A system for modeling the course trend of subarachnoid hemorrhage by integrating multi-source data

Abstract

This invention relates to the field of disease progression trend modeling technology, specifically a system for modeling the progression trend of subarachnoid hemorrhage by integrating multi-source data. This invention simultaneously collects and monitors four types of signals: intracranial pressure, blood flow velocity, cerebrospinal fluid pressure, and electroencephalogram (EEG). Instantaneous phase is extracted using Hilbert transform, and the time window is adaptively adjusted based on individual cerebrovascular conduction delay characteristics to calculate the phase-locking index between the three pairs of signals. A multivariate coupled oscillator model is established, and the topological invariant features of the phase trajectory are extracted. A comprehensive trend index is generated through tensor fusion. An individualized four-dimensional phase entropy baseline model is established, employing a two-layer early warning mechanism: a direct warning is issued when all three phase-locking indices show consecutive sign flips in their second derivatives; a comprehensive trend index confirmation is required when any two phase-locking indices flip. The system predicts the state level, trend level, and expected evolution trajectory based on a support vector regression model. This invention achieves accurate prediction and early warning of disease progression, providing a basis for clinical decision-making.

Description

Technical Field

[0001] This invention relates to the field of disease course trend modeling technology, specifically to a subarachnoid hemorrhage disease course trend modeling system that integrates multi-source data. Background Technology

[0002] Subarachnoid hemorrhage is a common and critical neurosurgical condition characterized by rapid and unpredictable changes in its course. Early and accurate prediction of the disease's progression is crucial for improving the prognosis of monitored patients. Traditional single-parameter monitoring methods struggle to capture the complexity of pathophysiological changes in a timely manner, necessitating more precise multi-source data fusion technologies.

[0003] Existing subarachnoid hemorrhage monitoring technologies have the following main shortcomings: Single-parameter monitoring has significant limitations, relying solely on single indicators such as intracranial pressure or blood flow velocity, which cannot comprehensively reflect changes in multi-system coupling and results in insufficient early warning sensitivity; There is a lack of individualized parameter adjustment mechanisms, with different monitoring subjects using the same monitoring standards and thresholds, ignoring individual differences in cerebrovascular physiological characteristics; Traditional methods often employ simple threshold judgments, lacking in-depth analysis of phase coupling relationships and high-dimensional topological features between physiological systems, making it difficult to extract the essential laws governing state evolution; The early warning time window is short, often only identifying the state trend of the monitored subject after clinical symptoms appear, missing the optimal intervention opportunity. Summary of the Invention

[0004] This invention provides a system for modeling the course trend of subarachnoid hemorrhage by integrating multi-source data, enabling accurate prediction of the condition and early warning, and improving the scientific nature of decision-making.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A system for modeling the course trend of subarachnoid hemorrhage by integrating multi-source data includes:

[0007] The multi-source signal acquisition module is used to simultaneously acquire intracranial pressure signals, blood flow velocity signals, cerebrospinal fluid pressure signals, and electroencephalogram (EEG) signals of subjects undergoing subarachnoid hemorrhage monitoring.

[0008] The phase-locking calculation module is used to extract the instantaneous phase of the four types of signals by performing Hilbert transform on them respectively. It adaptively adjusts the calculation time window of the phase-locking index according to the individual cerebrovascular conduction delay characteristics of the monitored object, and calculates the phase-locking index between intracranial pressure-blood flow, blood flow-cerebrospinal fluid pressure, and cerebrospinal fluid pressure-EEG signal respectively.

[0009] The tensor feature fusion module is used to establish a multi-element coupled oscillator model, extract the trajectory topological invariant features formed by the phases of four types of signals in high-dimensional space, and perform tensor fusion transformation on the phase locking index and topological invariant features to generate a comprehensive trend index.

[0010] The dual-layer early warning module is used to establish an individualized four-dimensional phase entropy baseline model. It adopts a dual-layer early warning mechanism: when all three phase locking indices show continuous sign reversal of the second derivative, it is directly identified as an abnormal state event; when any two reversals occur, it needs to be confirmed by a secondary early warning through a comprehensive trend indicator that deviates from the phase entropy baseline threshold.

[0011] The trend prediction output module is used to analyze the occurrence patterns and evolution characteristics of abnormal state events. It establishes a state trend prediction model based on the support vector regression model and outputs the current state level, trend level, and expected evolution trajectory.

