing and alerting the risk of stroke, provided as an embodiment of the present invention.

[0066] Figure 2 This is a functional block diagram of a stroke risk assessment and reminder system provided in an embodiment of the present invention. Detailed Implementation

[0067] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0068] See Figure 1 As shown, this embodiment of the invention provides a method for assessing and alerting the risk of stroke, including:

[0069] S1: Collect continuous medical structured test data of stroke risk prediction subjects. The medical structured test data includes carotid peak blood flow velocity, pulsatility index, resistance index, cerebral oxygen saturation, heart rate variability parameters, serum lactate concentration, arterial carbon dioxide partial pressure and blood pH value.

[0070] S2: Feature extraction is performed on medical structured test data, including: inputting carotid peak blood flow velocity, pulsatility index, and resistance index into a first bidirectional long short-term memory network; inputting cerebral oxygen saturation into a convolutional neural network structure containing differential operations; inputting heart rate variability parameters into a residual neural network structure based on wavelet transform; inputting serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH into a multilayer perceptron structure; outputting the corresponding nonlinear mapping features respectively; based on the nonlinear mapping features, constructing a neural functional coupling structure diagram of the stroke risk prediction object. The neural functional coupling structure diagram includes multiple nodes and multiple edges. Nodes represent the cerebral blood supply system, cerebral oxygen supply system, autonomic nervous regulation system, and metabolic homeostasis system, and edges represent the intensity of synergistic abnormalities between different nodes.

[0071] S3.1: Calculate the coupling anomaly coefficient between each node in the neural functional coupling structure diagram. The coupling anomaly coefficient is obtained by multiplying the Pearson correlation coefficient between the time series offset vectors of the corresponding index groups of adjacent nodes for a preset duration by the collaborative reference stability factor.

[0072] S3.2: The adjacency matrix of the neural functional coupling structure graph and the structural feature vector of each node are input into the mechanism-driven embedding network for joint modeling. The mechanism-driven embedding network includes a structure nested graph encoding layer, a side-gated linkage propagation layer and a graph global response convolutional layer.

[0073] S3.3: Based on the output of the mechanism-driven embedded network, determine whether the stroke risk prediction object is currently in a stroke risk state and output the assessed risk state.

[0074] It should be noted that the carotid peak blood flow velocity, pulsatility index, resistance index, cerebral oxygen saturation, heart rate variability parameters, serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH value selected in this embodiment are all important indicators that can stably reflect the nervous system, circulatory system, and biochemical metabolic state in the pre-stroke stage, and have clear medical significance and engineering feasibility. Carotid blood flow velocity, pulsatility index, and resistance index can reflect the dynamic stability of the cerebral blood supply pathway and are the basis for judging whether there are abnormalities in cerebral perfusion; cerebral oxygen saturation is an important parameter for non-invasive continuous monitoring of cerebral oxygen metabolism, and it usually decreases significantly in cerebral hypoxia before stroke onset; heart rate variability parameters can reflect the regulatory capacity of the autonomic nervous system, and its change pattern is highly correlated with the sympathetic-parasympathetic imbalance before stroke; serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH value constitute the core parameters of the metabolic homeostasis system, and their combined fluctuations often precede the appearance of clinical symptoms and manifest as cellular-level metabolic abnormalities, which have a prospective indicative role in stroke risk. All of the above parameters are structured data that are easy to obtain in hospital or home testing, and are feasible and scalable. The selection logic is clear and the coverage dimensions are comprehensive.

[0075] This embodiment introduces a coupled structure graph modeling approach based on the aforementioned multiple sets of indicators. Instead of analyzing each parameter in isolation, it constructs nodes representing the physiological functional systems to which the indicators belong, and edges representing the co-evolutionary trends between different systems, thus forming a quantifiable and computable coupled structure graph of neural functions. Compared to traditional models that rely solely on static feature inputs or classification algorithms, this structure graph can comprehensively express the interactive response states between the four major systems of "brain blood supply—brain oxygen supply—neural regulation—metabolic homeostasis." Combined with time-series trends, it uses Pearson correlation coefficients and historical stability factors to construct coupling anomaly coefficients, exhibiting time-varying and individual adaptability, effectively reducing false positive risks and improving early identification capabilities. The coupled structure graph is not used as static input but is input along with the structural feature vectors of each node into a mechanism-driven embedding network for joint modeling. The mechanism-driven embedding network extracts graph topological features using a nested graph encoding layer, simulates dynamic influence paths between systems using an edge-gated propagation layer, and integrates the overall structural response state using a global convolutional layer, thereby achieving a physically meaningful embedding representation within the neural network. This graph-based modeling method not only preserves the state information of each physiological system, but also establishes a numerical expression of the interaction mechanism between systems. It avoids the semantic confusion caused by feature splicing in traditional models, and significantly improves the interpretability and control accuracy of risk assessment while ensuring the predictive ability of the model.

[0076] In some embodiments of this application, when inputting the carotid peak blood flow velocity, pulsatility index, and resistance index into the first bidirectional long short-term memory network, the following steps are included:

[0077] The peak blood flow velocity, pulsatility index and resistance index of the carotid artery collected at fixed time intervals during a continuous monitoring period for the stroke risk prediction object are respectively constructed into a three-dimensional input vector sequence in chronological order;

[0078] The three-dimensional input vector sequence is input into the first bidirectional long short-term memory network, which is composed of forward LSTM units and backward LSTM units connected in parallel. Information extraction is performed in the forward time flow and the backward time flow respectively to obtain the forward hidden state vector and the backward hidden state vector corresponding to each time point.

