Web-based big data driven stroke risk prediction method

Abstract

The invention relates to the technical field of medical prediction, in particular to a Web-based big data driven stroke risk prediction method, which comprises the following steps: S1, multi-scale data acquisition and alignment: acquiring physiological data, behavior data and health state data of a user; s2, causal path identification: generating a causal feature sub-graph with confidence rating; s3, feature fusion modeling: performing response modeling on fluctuation features in the real-time physiological data, performing trend feature extraction on behavior data, and dynamically allocating fusion weights of the fluctuation features and the trend features to generate unified risk characterization features; s4, risk evolution simulation: simulating the influence of different intervention strategies on the stroke risk, and generating a risk evolution trajectory; and S5, intervention scheme generation: generating a personalized intervention scheme. According to the invention, the accuracy of risk prediction is improved, and the change of the health condition of the patient can be responded in real time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical prediction, and particularly relates to a Web-based big data driven stroke risk prediction method. BACKGROUND

[0002] Stroke is one of the major diseases causing death and disability worldwide, which has a serious impact on the health and quality of life of patients. With the aggravation of population aging, the incidence of stroke is rising. Early prediction and intervention are the key to effectively reduce the mortality and disability rate of stroke. Traditional stroke prediction methods usually rely on the subjective assessment of patients' symptoms and signs by clinicians, combined with routine health examination data such as blood pressure and blood sugar. These methods can predict the occurrence of stroke to some extent, but due to the lack of accurate real-time data support and personalized intervention strategies, their prediction accuracy and treatment effect still have great limitations.

[0003] At present, the prediction method of stroke mostly depends on a single data source or regular medical examination. These traditional methods face the following problems: first, they lack real-time collection and analysis of multi-dimensional data, which cannot fully reflect the dynamic changes of patients' health status; second, existing methods usually lack causal relationship analysis, which cannot deeply explore the mutual influence between different health factors, resulting in the inability to provide accurate personalized intervention suggestions; third, they lack real-time monitoring and intelligent feedback mechanism of patients' health status, making it difficult to intervene in a timely and effective manner according to changing risks. These deficiencies limit the effectiveness of existing technologies in practical applications, especially in the aspects of individualization, dynamic prediction and intervention. SUMMARY

[0004] The present application provides a Web-based big data driven stroke risk prediction method, which can respond to the health changes of patients in real time, improve the accuracy of stroke prediction and treatment effect, and reduce the incidence and disability rate of stroke.

[0005] A Web-based big data driven stroke risk prediction method, comprising the following steps:

[0006] S1, multi-scale data collection and alignment: collecting physiological data, behavior data and health status data of users, and realizing multi-scale data alignment across time scales through quantum time stamp;

[0007] S2, causal path identification: inputting the aligned multi-scale data into a causal discovery engine to identify the causal relationship between physiological abnormalities and health factors, and generating a causal feature subgraph with confidence rating;

[0008] S3, feature fusion modeling: based on the causal feature subgraph, a time-varying weight network is constructed, impulse neural network (SNN) is used to model the fluctuation features in real-time physiological data, time convolution network (TCN) is used to extract trend features from behavior data, and the fusion weight of fluctuation features and trend features is dynamically allocated according to the current health state data of the user, and a unified risk representation feature is generated;

[0009] S4, risk evolution simulation: the generated risk representation feature is input into the causal reinforcement learning framework to simulate the influence of different intervention strategies on the risk of stroke, and a risk evolution trajectory is generated;

[0010] S5, intervention scheme generation: based on the results of risk evolution simulation, a three-dimensional early warning display is triggered, and a personalized intervention scheme is generated.

[0011] Optionally, the multi-scale data acquisition and alignment in S1 comprises:

[0012] S11, multi-scale data classification acquisition: physiological data, behavior data and health state data of the user are collected, the physiological data includes cerebral blood flow, ECG RR interval and blood pressure waveform, the behavior data includes action acceleration, angular velocity and voice features, and the health state data comes from the electronic medical record system, including stroke score, disability assessment and medication compliance indicators;

[0013] S12, quantum timestamp generation and binding: obtain the time reference through the quantum key distribution network, and attach a unique quantum timestamp to each type of data;

[0014] S13, cross-scale time alignment: millisecond-level compensation is performed on physiological data, periodic correction is applied to behavior data, and health state data is aligned in calendar event mapping mode;

[0015] S14, timestamp verification and exception handling: time difference detection is performed on the collected multi-scale data, and the validity of the timestamp is verified through quantum signature, and the non-compliant data is removed.

[0016] Optionally, the cross-scale time alignment in S13 comprises:

[0017] S131, physiological data alignment: the physiological data acquisition time delay is corrected through a linear compensation model, the original time is aligned to the quantum time reference using the device calibration coefficient, and millisecond-level synchronization is achieved;

[0018] S132, behavior data alignment: for the periodic time error in behavior data, a sinusoidal function correction method is used for second-level alignment to eliminate the time deviation caused by environmental interference and device clock offset;

[0019] S133, health status data alignment: align health status data by calendar time segmentation, correct through time slice mapping and medical system working time offset.