[0012] As a preferred technical solution of the present invention, the synchronous acquisition specifically includes: implanting a pressure sensor in the brain of the monitored object to obtain intracranial pressure signal, using a transcranial Doppler ultrasound probe to detect the middle cerebral artery to obtain blood flow velocity signal, inserting a pressure catheter through lumbar puncture to obtain cerebrospinal fluid pressure signal, and using multi-lead electroencephalogram electrodes attached to the scalp to obtain electroencephalogram signals. The four types of sensors are connected to a unified multi-channel data acquisition system for synchronous digital processing.

[0013] As a preferred technical solution of the present invention, the adaptive adjustment based on the individual cerebrovascular conduction delay characteristics of the monitored object includes: calculating the cross-correlation function of intracranial pressure signal and blood flow velocity signal at different time offsets, determining the cerebrovascular conduction delay time by the time offset corresponding to the maximum value of the correlation coefficient, and setting an integer multiple of the delay time as the sliding time window length calculated by the phase locking index.

[0014] As a preferred embodiment of the present invention, the phase-locked index calculation step includes: calculating the phase difference between the instantaneous phases of each pair of signals, converting the phase difference sequence into a phase difference vector in complex exponential form, averaging the phase difference vector over time within the sliding time window to obtain an average complex vector, and calculating the magnitude of the average complex vector as the phase-locked index value of the corresponding signal pair.

[0015] As a preferred embodiment of the present invention, the establishment of the multi-element coupled oscillator model includes:

[0016] Using the instantaneous phases of the four types of signals as coordinate axes in a four-dimensional phase space, a four-dimensional phase vector sequence is constructed to form an initial trajectory point set in the four-dimensional space.

[0017] The phase space of the four-dimensional phase vector sequence is reconstructed using delayed embedding technology. The embedding dimension is determined by the pseudo nearest neighbor method, and the delay time parameter is determined by the mutual information method. The initial trajectory point set is reconstructed into a high-dimensional embedding space trajectory, and a multi-element coupled oscillator model describing the phase coupling relationship of four types of signals is established.

[0018] As a preferred embodiment of the present invention, the step of generating the comprehensive trend index includes:

[0019] The phase-locked indexes are used to form the first eigenvector, and the topological invariant features are used to form the second eigenvector.

[0020] Perform a tensor outer product operation on two eigenvectors to form a second-order tensor matrix, and then perform singular value decomposition on the tensor matrix;

[0021] The largest singular value is selected as the principal component, and combined with the norm information of the corresponding left and right singular vectors, a comprehensive trend index is constructed and output through weighted combination.

[0022] As a preferred technical solution of the present invention, the establishment of an individualized four-dimensional phase entropy baseline mode includes: collecting phase data of four types of signals during a continuous period of stable condition of the monitored object, extracting four instantaneous phase values ​​corresponding to each moment to form a four-dimensional phase vector, calculating the joint probability density distribution of the four-dimensional phase vector sequence, calculating the phase entropy value according to the joint probability distribution, and using the mean and standard deviation of the phase entropy value during the stable period as the individualized baseline parameters of the monitored object.

[0023] As a preferred embodiment of the present invention, the determination of the continuous sign reversal of the second derivative includes:

[0024] A first-order difference operation is performed on each phase-locking index to obtain a first-order derivative sequence, and a second-order difference operation is performed on the first-order derivative sequence to obtain a second-order derivative sequence.

[0025] The sign change of three consecutive sampling points in the second derivative sequence is detected. When the signs of the three consecutive points change alternately between positive and negative, it is determined that the phase-locked index has undergone continuous sign flipping of the second derivative.

[0026] As a preferred embodiment of the present invention, the occurrence mode and evolution characteristics of the analyzed abnormal events include:

[0027] The magnitude and rate of change of the phase-locking index before the occurrence of an abnormal statistical event;

[0028] Calculate the evolution trajectory parameters of topological invariant features during the occurrence of risk events, including the gradient of the Betty number change, the evolution rate of the correlation dimension, and the fluctuation amplitude of the maximum Lyapunov exponent;

[0029] Extract the volatility characteristics of the comprehensive trend indicators before and after the risk event is triggered, including volatility variance, kurtosis coefficient and skewness coefficient;

[0030] The risk event feature vector is composed of the phase-locked index change magnitude and rate of change, topological invariant evolution trajectory parameters, and comprehensive trend indicator fluctuation characteristics.

[0031] As a preferred embodiment of the present invention, the establishment of a state trend prediction model based on a support vector regression model includes:

[0032] The risk event feature vector is used as the input feature of the support vector regression model, and the state level label and trend level label are used as the output label.