[0079] The forward hidden state vector and the backward hidden state vector are concatenated to obtain the bidirectional context joint feature sequence;

[0080] The bidirectional context joint feature sequence is processed based on a sliding window. The mean of time period, maximum variation rate and local mutation gradient within each window are extracted and jointly encoded with the mean of the original input index of the corresponding time period to construct the time period feature vector of the cerebral blood supply system.

[0081] Arrange all time-segment feature vectors in chronological order to obtain a sequence of structural feature representations for nodes in the brain blood supply system in the neural functional coupling structure diagram.

[0082] The mean of the original input indicators is the average value of all sampling points calculated for the peak blood flow velocity, pulsatility index and resistance index of the carotid artery within each sliding window time period.

[0083] Specifically, within a continuous monitoring period, the peak carotid artery velocity, pulsatility index, and resistance index of the stroke risk prediction subjects are simultaneously collected at fixed time intervals (e.g., every 30 seconds or every minute). These data are then arranged chronologically to form a three-dimensional index vector for each time point, which is then stacked sequentially to form an input vector sequence, denoted as X = [x1, x2, ..., x...]. n ], where each x i It is a vector of dimension 3 containing three indices.

[0084] Subsequently, the input vector sequence X is fed into a first bidirectional long short-term memory network consisting of forward LSTM units and backward LSTM units connected in parallel. The forward LSTM processes X in chronological order to obtain the forward hidden state sequence H. + =[h1 + h2 + ,...,h n + The reverse LSTM processes X in reverse chronological order to obtain the reverse hidden state sequence H. - =[h1 - h2 - ,...,h n - ].

[0085] Then, for each time point i, the forward hidden state vector h is... i + With the reverse hidden state vector h i - The features are concatenated to form a bidirectional contextual joint feature vector h. i =[h i + h i - This forms a joint feature sequence H = [h1, h2, ..., h] n This sequence can simultaneously encode the forward and backward dynamic blood flow information at the current moment.

[0086] Next, a sliding window process is applied to the joint feature sequence H. A fixed-length sliding window (e.g., the number of sampling points corresponding to 5 minutes or 10 minutes) is set, and local time periods are extracted by sliding along the time axis. For each joint feature subsequence within the sliding window, all h values ​​within that time period are calculated.i The average value of the vector elements is used to obtain the mean vector for the time period; the maximum rate of variation (the ratio of the maximum change value to the initial value) and the local mutation gradient (the maximum first-order difference value within the current window) are calculated for this subsequence as an expression of the trend and abnormal mutations; at the same time, the mean values ​​of peak blood flow velocity, pulsatility index and resistance index are calculated in the original input index sequence corresponding to the sliding window to form the original mean vector.

[0087] The joint feature statistics (mean, maximum variability, mutation gradient) of the time period are concatenated with the mean of the original input index in terms of dimension to form a time period feature vector of the cerebral blood supply system of uniform length, which represents the blood supply dynamics characteristics and vascular function response status within the time period.

[0088] The aforementioned time-segment feature vectors are arranged sequentially according to their original time order to form a structurally stable time-segment feature sequence, which is used as a structural feature representation of the cerebral blood supply system nodes in the neural functional coupling structure diagram. This sequence not only preserves the dynamic blood supply change trend but also embeds the interaction information between abnormal mutation patterns and original vascular parameters, providing a feature expression with contextual meaning and physiological logic for subsequent inter-system coupling analysis and risk prediction.

[0089] Furthermore, the maximum rate of variation measures the relative change between points where the indicator changes most dramatically within a certain time period, reflecting extreme trends of rapid local increases or decreases. Local mutation gradients are used to capture the slope characteristics of signals that exhibit significant jumps within a short period, quantifying the steepness of the slope of sudden abnormal trends. The steps for obtaining this gradient can be: using the current sliding window length n as a baseline, perform a difference operation on the sampled sequence; in the differenced sequence, calculate the maximum value of the absolute slope.

[0090] It should be noted that the first bidirectional long short-term memory network in this embodiment is a standard Bi-LSTM network structure, which is a widely disclosed general model in the field of deep learning, especially in time series modeling, where it has been supported by a large number of application documents and frameworks.

[0091] It should be noted that, in this embodiment, joint encoding refers to extracting multiple different types of feature information from the input data and model output within each time sliding window, and integrating this information into a unified vector through concatenation and compression to represent the state of the cerebral blood supply system during that time period. The entire process includes the following three steps:

[0092] First, for each sliding window, four sets of information are extracted: 1) the average value of the original input data within the window (i.e., carotid peak blood flow velocity, pulsatility index, and resistance index) during that time period; 2) the difference corresponding to the moment when these three parameters show the largest change within that time period, used to characterize the maximum variability; 3) the degree of rapid mutations occurring within the window, reflected by the local mutation gradient; and 4) the contextual feature information of the window output by the bidirectional long short-term memory network, representing its temporal dependence. These four sets of information all represent important dynamic features within the window.

[0093] Next, the four sets of information are arranged and combined in a fixed order, directly concatenating them into a complete feature vector. This concatenation does not involve weighting or merging; it simply places each value in a corresponding position within the vector to ensure that all features are completely preserved.

[0094] Finally, to control the overall dimensionality of the feature vectors and make them suitable for subsequent neural network input, the concatenated vectors need to be compressed. This can be done in a simple way, such as by adding a fully connected neural network layer to map it to a fixed-dimensional vector. Alternatively, depending on system requirements, dimensionality reduction techniques can be used to extract the most representative information.

[0095] Through the above splicing and compression operations, each time period is encoded into a vector with a unified structure to describe the functional state of the brain's blood supply system during that time period, thus providing basic data for the subsequent construction of a neural functional coupling structure diagram.