[0020] Optionally, the causal path identification in S2 includes:

[0021] S21, causal feature space construction: standardize the output aligned multi-scale data and construct a three-dimensional causal feature tensor;

[0022] S22, mixed causal structure learning: discover the static causal relationship between variables through constraint-based Bayesian network, and verify the time series causality between variables through time series Granger causality test;

[0023] S23, false association elimination: eliminate false associations through counterfactual causal strength Analysis and screening out weakly related or pseudo-related paths, only retaining causal edges that meet And at the same time, environmental variable control test;

[0024] S24, causal subgraph generation: based on the time lag range of the causal edge, construct a multi-layer causal network structure, which specifically includes:

[0025] The first layer: physiological Physiological causal edge, delay ;

[0026] The second layer: behavior Physiological causal edge, ;

[0027] The third layer: health status Behavioral / physiological causal edge, ;

[0028] The weight of each causal edge of the multi-layer causal network structure is calculated according to the normalized causal strength and Granger test strength;

[0029] S25, confidence rating: fuse counterfactual causal strength, time series Granger causality test results and causal structure stability to generate confidence score And hierarchical evaluation, when The confidence rating is A level, when The confidence rating is B level, when The confidence rating is C level (manual review is required).

[0030] Optionally, the mixed causal structure learning in S22 includes:

[0031] S221, static causal discovery: constraint-based Bayesian network is used for causal graph structure learning, conditional mutual information (MI) is calculated to identify the causal relationship between variables, and sparsity control coefficient is used optimization is performed;

[0032] S222, dynamic causal verification: Granger causality test is used to verify the causal path, and F test statistic is calculated , and the path with time sequence causal relationship is screened out.

[0033] Optionally, the feature fusion modeling in S3 comprises:

[0034] S31, parallel extraction of multi-modal features: dual-channel architecture is used to extract multi-modal features in parallel, real-time fluctuation features are processed by using the leaky integrate-and-fire (LIF) neuron model through the spiking neural network (SNN) channel, and trend features are extracted by using a 6-layer residual dilated convolutional network through the temporal convolution network (TCN) channel;

[0035] S32, dynamic weight adaptive fusion: the health state vector in the health state data is extracted, the dynamic fusion weight of the output of the spiking neural network (SNN) channel and the temporal convolution network (TCN) channel is generated by using a multi-layer perceptron (MLP), and the final risk representation is generated by combining the calculated dynamic fusion weight.

[0036] Optionally, the parallel extraction of multi-modal features in S31 comprises:

[0037] S311, real-time fluctuation modeling: synapse connection is initialized based on the edge weight of the causal subgraph, membrane potential is updated by using the leaky integrate-and-fire (LIF) neuron model, and whether the neuron fires is determined by setting an adaptive threshold to capture fluctuation features;

[0038] S312, trend feature extraction: a 6-layer residual dilated convolutional network is configured, each layer uses a dilated convolution operation to extract trend information of different scales, and a jump connection mechanism is used to combine the output of each layer of convolution with the input to strengthen the feature representation capability.

[0039] Optionally, the dynamic weight adaptive fusion in S32 comprises:

[0040] S321, health state space-time coding: features of health state data are extracted by multi-scale convolution, different time scale information is captured by using dilated convolution, and a health state vector is generated by pooling to extract periodic health patterns;

[0041] S322, Gating weight generation: using a multi-layer perceptron (MLP) to generate dynamic fusion weights of the pulse neural network (SNN) channel and the temporal convolution network (TCN) channel output, combining the health state vector to calculate the weight through the gating operation, and adjusting the weight by using the time period modulation factor;

[0042] S323, Risk representation synthesis: weighting and fusing the output feature maps of the pulse neural network (SNN) channel and the temporal convolution network (TCN) channel according to the adjusted dynamic weights to generate the final risk representation feature.

[0043] Optionally, the risk evolution simulation in S4 includes:

[0044] S41, Virtual patient environment construction: mapping the generated final risk representation feature to the environment state of the virtual patient, mapping the high-dimensional feature to a 512-dimensional space through a self-encoder dimension reduction, and setting constraint conditions for multi-dimensional intervention measures through a multi-dimensional intervention strategy space including drug dosage, rehabilitation intensity and diet scheme;

[0045] S42, Causal transfer function modeling: constructing a neural differential equation (NDE) model to describe the causal relationship of state transfer in the virtual patient environment, defining a state transfer equation to simulate random physiological fluctuations and combining a causal transfer function to analyze the change of health state over time, and evaluating the effect of each intervention measure through a composite reward function, including risk change, intervention cost and compliance;

[0046] S43, Strategy optimization and simulation: generating a behavior strategy using a causal guided policy gradient algorithm, and optimizing the intervention strategy by running multiple instances in parallel to update the health state of the patient until the simulation time reaches the set month;

[0047] S44, Risk trajectory generation: performing kernel density estimation on the simulation results to analyze the evolution trajectory of the risk, calculating the risk distribution at each time point, and extracting the risk threshold crossing event to mark the critical risk change moment.