[0096] In some embodiments of this application, brain oxygen saturation is input into a convolutional neural network structure that includes differential operations, including:

[0097] Brain oxygen saturation data were constructed as a one-dimensional time series; first-order difference operation was performed on the brain oxygen saturation time series to generate a difference series representing the change amplitude of adjacent sampling points;

[0098] The one-dimensional time series and the difference series are concatenated along the channel dimension to obtain a two-dimensional input matrix;

[0099] The two-dimensional input matrix is ​​sequentially fed into the first, second, and third one-dimensional convolutional layers. The first one-dimensional convolutional layer uses multiple convolutional kernels of length 3, the second one-dimensional convolutional layer uses multiple convolutional kernels of length 5, and the third one-dimensional convolutional layer uses multiple convolutional kernels of length 7. The output of each convolutional layer is processed by batch normalization and ReLU activation function before being passed to the next convolutional layer. After the output of the third one-dimensional convolutional layer, a one-dimensional global average pooling layer is connected to output a feature vector of fixed dimensions.

[0100] The feature vectors are used as structural feature vectors of the brain oxygen supply system in the neural functional coupling structure diagram.

[0101] In this embodiment, brain oxygen saturation is processed as a time series signal. First, based on the raw brain oxygen saturation data obtained from continuous monitoring, a one-dimensional time series is constructed in the order of sampling time. Each data point in this series corresponds to a brain oxygen saturation value at a fixed sampling time point. Then, a first-order difference operation is performed on this one-dimensional time series, that is, the difference between the t-th sample value and the (t-1)-th sample value in the series is calculated to form a new difference sequence, which is used to represent the change in brain oxygen saturation between two consecutive sampling times. This difference process is performed point by point, and the first point can be set to 0 or discarded if there is no previous value.

[0102] To preserve the original trends and variation characteristics, the original one-dimensional time series and the difference series are concatenated along the channel dimension to obtain a two-dimensional input matrix. Each row corresponds to a time point, and each column represents the original brain oxygen saturation value and its difference at that time point. This two-dimensional input matrix serves as the input to the model, which is then sequentially fed into three one-dimensional convolutional neural network structures.

[0103] The first one-dimensional convolutional layer uses multiple convolutional kernels of length 3 and performs sliding convolution with a stride of 1 to extract local features for each time segment of the input matrix. After the convolution operation, batch normalization is performed on the resulting feature map to suppress internal covariate shifts, and then the ReLU activation function is used to introduce nonlinear mapping capabilities. The second one-dimensional convolutional layer uses multiple convolutional kernels of length 5 on the output of the previous layer to further extract mesoscale temporal features, and repeats the convolution, normalization and activation function processing. The third one-dimensional convolutional layer uses multiple convolutional kernels of length 7 to expand the receptive field and extract longer temporal change patterns, and continues to complete the same operation process as the previous layer.

[0104] After the third convolutional layer output, to avoid introducing redundant parameters and improve the model's generalization ability, a one-dimensional global average pooling layer is applied. This averages the output sequence of each channel, compressing the time dimension into a single value, thus generating a fixed-dimensional feature vector. The length of this vector is equal to the number of output channels of the last convolutional layer. Ultimately, this feature vector serves as the structural feature vector of the brain oxygen supply system in the neural functional coupling structure diagram, representing the dynamic change trend and local fluctuation characteristics of the system during the monitoring period. It is used for subsequent modeling of coupling relationships with other functional system nodes and anomaly analysis.

[0105] It should be noted that the first convolutional layer has a kernel length of 3, used to capture rapid fluctuations, abrupt changes, and micro-oscillations in brain oxygen saturation within very short timescales (e.g., within seconds), such as transient hypoxia and noise disturbances. This small receptive field design helps the model to be highly sensitive to local changes, thus forming a high response to short-term anomalies. The second convolutional layer has a kernel length of 5, used to extract trend structures within medium timescales, such as respiratory rhythm fluctuations, local slow declines, or recovery phases. This layer acts as a bridge between local responses and global trends, enhancing the model's perception of transitional regions between adjacent states. The third convolutional layer has a kernel length of 7, with a receptive field covering a longer time period, used to identify chronic trends in brain oxygen saturation, phased plateaus, or overall downward trends. This convolutional layer plays a core role in understanding the stability of large-scale features and the segmentation of phase states.

[0106] In some embodiments of this application, the heart rate variability parameters are input into a wavelet transform-based residual neural network structure, including:

[0107] Heart rate variability parameters were constructed into time series data, and wavelet multi-scale decomposition was performed to extract two high-frequency sub-band signals, which were used as the first sub-band sequence and the second sub-band sequence, respectively.

[0108] Each sub-band sequence is divided into time sliding windows of a preset length. Within each time sliding window, the absolute value sequence of the difference between two adjacent sampling points is calculated to construct several local fluctuation amplitude sequences.

[0109] For each local fluctuation amplitude sequence, calculate the mean and standard deviation. If the mean is greater than the mean of the corresponding sub-band sequence in the complete monitoring period plus twice the standard deviation, then the current time sliding window is determined to be the heart rate mutation window.

[0110] The data from all heart rate mutation windows are combined into a two-dimensional tensor in chronological order and then input into a one-dimensional convolutional network containing multiple residual structures for feature extraction to obtain the residual feature tensor.

[0111] The first and second sub-band sequences within the complete monitoring period are respectively input into a one-dimensional convolutional network with multiple kernel sizes to extract periodic variation features, and then concatenated after average pooling to form a global high-frequency sub-band feature vector.

[0112] The residual feature tensor is concatenated with the global high-frequency subband feature vector to form a fused feature tensor. The fused feature tensor is then subjected to max pooling and vector compression to obtain the structural feature vector representing the autonomic nervous system.