[0048] Optionally, the intervention scheme generation in S5 includes:

[0049] S51, Analyzing risk changes: extracting the risk change and the critical risk change moment according to the results of the risk evolution simulation;

[0050] S52, Triggering three-dimensional early warning display: when the risk change is detected to exceed 0.2 or reach the critical risk change moment, automatically triggering the three-dimensional early warning display to display the patient's risk state in real time through a three-dimensional visualization graph;

[0051] S53, generating a personalized intervention plan: according to the results of the three-dimensional early warning display, combined with the personalized information of the patient (such as age, gender, medical history, etc.), a personalized intervention plan is generated, including drug treatment, rehabilitation training, diet adjustment.

[0052] The beneficial effects of the present application are:

[0053] The present application, through multi-modal data acquisition, including physiological data, behavior data and health status data, uses quantum timestamp technology to accurately align multi-source data across time scales, ensuring the high quality and timeliness of the data, enabling the synchronous processing of various types of data, ensuring the comparability and usability of the data.

[0054] The present application, through causal path identification and feature fusion modeling, can identify the causal relationship between different health factors and generate a unified risk characterization feature, combining pulse neural networks and time convolution networks for dynamic fusion of multiple features, not only improving the accuracy of risk prediction, but also responding to changes in patient health status in real time, providing strong support for clinical decision-making.

[0055] The present application, by combining drug treatment, rehabilitation training and diet adjustment, can effectively reduce the risk of stroke in patients and improve treatment effect, the combination of three-dimensional early warning display and personalized intervention plan can provide medical personnel with real-time, intuitive risk assessment and intervention suggestions, ensuring timely and accurate treatment. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only illustrate the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0057] Fig. 1 The prediction method flowchart of the present application embodiment is shown in the figure.

[0058] Fig. 2 The causal path identification diagram of the present application embodiment is shown in the figure. DETAILED DESCRIPTION

[0059] The present application will be described in detail below in conjunction with the drawings and specific embodiments. It should be noted here that in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be used by those skilled in the art to implement them; and the drawings are only used to describe the embodiments more specifically, and are not intended to specifically limit the present application.

[0060] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0061] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0062] like Figs. 1-2 As shown, a web-based big data-driven stroke risk prediction method includes the following steps:

[0063] S1, Multi-scale data acquisition and alignment: Collect users' physiological data, behavioral data and health status data, and achieve multi-scale data alignment across time scales through quantum timestamps;

[0064] S2, Causal Path Identification: Aligned multi-scale data is input into the causal discovery engine to identify the causal relationship between physiological abnormalities and health factors, and generate a causal feature subgraph with confidence rating.

[0065] S3, Feature Fusion Modeling: Based on the causal feature subgraph, a time-varying weighted network is constructed. The spiking neural network (SNN) is used to model the response of fluctuation features in real-time physiological data. The temporal convolutional network (TCN) is combined to extract trend features from behavioral data. The fusion weights of fluctuation features and trend features are dynamically allocated according to the user's current health status data to generate a unified risk characterization feature.

[0066] S4, Risk Evolution Simulation: Input the generated risk representation features into a causal reinforcement learning framework to simulate the impact of different intervention strategies on stroke risk and generate risk evolution trajectories;

[0067] S5, Intervention Plan Generation: Based on the results of risk evolution simulation, a three-dimensional early warning display is triggered, and a personalized intervention plan is generated.

[0068] Multi-scale data acquisition and alignment in S1 includes:

[0069] S11, multi-scale data classification collection: collect user's physiological data, behavior data and health status data, physiological data includes cerebral blood flow, ECG RR interval and blood pressure waveform, behavior data includes motion acceleration, angular velocity and speech features, health status data comes from electronic medical record system, including stroke score, disability assessment and medication compliance indicators, specifically including:

[0070] Physiological data collection: physiological data is obtained by a medical-grade wearable device at a sampling rate of 1000 Hz;

[0071] Behavior data collection: nine-axis IMU sensors are used to collect limb motion acceleration , angular velocity and speech features at a sampling rate of 100 Hz, wherein the speech features are extracted by the Mel-frequency cepstral coefficient (MFCC) method and represented as:

[0072] ;

[0073] wherein, is the discrete cosine transform, is the th Mel-frequency cepstral coefficient, is the Mel filter bank, is the power spectrum of the speech frame, is the fast Fourier transform;

[0074] Health status data collection: health status data is extracted from the electronic health record (EHR) system, specifically including:

[0075] Stroke score: ;

[0076] Disability assessment: ;

[0077] Medication compliance indicators: ;

[0078] S12, quantum timestamp generation and binding: obtain time reference through quantum key distribution network, and attach unique quantum timestamp to each type of data, represented as:

[0079] ;

[0080] wherein, is the initial deviation of the device and the quantum clock source, is the duration of the synchronization operation, is the time synchronization decay constant, is the quantum timestamp, is coordinated universal time;

[0081] ;

[0082] wherein, is the final generated timestamp, including time and hash, is the hash value of the data block;

[0083] S13, cross-scale time alignment: millisecond-level compensation is performed on physiological data, periodic correction is applied to behavior data, and health status data is aligned in a calendar event mapping manner, ensuring time consistency of multi-source data;

[0084] S14, timestamp verification and abnormality processing: time difference detection is performed on the collected multi-scale data, and the effectiveness of the timestamp is verified through quantum signature, and non-compliant data is removed to ensure data quality and synchronization accuracy, which is expressed as:

[0085] ;

[0086] wherein, is the maximum allowed time interval deviation, (physiological data), (behavior data), (health data), , is the time interval between adjacent two collected data;

[0087] ;

[0088] wherein, is the verification result, a Boolean type (True / False), is a signature verification algorithm based on quantum key, is a public key.

[0089] The cross-scale time alignment in S13 includes:

[0090] S131, physiological data alignment: the linear compensation model is used to correct the physiological data collection time delay, the original time is aligned to the quantum time reference by using the device calibration coefficient, and millisecond-level synchronization is realized, which is expressed as:

[0091] ;

[0092] wherein, is the biological sensor time delay coefficient, is the corrected physiological data timestamp, is the original data record time;

[0093] S132, behavior data alignment: for the periodic time error in behavior data, a sinusoidal function correction method is used for second-level alignment to eliminate the time deviation caused by environmental interference and device clock offset, expressed as:

[0094] ;

[0095] wherein, is the environmental interference intensity factor, is the sensor crystal frequency, is the corrected behavior data timestamp;

[0096] S133, health status data alignment: the health status data is aligned according to the calendar time segmentation, and the time slice mapping is corrected with the medical system working time offset, to ensure the unified time reference on the day level, expressed as:

[0097] ;

[0098] wherein, is the single-day time slice length, is the medical institution standard working time offset, is the aligned health data timestamp.

[0099] The causal path identification in S2 includes:

[0100] S21, causal feature space construction: the output aligned multi-scale data is standardized, and a three-dimensional causal feature tensor is constructed, expressed as:

[0101] ;

[0102] wherein, is the multi-scale data The mean value in the sliding window second, is the dynamic standard deviation of the multi-scale data in the statistical period hours, is the standardized multi-scale data;

[0103] ;

[0104] wherein, is the three-dimensional causal feature tensor, is the physiological feature index set, is the behavior feature index set, is the health status feature index set, , , respectively, are the corresponding standardized feature values;

[0105] S22, mixed causal structure learning: discover static causal relationships among variables through constraint-based Bayesian networks, and verify the time series causality between variables through time series Granger causality test, to improve the time series interpretation of causal structure;

[0106] S23, false association elimination: through counterfactual causal strength analysis to filter out weakly related or spurious related paths, only keep the causal edges that meet , and at the same time conduct environmental variable control test, so as to retain the core causal chain link with actual intervention value, expressed as:

[0107] ;

[0108] ;

[0109] Among them, is the causal variable in causal path analysis, is the target variable in causal analysis, represents the confounding variable used for control or adjustment, , is the two different states of the causal variable , is Pearl's intervention symbol, is the set of all observed variables, represents the subset of confounding variables used to test conditional independence, represents and are independent under condition;

[0110] S24, causal subgraph generation: based on the time lag range of causal edges, construct a multi-layer causal network structure, including:

[0111] The first layer: physiological physiological causal edges, delay ;

[0112] The second layer: behavioral physiological causal edges, ;

[0113] The third layer: health status behavioral / physiological causal edges, ;

[0114] The weight of each causal edge in the multi-layer causal network structure is calculated according to the normalized causal strength and Granger test strength, expressed as:

[0115] ;

[0116] wherein, is the weight of the causal edge, is the strength of the causal edge, is the maximum causal strength among all edges, is the F-test statistic of Granger causality test; S25, confidence rating: fuse the counterfactual causal strength, the temporal Granger causality test result and the causal structure stability to generate the confidence score

[0117] and evaluate, when the confidence rating is A level, when the confidence rating is B level, and when the confidence rating is C level (manual review is required), which is represented as:

[0118] ;

[0119] wherein, are the corresponding weighting coefficients, is the structural entropy of the parent node of the target node, is the probability distribution of the parent node , is the logarithmic probability of each variable value of the parent node .

[0120] The mixed causal structure learning in S22 includes:

[0121] S221, static causal discovery: a constraint-based Bayesian network is used to learn the causal graph structure, the conditional mutual information (MI) is calculated to identify the causal relationship between variables, and the sparsity control coefficient is used for optimization to retain the most important causal path, which is represented as:

[0122] ;

[0123] wherein, is the Bayesian causal graph structure, is the input data set, is the posterior probability used to learn the optimal graph structure, is the variable has a causal impact on , is the parent node set of the variable , is the conditional mutual information, which measures the degree of association between and given .​

[0124] S222, dynamic causal verification: verify the causal path by Granger causality test, calculate the F test statistic , and filter out the path with time causal relationship, expressed as:

[0125] ;

[0126] wherein, is the residual sum of squares of the restricted model (excluding ), is the residual sum of squares of the unrestricted model (including ), is the lag order corresponding to three time granularities (seconds, minutes, hours), is the effective sample size, which needs to satisfy ;

[0127] The screening criterion of the path with time causal relationship is:

[0128] ;

[0129] wherein, is the critical value of F distribution, the confidence level , and the degree of freedom .