[0113] In this embodiment, heart rate variability parameters are processed as one-dimensional time-series data to construct a continuous time series within a complete monitoring period. The time step is fixed to ensure consistent sampling frequency. Next, wavelet transform is used to perform multi-scale decomposition of the heart rate variability time series. A suitable wavelet basis function (such as the Daubechies wavelet or Symlet wavelet) is selected, and first-order or second-order decomposition is performed to extract two high-frequency sub-band signals, labeled as the first sub-band sequence and the second sub-band sequence, respectively. This operation aims to isolate the high-frequency detail information in the original signal to identify local burst variation patterns.

[0114] Subsequently, each high-frequency sub-band sequence is divided using a fixed-length sliding window, with each window covering a continuous time segment. Within each sliding window, the absolute value sequence of the differences between all adjacent sampling points within that window is calculated. This difference sequence constitutes the local fluctuation amplitude sequence, used to reflect the drastic short-term changes in heart rate variability.

[0115] Based on this, the mean and standard deviation of each local fluctuation amplitude sequence are calculated and compared with the overall mean and standard deviation of that sub-band sequence over the complete monitoring period. If the mean fluctuation amplitude of a certain sliding window is greater than the threshold of "mean of the complete period + twice the standard deviation", the window is determined to be a "heart rate mutation window". All data identified as mutation windows are collected and arranged in chronological order to form a two-dimensional tensor, the first dimension of which is the time window number and the second dimension is the sampled data sequence in that window.

[0116] The aforementioned two-dimensional tensor is input into a one-dimensional convolutional neural network containing multiple residual structures for feature extraction. The residual structure typically consists of two adjacent one-dimensional convolutional layers, with kernel sizes uniformly set to 3 or 5. Each convolutional layer is followed by BatchNorm and ReLU activation. The residual connections skip one convolutional layer, making feature gradients less prone to decay in deep networks, thus improving training stability and deep semantic extraction capabilities. The network ultimately outputs a residual feature tensor, representing the extracted temporal mutation patterns within the heart rate mutation window.

[0117] Simultaneously, the first sub-band sequence and the second sub-band sequence (without sliding window filtering) are input into another convolutional network. This convolutional network contains multiple parallel one-dimensional convolutional layers, each using convolutional kernels of different sizes (e.g., 3, 5, 7, etc.) to extract the periodic variation features of the high-frequency sub-bands. Each convolutional output is followed by a pooling operation, and an average pooling operation is performed in the final layer to compress the convolutional results at each scale to a uniform dimension before concatenation, forming a global high-frequency sub-band feature vector.

[0118] Next, the residual feature tensor and the global high-frequency subband feature vector are concatenated along the channel dimension to form a fused feature tensor. To further reduce dimensionality and extract the most representative comprehensive features, max pooling is performed on the fused feature tensor, and a vector compression strategy (such as fully connected layers or global average pooling) is used to transform it into a fixed-length structural feature vector.

[0119] Ultimately, this structural feature vector is used to characterize the state of the autonomic nervous system node of the current stroke risk prediction object in the neural functional coupling structure diagram, serving as the structural feature representation of that node. The entire process achieves multi-scale joint modeling and feature extraction of heart rate variability signals from local mutations to global cycles.

[0120] In some embodiments of this application, serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH are input into the multilayer sensor structure, including:

[0121] Serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH were constructed into a three-dimensional time series vector. The three-dimensional time series vector was normalized and divided into several time segments using an overlapping sliding window method.

[0122] In each time segment, the covariance amplitude between lactate concentration and arterial carbon dioxide partial pressure, the correlation direction coefficient between serum lactate concentration and blood pH, and the range index of blood pH change rate are calculated to construct a metabolic synergistic state characteristic group.

[0123] The metabolic co-state feature set is input into a multilayer perceptron structure containing three hidden layers. Each hidden layer of the multilayer perceptron structure has 32, 16 and 8 nodes respectively. The ReLU activation function is used to output a mapping feature with consistent dimension.

[0124] The perceptron outputs for all time segments are averaged over time to obtain the full-cycle state representation vector of the metabolic homeostasis system. This state representation vector is then used as the structural feature vector of the metabolic homeostasis system in the neural functional coupling structure diagram.

[0125] Specifically, serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH value, collected at fixed time intervals during a continuous monitoring period for stroke risk prediction subjects, were constructed into three sets of time series data of equal length. Subsequently, these three sets of data were concatenated into a single three-dimensional time series vector sequence using the same timestamp, and this three-dimensional vector sequence was normalized to eliminate the influence of dimensional differences between different indicators on subsequent feature extraction.

[0126] Next, the normalized 3D vector sequence is divided into several time segments using an overlapping sliding window approach. Each sliding window contains data from several consecutive time points, and adjacent sliding windows overlap to ensure temporal continuity and boundary stability. Within each time segment, the following three features are calculated:

[0127] Covariance amplitude: For the corresponding numerical sequences of serum lactate concentration and arterial carbon dioxide partial pressure in the current time period, calculate the absolute value of the covariance and normalize it in combination with the standard deviation of these two indicators in the overall monitoring period to obtain the covariance amplitude that reflects the intensity of their coordinated change.

[0128] Directional correlation coefficient: For the serum lactate concentration and blood pH value in the current time period, calculate the Pearson correlation coefficient and give a signed directional correlation scalar based on the direction of change (increasing or decreasing) of the two.

[0129] Range index: The difference between the maximum and minimum values ​​of blood pH change rate within the current time period, i.e. the difference sequence between two adjacent time points, is extracted as the range measure of pH fluctuation during that time period.

[0130] The three features mentioned above are concatenated according to time periods to form a metabolic coordination state feature group, which is then used as the input feature for that time period and sequentially input into a multilayer perceptron structure. This multilayer perceptron structure contains three fully connected hidden layers with the number of nodes set to 32, 16, and 8 respectively. All hidden layers use the ReLU activation function and output a mapped feature vector of the same dimension.