[0130] The feature fusion modeling in S3 includes:

[0131] S31, parallel extraction of multi-modal features: parallel extraction of multi-modal features is performed by using a dual-channel architecture, real-time fluctuation features are processed by using a leaky integrate-and-fire (LIF) neuron model through a spiking neural network (SNN) channel, and trend features are extracted by using a 6-layer residual dilated convolutional network through a temporal convolutional network (TCN) channel;

[0132] S32, dynamic weight adaptive fusion: a health state vector in the health state data is extracted, dynamic fusion weights of outputs of the spiking neural network (SNN) channel and the temporal convolutional network (TCN) channel are generated by using a multi-layer perceptron (MLP), and a final risk representation is generated by combining the calculated dynamic fusion weights.

[0133] The parallel extraction of multi-modal features in S31 includes:

[0134] S311, real-time fluctuation modeling: synapse connections are initialized based on edge weights of a causal subgraph, membrane potentials are updated by using a leaky integrate-and-fire (LIF) neuron model, and whether a neuron fires a pulse is determined by setting an adaptive threshold, so as to capture fluctuation features, and specifically includes:

[0135] S3111, LIF neuron initialization: Synaptic connections are configured based on the edge weights of the causal subgraph, represented as:

[0136] ;

[0137] in, For the first The first neuron and the second Synaptic connection weights between neurons Let the causal edge weights come from the causal subgraph. , representing the randomness of connections in each neuron, follows a uniform distribution. ;

[0138] S3112, Dynamic Update of Membrane Potential: The leaky integral ignition (LIF) neuron model is used to dynamically update the neuronal membrane potential, expressed as:

[0139] ;

[0140] in, For the first The membrane potential of a neuron The membrane time constant is used to match the cerebral vascular pressure fluctuation cycle. This is the leakage compensation coefficient, representing the degree of membrane potential decay. The input current represents the current generated by the activity of the presynaptic neuron. This is the output signal of the presynaptic neuron;

[0141] S3113, pulse triggering mechanism: ;

[0142] in, For the first The pulse output state of each neuron, where 1 indicates firing a pulse and 0 indicates not firing a pulse. As an adaptive threshold, a pulse is fired when the neuron's membrane potential reaches this value;

[0143] S312, Trend Feature Extraction: A 6-layer residual dilated convolutional network is configured, with each layer using dilated convolution operations to extract trend information at different scales. The convolution operations ensure feature dimension matching, and a skip connection mechanism combines the output and input of each convolutional layer to enhance feature representation capabilities. Specifically, this includes:

[0144] S3121, Residual Block Structure Configuration: (The following is a continuation of the previous sentence) layer The parameter configuration includes the expansion rate. kernel size Number of output channels ;

[0145] The input and output dimensions of the convolution operation are matched, which is represented as:

[0146] ;

[0147] Wherein, is the time length (or length dimension) of the output feature map, is the time length (or length dimension) of the input feature map, is the padding length, used to ensure the boundary alignment in the convolution operation, is the stride of the convolution operation, which controls the step length of the convolution kernel sliding;

[0148] S3122, dilated convolution calculation: ;

[0149] Wherein, is the output feature map of the layer, is the convolution kernel weight of the layer, is the output feature map of the layer, is the bias term of the layer;

[0150] S3123, skip connection mechanism: ;

[0151] Wherein, is the weighted output feature map of the layer of the skip connection, is a 1x1 convolution operation used to adjust the number of channels, is the input feature map of the layer.

[0152] The dynamic weight adaptive fusion in S32 includes:

[0153] S321, health status spatiotemporal encoding: features of health status data are extracted by multi-scale convolution, information of different time scales is captured by dilated convolution, and health status vectors are generated by pooling to extract periodic health patterns, which specifically includes:

[0154] S3211, multi-scale convolution feature extraction: input health status data (wherein hours, dimensional features) are encoded by three-level dilated convolution, which is represented as:

[0155] The dilation rate , the output dimension: ;

[0156] dilation rate , output dimension: ;

[0157] dilation rate , output dimension: ;

[0158] wherein, is the feature encoding convolution kernel size, is the feature encoding dilation rate, is the output channel number, is the gated linear unit, , is the sigmoid activation function;

[0159] S3212, time-dependent modeling: a convolutional neural network (CNN) is used to capture the periodic pattern of health status data, represented as:

[0160] daily periodic pattern extraction;

[0161] three-day periodic pattern extraction;

[0162] ;

[0163] wherein, is the daily periodic pattern feature extracted by convolution operation, is the three-day periodic pattern feature extracted by convolution operation, is the feature map obtained after layer normalization and GeLU activation on , is the convolution kernel size of the convolutional neural network, is the dilation rate of the convolutional neural network;

[0164] ;

[0165] wherein, is the final health status vector, is an adaptive max-pooling operation that compresses the dimension of the feature map to a fixed size (512);

[0166] S322, gated weight generation: a multi-layer perceptron (MLP) is used to generate dynamic fusion weights of the pulse neural network (SNN) channel and the temporal convolution network (TCN) channel output, the weights are calculated by combining the health status vector through the gating operation, and the weights are adjusted by using the time period modulation factor, specifically including:

[0167] S3221, Multi-layer perceptron architecture: build a gating projection network, generate dynamic weights, denoted as:

[0168] ;

[0169] ;

[0170] ;

[0171] wherein, is the dynamic weight output by the gating mechanism, is the weight matrix of the gating network, is the bias term of the gating network, is the feature output by the projection network, is the weight matrix of the projection network, is the bias term of the projection network, is the dynamic weight of the SNN channel, is the dynamic weight of the TCN channel;

[0172] S3222, Phase-sensitive adjustment: introduce a time period modulation factor to dynamically adjust the weight, denoted as:

[0173] ;

[0174] ;

[0175] wherein, , are the dynamic weights of the SNN channel and the TCN channel after adjustment respectively, is the adjustment factor, which controls the influence of health status change on the weight, is the current timestamp;

[0176] S323, Risk representation synthesis: the output feature maps of the spiking neural network (SNN) channel and the temporal convolution network (TCN) channel are weighted and fused according to the adjusted dynamic weights to generate the final risk representation feature, ensuring that the fusion process meets the predetermined constraint conditions, including:

[0177] S3231, Feature dimension alignment: align the dimensions of SNN and TCN outputs to ensure that the number of channels matches, denoted as:

[0178] ;

[0179] ;

[0180] wherein, , Output feature maps of SNN and TCN channels respectively, 、 Linearly transformed feature maps for fusion, The desired dimension number of SNN and TCN output feature maps after linear transformation;

[0181] S3232, dynamic weighted fusion: weighted fusion according to the adjusted dynamic weight, generate the final risk representation, denoted as:

[0182] ;

[0183] Wherein, The final generated risk representation feature;

[0184] Constraint condition: ;

[0185] Wherein, 、 The trace of SNN and TCN fusion weight respectively, ensure their sum is 1.

[0186] The risk evolution simulation in S4 includes:

[0187] S41, virtual patient environment construction: map the generated final risk representation feature to the virtual patient's environment state, map the high-dimensional feature to 512-dimensional space through autoencoder dimension reduction, and set constraint conditions for multi-dimensional intervention measures through multi-dimensional intervention strategy space, including drug dosage, rehabilitation intensity and diet scheme, so as to effectively control in the simulation process, including:

[0188] S411, state space definition: map the generated final risk representation feature to the environment state, denoted as:

[0189] ;

[0190] Wherein, The environment state vector, The dimension reduction projection based on autoencoder;

[0191] S412, action space design: define the multi-dimensional intervention strategy space, denoted as:

[0192] ;

[0193] Wherein, Intervention strategy;

[0194] Constraint condition: ;

[0195] ;

[0196] S42, Causal Transition Function Modeling: Construct a Neural Differential Equation (NDE) model to describe the causal relationship of state transition in the virtual patient environment, simulate random physiological fluctuations by defining state transition equations, analyze the changes of health status over time by combining causal transition functions, and evaluate the effectiveness of each intervention measure through a composite reward function, including risk change, intervention cost, and compliance, including:

[0197] S421, State Transition Equation: Construct a causal model based on neural differential equations, represented as:

[0198] ;

[0199] where, is a three-layer multi-layer perception (MLP) network with a structure of , is a learnable parameter, is a heteroscedastic noise network, is an independent parameter, is a Wiener process increment used to simulate random physiological fluctuations;

[0200] S422, Reward Function Design: Calculate the composite reward, represented as:

[0201] ;

[0202] where, is the risk change, is the standardization of drug cost and rehabilitation cost, , is the compliance, based on the predicted value of historical compliance, ranging from , , , is a tunable hyperparameter, is the value of the reward function at time step ;

[0203] S43, Strategy Optimization and Simulation: Use the causal-guided policy gradient algorithm to generate the behavior strategy, and run multiple instances through parallel simulation to optimize the intervention strategy, update the patient's health status until the simulation time reaches the set month, including:

[0204] S431, Behavior Strategy Generation: Use the causal-guided policy gradient algorithm to generate the behavior strategy, represented as:

[0205] ;

[0206] where, is the policy network, representing the state Select action The probability distribution, Let be the advantage function, representing the advantage of each action in a given state. The parameters of the policy network determine the policy for selecting control actions;

[0207] S432, Virtual Clinical Trial Execution: Parallel Execution Each simulation instance executes the following:

[0208] ;

[0209] ;

[0210] ;

[0211] ;