[0131] The perceptron outputs for all time segments are averaged along the time dimension to obtain a full-cycle state expression vector representing the metabolic homeostasis characteristics within the complete monitoring period. Finally, this state expression vector is used as the structural feature vector of the metabolic homeostasis system in the neural functional coupling structure diagram for subsequent stroke risk prediction.

[0132] In some embodiments of this application, the steps for obtaining the covariance amplitude are as follows: within a fixed-length time sliding window, continuous sampling values ​​of lactic acid concentration and carbon dioxide partial pressure are extracted respectively, the mean values ​​of the two sets of sampling values ​​within the current window are calculated respectively, the difference between each sampling point value and the corresponding mean value is calculated one by one, and the average value of the two sets of differences is calculated by multiplying the corresponding positions of the differences to obtain the covariance amplitude within the current time window.

[0133] The correlation direction coefficient is obtained by calculating the cosine similarity.

[0134] The range index is the difference between the maximum and minimum values ​​within a time sliding window.

[0135] In this embodiment, the steps for calculating the relevant direction coefficients are as follows: First, perform first-order differences on the serum lactate concentration and blood pH value sequences respectively to construct the direction vectors:

[0136] ΔL={l 2- l1, l3-l2, ..., l n -l n-1};

[0137] ΔP={p2-p1,p3-p2,...,p n -p n-1};

[0138] but,

[0139] Where ΔL is the direction vector of serum lactate concentration, ΔP is the direction vector of blood pH value, and l n p represents the serum lactate concentration from the nth collection. n Δl represents the pH value of the blood collected for the nth time. i Indicate l i -l i-1 , Δp i p i -p i-1 The closer the correlation coefficient is to 1, the more synchronous the pH value and lactic acid change.

[0140] Where n is the number of data points for serum lactate concentration or arterial carbon dioxide partial pressure, l i Let c be the serum lactate concentration from the i-th collection. i The partial pressure of arterial carbon dioxide is measured in the i-th data collection. and These represent the average values ​​of serum lactate concentration and arterial carbon dioxide partial pressure, respectively.

[0141] In some embodiments of this application, obtaining the coupling anomaly coefficient includes:

[0142] The structural feature vector sequences collected at fixed time intervals within a preset time period of the coupled node pairs are extracted into index time series groups; the coupled node pairs are the pairwise combinations of four nodes.

[0143] Perform first-order differencing on each set of indicator time series to construct the corresponding offset vector;

[0144] Two offset vectors are selected to calculate the Pearson correlation coefficient, which serves as the correlation trend factor for the coupled node pair in the current time period;

[0145] Take the historical Pearson correlation coefficient and obtain the corresponding correlation trend factor. Take the reciprocal of the variance of the corresponding correlation trend factor as the stability factor.

[0146] Multiply the relevant trend factor and the stability factor for the current time period to obtain the coupling anomaly coefficient;

[0147] If the coupling anomaly coefficient is lower than the mean of the same coupled node pair over consecutive periods minus two standard deviations, then the node pair is marked as having a coupling anomaly.

[0148] It should be noted that for the four key system nodes defined in the neural functional coupling structure diagram—namely, the cerebral blood supply system node, the cerebral oxygen supply system node, the autonomic nervous regulation system node, and the metabolic homeostasis system node—structural feature vectors are obtained from their respective feature extraction modules. For each system node, its structural feature vector is collected at set fixed time intervals (e.g., every minute or every five minutes) within a continuous monitoring period, constructing the corresponding time series. Let the total collection duration be T minutes and the sampling interval be Δt; then each node forms a feature time series of length N = T / Δt.

[0149] Then, for the four nodes mentioned above, all pairwise coupled node pairs are constructed, forming six sets of coupled node pairs. For each set of node pairs, its corresponding structural feature vector sequence is extracted, and the structural feature vector of each node is expanded into a set of index time series with consistent dimensions, that is, the numerical change sequence of each dimension index within the entire sampling time window.

[0150] Next, a first-order differencing operation is performed on each indicator time series to obtain a sequence of changes between adjacent time points, denoted as an offset vector, reflecting the dynamic trend of the indicator within that node. For example, if an indicator sequence is [x1,x2,x3,…,xN], then the corresponding first-order differencing vector is [x2-x1,x3...xN]. - x2,…,xN-xN-1].

[0151] For a given pair of coupled nodes, within the given time window, select any one-dimensional offset vector (or a one-dimensional principal trend offset vector after principal component compression) corresponding to the two nodes in the pair, and calculate the Pearson correlation coefficient between these two offset vectors. This coefficient is denoted as the correlation trend factor of the coupled node pair under the current time window. This factor is used to measure whether the linkage change trends of the two systems are consistent. The value range is between [-1,1][-1,1][-1,1]. A positive correlation indicates synchronous rise or fall, while a negative correlation indicates opposite changes.

[0152] Next, multiple relevant trend factors calculated for the coupled node pair across multiple consecutive historical time windows (e.g., 12 time windows sliding in 5-minute increments over the past hour) are extracted to construct a historical set of relevant trend factors. The variance of this set is calculated to represent the fluctuation of the relevant trend. The reciprocal of this variance is then taken as the stability factor of the coupled node pair in the current time period. A larger factor value indicates more stable past trend changes, while a smaller value indicates greater trend fluctuations.

[0153] The coupling anomaly coefficient of the coupled node pair is obtained by multiplying the relevant trend factor calculated in the current time window with the corresponding stability factor. In this scheme, this coefficient is designed as a comprehensive metric that integrates instantaneous trend and long-term stability. The lower the value, the more the current linkage state deviates from the historical stable collaborative trend.