[0212] in, This refers to the number of parallel simulation instances used to run multiple virtual patient environment instances simultaneously. The time step is set to 1 month, representing the time interval between each simulation step. The noise is Gaussian, simulating random physiological fluctuations at each time step, derived from a standard normal distribution. The mean is 0, and the covariance matrix is ​​the identity matrix. , For the current moment Intervention strategies, Given a state Generate an action The probability distribution, For noise model, These are the parameters of the noise model;

[0213] S44, Risk Trajectory Generation: Kernel density estimation is performed on the simulation results to analyze the evolution trajectory of the risk. By calculating the risk distribution at each time point and extracting risk threshold crossing events, key risk change moments are marked, represented as:

[0214] ;

[0215] in, In time At any given moment, the probability density function of risk. The number of simulation instances, For kernel function, , This is the bandwidth parameter of the kernel function. The risk value at the current moment. For the first an individual simulation instance at time a risk value at time

[0216] ;

[0217] wherein, the time of the key event, i.e. the earliest time at which the risk value exceeds the clinical warning threshold, a risk value at time a risk value at time a risk value at time a clinical warning threshold.

[0218] The intervention scheme generation in S5 includes:

[0219] S51, analyzing the risk change: according to the results of the risk evolution simulation, the risk change amount and the key risk change time are extracted;

[0220] S52, triggering three-dimensional warning display: when it is detected that the risk change amount exceeds 0.2 or reaches the key risk change time, the three-dimensional warning display is automatically triggered, and the risk state of the patient is displayed in real time through the three-dimensional visualization graph;

[0221] S53, generating an individualized intervention scheme: according to the results of the three-dimensional warning display, combined with the individualized information of the patient (such as age, gender, medical history, etc.), an individualized intervention scheme is generated, including drug treatment, rehabilitation training, diet adjustment, to reduce the risk of the patient and improve the treatment effect.

[0222] The present application encompasses any substitutions, modifications, equivalent methods and schemes made on the essence and scope of the present application. In order for the public to have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without the description of these details to those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0223] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A web-based big data driven stroke risk prediction method, characterized in that, The method comprises the following steps: S1, multi-scale data acquisition and alignment: physiological data, behavior data and health status data of the user are collected, and multi-scale data alignment across time scales is realized through quantum timestamps; S2, causal path identification: the aligned multi-scale data is input into a causal discovery engine to identify the causal relationship between physiological abnormalities and health factors, and a causal feature subgraph with a confidence rating is generated; S3, feature fusion modeling: a time-varying weight network is constructed based on the causal feature subgraph, impulse neural networks are used to model the fluctuation features in real-time physiological data, time convolution networks are used to extract trend features from behavior data, and the fusion weight of fluctuation features and trend features is dynamically allocated according to the current health status data of the user, to generate unified risk representation features; S4, risk evolution simulation: the generated risk representation features are input into a causal reinforcement learning framework to simulate the influence of different intervention strategies on stroke risk, and a risk evolution trajectory is generated; S5, intervention scheme generation: based on the results of risk evolution simulation, a three-dimensional early warning display is triggered, and a personalized intervention scheme is generated.

2. The web-based big data driven stroke risk prediction method of claim 1, wherein, The multi-scale data acquisition and alignment in S1 comprises: S11, multi-scale data classification acquisition: physiological data, behavior data and health status data of the user are collected, the physiological data includes cerebral blood flow, ECG RR interval and blood pressure waveform, the behavior data includes action acceleration, angular velocity and voice features, and the health status data comes from an electronic medical record system, including stroke score, disability assessment and medication compliance indicators; S12, quantum timestamp generation and binding: a time reference is obtained through a quantum key distribution network, and a unique quantum timestamp is attached to each type of data; S13, cross-scale time alignment: physiological data is compensated at the millisecond level, behavior data is corrected periodically, and health status data is aligned according to calendar event mapping; S14, timestamp verification and exception handling: time difference detection is performed on the collected multi-scale data, and the effectiveness of the timestamp is verified through quantum signature, and the non-compliant data is removed. 3.The Web-based big data-driven stroke risk prediction method of claim 2, wherein, The cross-scale time alignment in S13 comprises: S131, physiological data alignment: the physiological data acquisition delay is corrected through a linear compensation model, the original time is aligned to the quantum time reference using device calibration coefficients, and millisecond-level synchronization is realized; S132, behavior data alignment: for the periodic time error in behavior data, a sinusoidal function correction method is used for second-level alignment to eliminate time deviation caused by environmental interference and device clock offset; S133, health status data alignment: health status data is aligned by calendar time segmentation, and corrected by time slice mapping and medical system working time offset. 4.The Web-based big data driven stroke risk prediction method of claim 1, wherein, The causal path identification in S2 comprises: S21, causal feature space construction: the output aligned multi-scale data is standardized, and a three-dimensional causal feature tensor is constructed; S22, mixed causal structure learning: static causal relationships between variables are discovered through a constraint-based Bayesian network, and time series causality between variables is verified through a time series Granger causality test; S23, false association elimination: by counterfactual causal strength weakly related or spurious paths, only keeping the causal edges that satisfy the causal edge, and at the same time, the environmental variable control test; S24, Causal subgraph generation: based on the time lag range of the causal edge, a multi-layer causal network structure is constructed, specifically including: First tier: physiology physiological causal edge, delay ; Second layer: Behavior physiological cause-effect edge, ; Third tier: Health status Behavioral / physiological cause-effect edges, ; The weight of each causal edge in the multi-layer causal network structure is calculated based on the normalized causal strength and Granger test strength; S25, confidence rating: fuse counterfactual causal strength, temporal Granger causality test results and causal structure stability to generate confidence score and hierarchical evaluation, when the confidence rating is A level, when the confidence rating is B level, when the confidence rating is C level. 5.The Web-based big data-driven stroke risk prediction method of claim 4, wherein, The hybrid causal structure learning in S22 includes: S221, static causal discovery: constraint-based Bayesian network is used for learning the structure of causal graph, the causal relationship between variables is identified by calculating conditional mutual information, and the sparsity control coefficient is used optimization; S222, dynamic causal verification: adopt Granger causality test to verify the causal path, calculate F test statistic and screen out the path with time series causal relationship.