[0154] Finally, let the mean of the coupling anomaly coefficient of the coupled node pair over the past multiple periods be μ and the standard deviation be σ. If the currently calculated coupling anomaly coefficient is less than μ-2σ, that is, significantly lower than the historical average level and exceeds the two-fold fluctuation limit, then it is determined that the coupled node pair has a structural coupling anomaly in the current time window and is marked for subsequent risk reasoning or path diagnosis processes.

[0155] In some embodiments of this application, determining whether a stroke risk prediction object is currently in a stroke risk state includes:

[0156] The embedding vectors of each functional system node in the mechanism-driven embedded network output are concatenated to construct a comprehensive state representation tensor.

[0157] The discrimination index is calculated based on the comprehensive state representation tensor. The discrimination index includes global similarity offset, local edge weight anomaly density and clustering consistency loss value between nodes.

[0158] The global similarity offset is the cosine distance between the current integrated state tensor and the integrated state representation tensor of the normal samples;

[0159] Local edge weight anomaly density is the percentage of edges with edge weights greater than a set upper limit among all node pairs;

[0160] The clustering consistency loss between nodes is the ratio of the cohesion loss of the current embedded vector in the preset functional cluster to the cross-cluster distance.

[0161] A risk status score is generated based on the discrimination index. If the score is greater than the high-risk threshold, the patient is determined to be in a high-risk state for stroke. If the risk status score is less than or equal to the high-risk threshold but greater than the medium-risk threshold, the patient is determined to be in a medium-risk state. If the risk status score is less than or equal to the medium-risk threshold, the patient is determined to be in a low-risk state.

[0162] In this embodiment, the process of determining whether a stroke risk prediction subject is currently in a stroke risk state first relies on the output of a mechanism-driven embedding network. This mechanism-driven embedding network has, in a previous step, converted multiple functional system nodes (such as cerebral blood supply system nodes, cerebral oxygen supply system nodes, autonomic nervous system regulation system nodes, and metabolic homeostasis system nodes) into corresponding embedding vectors. These embedding vectors are essentially high-dimensional feature vectors representing the current state of each functional system. The embedding vectors of these system nodes are then concatenated in a fixed order to obtain a unified-dimensional comprehensive state representation tensor, which serves as the overall functional state representation of the individual across multiple system interaction dimensions.

[0163] Subsequently, based on this comprehensive state representation tensor, three discriminant indicators for characterizing the current risk level are calculated: global similarity shift, local edge weight anomaly density, and node clustering consistency loss. The global similarity shift is calculated by performing a cosine similarity calculation between the current individual's comprehensive state representation tensor and the pre-stored normal sample state tensor in the system. Specifically, the complementary distance is obtained by calculating the cosine of the angle between the two vectors, reflecting the degree of deviation between the current state and the normal baseline state; a larger value indicates a more abnormal state.

[0164] The calculation of local edge weight anomaly density is based on the edge weights between nodes in the neural functional coupling structure graph. It counts the number of edges whose weights exceed a set upper threshold among all node pairs, and uses this percentage as the local edge weight anomaly density. This index reflects the density of anomalies in interactions between local systems. Edge weights are typically defined based on the strength of cooperative relationships in preceding steps, such as coupling anomaly coefficients or eigenvector similarity.

[0165] The clustering consistency loss value between nodes is calculated as follows: First, based on the preset functional cluster division, the corresponding embedding vectors in the comprehensive state representation tensor are assigned to different functional clusters, and the average Euclidean distance between the embedding vectors within each functional cluster is calculated as the cohesion index; at the same time, the average Euclidean distance between cross-cluster vectors is calculated as the separation index; the ratio of the cohesion index to the separation index is used as the clustering consistency loss value. The larger this value is, the more blurred the boundary of the system's functional division and the more disordered the internal structure.

[0166] After the three discriminant indicators are calculated, they are synthesized using a weighted scoring function to form the final risk status score. The weights of each indicator can be adjusted based on the sensitivity and specificity of stroke incidence in historical training data. Finally, the risk status score is compared with a pre-set risk level threshold. If the score is higher than the high-risk threshold (e.g., set at 0.85), the individual is determined to be at high risk of stroke; if the score is between the medium-risk and high-risk thresholds (e.g., 0.65 to 0.85), the individual is determined to be at medium risk; if the score is less than or equal to the medium-risk threshold, the individual is determined to be at low risk.

[0167] In some embodiments of this application, the method further includes: after determining the risk status, mapping the risk assessment level to the corresponding prompt information and color level to remind the user.

[0168] See Figure 2 As shown, this embodiment of the invention provides a system for assessing and alerting the risk of stroke, comprising:

[0169] The acquisition module is configured to collect continuous medical structured test data from stroke risk prediction subjects;

[0170] The feature module is configured to extract features from medical structured test data, including: inputting carotid peak blood flow velocity, pulsatility index, and resistance index into a first bidirectional long short-term memory network; inputting brain oxygen saturation into a convolutional neural network structure containing differential operations; inputting heart rate variability parameters into a residual neural network structure based on wavelet transform; inputting serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH into a multilayer perceptron structure; outputting corresponding nonlinear mapping features respectively; and constructing a neural functional coupling structure diagram of the stroke risk prediction object based on the nonlinear mapping features.

[0171] The state assessment module is configured to calculate the coupling anomaly coefficient between nodes in the neural functional coupling structure graph; input the adjacency matrix of the neural functional coupling structure graph and the structural feature vector of each node into the mechanism-driven embedding network for joint modeling; and based on the output of the mechanism-driven embedding network, determine whether the stroke risk prediction object is currently in a stroke risk state and output the assessed risk state.