6. The web-based big data driven stroke risk prediction method of claim 5, wherein, The feature fusion modeling in S3 includes: S31, Parallel extraction of multi-modal features: a dual-channel architecture is used to extract multi-modal features in parallel, real-time fluctuation features are processed using a leaky integral firing neuron model through a pulse neural network channel, and trend features are extracted using a 6-layer residual dilation convolutional network through a time convolution network channel; S32, Dynamic weight adaptive fusion: a health state vector is extracted from the health state data, a dynamic fusion weight of the output of the pulse neural network channel and the time convolution network channel is generated using a multi-layer perceptron, and a final risk representation is generated by combining the calculated dynamic fusion weight.

7. The web-based big data driven stroke risk prediction method of claim 6, wherein, The parallel extraction of multi-modal features in S31 includes: S311, Real-time fluctuation modeling: synapse connections are initialized based on the edge weight of the causal subgraph, membrane potential is updated through a leaky integral firing neuron model, and an adaptive threshold is set to determine whether the neuron fires a pulse to capture fluctuation features; S312, Trend feature extraction: a 6-layer residual dilation convolutional network is configured, each layer uses a dilation convolution operation to extract trend information at different scales, and a jump connection mechanism is used to combine the output of each convolution layer with the input to enhance feature representation capability. 8.The Web-based big data-driven stroke risk prediction method of claim 7, wherein, The dynamic weight adaptive fusion in S32 includes: S321, Health state spatio-temporal coding: features of the health state data are extracted through multi-scale convolution, different time scales of information are captured using dilation convolution, and a health state vector is generated through pooling to extract periodic health patterns; S322, Gating weight generation: a multi-layer perceptron is used to generate a dynamic fusion weight of the output of the pulse neural network channel and the time convolution network channel, the weight is calculated by combining the health state vector through gating operation, and the weight is adjusted using a time period modulation factor; S323, Risk representation synthesis: the output feature maps of the pulse neural network channel and the time convolution network channel are weighted and fused according to the adjusted dynamic weight to generate the final risk representation feature. 9.The Web-based big data-driven stroke risk prediction method of claim 8, wherein, The risk evolution simulation in S4 includes: S41, Virtual patient environment construction: the generated final risk representation feature is mapped to the environment state of the virtual patient, high-dimensional features are mapped to a 512-dimensional space through autoencoder dimension reduction, and constraints are set for multi-dimensional intervention measures through a multi-dimensional intervention strategy space including drug dosage, rehabilitation intensity and diet scheme; S42, Causal transfer function modeling: a neural differential equation model is constructed to describe the causal relationship of state transfer in the virtual patient environment, a state transfer equation is defined to simulate random physiological fluctuations, the change of health state over time is analyzed by combining the causal transfer function, and the effect of each intervention measure is evaluated through a composite reward function, including risk change, intervention cost and compliance; S43, policy optimization and simulation: generate behavior policy using causal-guided policy gradient algorithm, optimize intervention policy by running multiple instances in parallel simulation, update patient's health status until simulation time reaches the set month; S44, risk trajectory generation: perform kernel density estimation on simulation results, analyze the evolution trajectory of risk, calculate the risk distribution at each time point, and extract the risk threshold crossing event to mark the key risk change moment. 10.The Web-based big data-driven stroke risk prediction method of claim 9, wherein, The intervention scheme generation in S5 includes: S51, analyze risk changes: according to the results of risk evolution simulation, extract the risk change and the key risk change moment; S52, trigger three-dimensional early warning display: when the risk change is detected to exceed 0.2 or reach the key risk change moment, automatically trigger the three-dimensional early warning display, and real-time display the patient's risk state through three-dimensional visualization graphics; S53, generate personalized intervention scheme: according to the results of three-dimensional early warning display, combined with the personalized information of the patient, generate a personalized intervention scheme, including drug treatment, rehabilitation training, and diet adjustment.

Images