[0172] It should be noted that this embodiment may also include: a reminder and interaction management module, which is used to manage the treatment progress and follow-up reminder process for stroke patients, including: recording the follow-up time nodes and corresponding items set by the doctor for the patient; judging whether the follow-up reminder time has been reached based on the dynamic comparison of system time, and reminding the patient in the form of pop-up window, voice or vibration prompts if the conditions are met; displaying the patient's current treatment progress, assessment status change trend, prevention and control measures suggestions and precautions through a visual interface based on the treatment plan and current assessment status; and supporting the retrieval of historical treatment data and assessment reports according to the request of the patient or doctor, so as to realize self-service viewing and intelligent prompts of the treatment trajectory.

[0173] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for assessing and alerting the risk of stroke, characterized in that, include: Collect continuous medical structured test data of stroke risk prediction subjects. The medical structured test data includes carotid peak blood flow velocity, pulsatility index, resistance index, cerebral oxygen saturation, heart rate variability parameters, serum lactate concentration, arterial carbon dioxide partial pressure and blood pH value. Feature extraction is performed on the medical structured detection data, including: inputting carotid artery peak blood flow velocity, pulsatility index, and resistance index into a first bidirectional long short-term memory network; inputting cerebral oxygen saturation into a convolutional neural network structure containing differential operations; inputting heart rate variability parameters into a residual neural network structure based on wavelet transform; inputting serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH into a multilayer perceptron structure; outputting corresponding nonlinear mapping features respectively; and constructing a neural functional coupling structure diagram of the stroke risk prediction object based on the nonlinear mapping features. The neural functional coupling structure diagram includes multiple nodes and multiple edges. Nodes represent the cerebral blood supply system, cerebral oxygen supply system, autonomic nervous regulation system, and metabolic homeostasis system, and edges represent the intensity of synergistic abnormalities between different nodes. Calculate the coupling anomaly coefficient between each node in the neural functional coupling structure diagram. The coupling anomaly coefficient is obtained by multiplying the Pearson correlation coefficient between the time series offset vectors of the corresponding index groups of adjacent nodes for a preset duration by a collaborative reference stability factor. The adjacency matrix of the neural functional coupling structure graph and the structural feature vector of each node are input into the mechanism-driven embedding network for joint modeling. The mechanism-driven embedding network includes a structure nested graph encoding layer, a side-gated linkage propagation layer and a graph global response convolutional layer. Based on the output of the embedded network driven by the aforementioned mechanism, it is determined whether the stroke risk prediction object is currently in a stroke risk state, and the assessed risk state is output.

2. The method for assessing and reminding about the risk of stroke according to claim 1, characterized in that, When inputting carotid peak blood flow velocity, pulsatility index, and resistance index into the first bidirectional long short-term memory network, the following are included: The peak blood flow velocity, pulsatility index and resistance index of the carotid artery collected at fixed time intervals during a continuous monitoring period for the stroke risk prediction object are respectively constructed into a three-dimensional input vector sequence in chronological order; The three-dimensional input vector sequence is input into the first bidirectional long short-term memory network, which is composed of forward LSTM units and backward LSTM units connected in parallel. Information extraction is performed in the forward time flow and the backward time flow respectively to obtain the forward hidden state vector and the backward hidden state vector corresponding to each time point. The forward hidden state vector and the backward hidden state vector are concatenated to obtain the bidirectional context joint feature sequence; The bidirectional context joint feature sequence is processed based on a sliding window. The mean value of the time period, the maximum variation rate and the local mutation gradient within each window are extracted and jointly encoded with the mean value of the original input index of the corresponding time period to construct the time period feature vector of the cerebral blood supply system. Arrange all time-segment feature vectors in chronological order to obtain a sequence of structural feature representations for nodes in the brain blood supply system in the neural functional coupling structure diagram. The average value of the original input index is the average value of all sampling points calculated for the peak blood flow velocity, pulsatility index and resistance index of the carotid artery within each sliding window time period.

3. The method for assessing and reminding about the risk of stroke according to claim 2, characterized in that, Brain oxygen saturation is input into a convolutional neural network structure that includes differential operations, including: Brain oxygen saturation data are constructed into a one-dimensional time series; a first-order difference operation is performed on the brain oxygen saturation time series to generate a difference series representing the change amplitude of adjacent sampling points; The one-dimensional time series and the difference series are concatenated along the channel dimension to obtain a two-dimensional input matrix; The two-dimensional input matrix is ​​sequentially input into a first one-dimensional convolutional layer, a second one-dimensional convolutional layer, and a third one-dimensional convolutional layer. The first one-dimensional convolutional layer has multiple convolutional kernels of length 3, the second one-dimensional convolutional layer has multiple convolutional kernels of length 5, and the third one-dimensional convolutional layer has multiple convolutional kernels of length 7. The output of each convolutional layer is processed by batch normalization and ReLU activation function and then passed to the next convolutional layer. After the output of the third one-dimensional convolutional layer, a one-dimensional global average pooling layer is connected to output a feature vector of fixed dimension. The aforementioned feature vector is used as the structural feature vector of the brain oxygen supply system in the neural functional coupling structure diagram.

4. The method for stroke risk assessment and alerting according to claim 3, characterized in that, The heart rate variability parameters are input into a wavelet transform-based residual neural network structure, including: Heart rate variability parameters were constructed into time series data, and wavelet multi-scale decomposition was performed to extract two high-frequency sub-band signals, which were used as the first sub-band sequence and the second sub-band sequence, respectively. Each sub-band sequence is divided into time sliding windows of a preset length. Within each time sliding window, the absolute value sequence of the difference between two adjacent sampling points is calculated to construct several local fluctuation amplitude sequences. For each local fluctuation amplitude sequence, calculate the mean and standard deviation. If the mean is greater than the mean of the corresponding sub-band sequence in the complete monitoring period plus twice the standard deviation, then the current time sliding window is determined to be a heart rate mutation window. The data from all heart rate mutation windows are combined into a two-dimensional tensor in chronological order and then input into a one-dimensional convolutional network containing multiple residual structures for feature extraction to obtain the residual feature tensor. The first and second sub-band sequences within the complete monitoring period are respectively input into a one-dimensional convolutional network with multiple kernel sizes to extract periodic variation features, and then concatenated after average pooling to form a global high-frequency sub-band feature vector. The residual feature tensor is concatenated with the global high-frequency subband feature vector to form a fused feature tensor. The fused feature tensor is then subjected to max pooling and vector compression to obtain the structural feature vector representing the autonomic nervous system.

5. The method for assessing and reminding about the risk of stroke according to claim 4, characterized in that, Serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH are input into a multilayer sensor structure, including: Serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH were constructed into a three-dimensional time series vector; the three-dimensional time series vector was normalized and divided into several time segments using an overlapping sliding window method; In each time segment, the covariance amplitude between lactate concentration and arterial carbon dioxide partial pressure, the correlation direction coefficient between serum lactate concentration and blood pH, and the range index of blood pH change rate are calculated to construct a metabolic synergistic state characteristic group. The metabolic co-state feature set is input into a multilayer perceptron structure containing three hidden layers. Each hidden layer of the multilayer perceptron structure has 32, 16 and 8 nodes respectively. The ReLU activation function is used to output a mapping feature with consistent dimension. The perceptron outputs for all time segments are averaged over time to obtain the full-cycle state expression vector of the metabolic homeostasis system; this state expression vector is then used as the structural feature vector of the metabolic homeostasis system in the neural functional coupling structure diagram.

6. The method for assessing and reminding about the risk of stroke according to claim 5, characterized in that, The steps for obtaining the covariance amplitude are as follows: within a fixed-length time sliding window, continuously sampled values ​​of lactic acid concentration and carbon dioxide partial pressure are extracted respectively, the mean values ​​of the two sets of sampled values ​​within the current window are calculated respectively, the difference between each sampled value and the corresponding mean value is calculated one by one, and the average value of the two sets of differences is calculated by multiplying the corresponding positions of the differences to obtain the covariance amplitude within the current time window. The relevant direction coefficients are obtained by calculating cosine similarity. The range index is the difference between the maximum and minimum values ​​within a time sliding window.

7. The method for stroke risk assessment and alerting according to claim 6, characterized in that, The acquisition of the coupling anomaly coefficient includes: The structural feature vector sequences collected at fixed time intervals within a preset time period from the coupled node pairs are extracted into index time series groups; the coupled node pairs are pairwise combinations of four nodes. Perform first-order differencing on each set of indicator time series to construct the corresponding offset vector; Two offset vectors are selected to calculate the Pearson correlation coefficient, which serves as the correlation trend factor for the coupled node pair in the current time period; Take the historical Pearson correlation coefficient and obtain the corresponding correlation trend factor. Take the reciprocal of the variance of the corresponding correlation trend factor as the stability factor. The coupling anomaly coefficient is obtained by multiplying the relevant trend factor and the stability factor for the current time period. If the coupling anomaly coefficient is lower than the mean of the same coupled node pair over consecutive periods minus two standard deviations, then the node pair is marked as having a coupling anomaly.

8. The method for assessing and reminding about the risk of stroke according to claim 7, characterized in that, Determining whether a person is currently at risk of stroke includes: The embedding vectors of each functional system node in the mechanism-driven embedded network output are concatenated to construct a comprehensive state representation tensor. The discrimination index is calculated based on the comprehensive state representation tensor. The discrimination index includes global similarity offset, local edge weight anomaly density and clustering consistency loss value between nodes. The global similarity offset is the cosine distance between the current integrated state tensor and the integrated state representation tensor of the normal sample; Local edge weight anomaly density is the percentage of edges with edge weights greater than a set upper limit among all node pairs; The clustering consistency loss between nodes is the ratio of the cohesion loss of the current embedded vector in the preset functional cluster to the cross-cluster distance. A risk status score is generated based on the discrimination index. If the score is greater than the high-risk threshold, the patient is determined to be in a high-risk state for stroke. If the risk status score is less than or equal to the high-risk threshold but greater than the medium-risk threshold, the patient is determined to be in a medium-risk state. If the risk status score is less than or equal to the medium-risk threshold, the patient is determined to be in a low-risk state.

9. The method for assessing and reminding about the risk of stroke according to claim 8, characterized in that, Also includes: Once the risk status is determined, the risk assessment level is mapped to a corresponding prompt message and color level to remind the user.

10. A system for assessing and alerting the risk of stroke, used to implement the method described in any one of claims 1-9, characterized in that, The system includes: The acquisition module is configured to collect continuous medical structured test data from stroke risk prediction subjects; The feature module is configured to extract features from the medical structured detection data, including: inputting carotid peak blood flow velocity, pulsatility index, and resistance index into a first bidirectional long short-term memory network; inputting brain oxygen saturation into a convolutional neural network structure containing differential operations; inputting heart rate variability parameters into a residual neural network structure based on wavelet transform; inputting serum lactate concentration, arterial carbon dioxide partial pressure, and blood pH into a multilayer perceptron structure; outputting corresponding nonlinear mapping features respectively; and constructing a neural functional coupling structure diagram of the stroke risk prediction object based on the nonlinear mapping features. The state assessment module is configured to calculate the coupling anomaly coefficient between nodes in the neural functional coupling structure graph; input the adjacency matrix of the neural functional coupling structure graph and the structural feature vector of each node into the mechanism-driven embedding network for joint modeling; and, based on the output of the mechanism-driven embedding network, determine whether the stroke risk prediction object is currently in a stroke risk state and output the assessed risk state.

Images