[0033] The radial basis function is used as the kernel function to train the support vector regression model. The regularization parameters and kernel function parameters of the model are optimized by cross-validation, and a mapping relationship from feature vectors to state prediction results is established.

[0034] The beneficial effects of this invention are:

[0035] 1. This invention combines Hilbert phase extraction of four types of physiological signals with individual cerebrovascular conduction delay characteristics, and achieves individualized multi-system phase coupling monitoring by adaptively adjusting the phase-locking index calculation time window. This technology overcomes the limitations of traditional single-parameter monitoring, can capture deep coordination relationships between physiological systems, and significantly improves the early identification capability of the monitored object's state.

[0036] 2. This invention fuses the topological invariant features of a multi-element coupled oscillator model with the phase-locking index through tensor outer product, and generates a comprehensive trend index through singular value decomposition. This deep fusion method of multimodal features can comprehensively characterize the dynamic evolution of the four-dimensional phase space, providing a more sensitive and stable basis for judging state trends than traditional methods.

[0037] 3. This invention establishes a two-layer early warning system based on an individualized four-dimensional phase entropy baseline. By combining second-derivative continuous sign flip detection and comprehensive trend index deviation analysis, a balance between high sensitivity and high specificity is achieved. The organic integration of this mechanism with a support vector regression prediction model not only enables advanced early warning but also outputs quantified state levels, trend levels, and expected evolution trajectories, providing forward-looking decision support for precise clinical intervention. Attached Figure Description

[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0039] Figure 1 This is a schematic diagram of the structure of a subarachnoid hemorrhage disease progression trend modeling system that integrates multi-source data according to the present invention;

[0040] Figure 2 This is a flowchart illustrating the process of a subarachnoid hemorrhage disease progression modeling system that integrates multi-source data, as per the present invention. Detailed Implementation

[0041] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0042] Example 1: As Figure 1 As shown, the present invention provides a subarachnoid hemorrhage disease progression trend modeling system that integrates multi-source data, comprising:

[0043] The multi-source signal acquisition module is used to simultaneously acquire intracranial pressure signals, blood flow velocity signals, cerebrospinal fluid pressure signals, and electroencephalogram (EEG) signals of subjects undergoing subarachnoid hemorrhage monitoring.

[0044] Furthermore, the synchronous acquisition specifically includes: implanting a pressure sensor in the brain of the monitored subject to obtain intracranial pressure signals; using a transcranial Doppler ultrasound probe to detect the middle cerebral artery to obtain blood flow velocity signals; inserting a pressure catheter through lumbar puncture to obtain cerebrospinal fluid pressure signals; and using multi-lead electroencephalogram electrodes attached to the scalp to obtain electroencephalogram signals. The four types of sensors are connected to a unified multi-channel data acquisition system for synchronous digital processing.

[0045] Specifically, in this embodiment, four different sensors are used to simultaneously collect physiological signals from subjects monitoring for subarachnoid hemorrhage. A minimally invasive pressure sensor is implanted intracranially to acquire intracranial pressure signals. This sensor is inserted through a burr hole into the skull, either epidurally or within the brain parenchyma, to monitor intracranial pressure changes in real time. A transcranial Doppler ultrasound probe is positioned at the temporal window to detect blood flow in the middle cerebral artery and acquire blood flow velocity signals. The ultrasound probe measures the frequency shift caused by red blood cell movement using the Doppler effect. A pressure catheter is inserted into the subarachnoid space via lumbar puncture at the L3-L4 intervertebral space to acquire cerebrospinal fluid pressure signals. Multi-lead electroencephalography (EEG) electrodes arranged in a standard 10-20 system are attached to the scalp surface to acquire EEG signals.

[0046] The four types of sensors are connected to a unified multi-channel data acquisition system via dedicated interfaces. This system uses a unified clock source to ensure time synchronization of signals from each channel. The system is set to a sampling frequency of 1000Hz and uses a 16-bit A / D converter for digital processing. To ensure signal quality, the intracranial pressure sensor uses a differential amplifier circuit, the blood flow velocity signal is filtered by a bandpass filter to remove motion artifacts, the cerebrospinal fluid pressure catheter is equipped with a pressure transmitter to achieve analog output, and the electroencephalogram (EEG) signal uses a high input impedance preamplifier to reduce the influence of electrode contact impedance.

[0047] The four types of raw signals acquired were first subjected to baseline drift correction, and a high-pass filter was used to remove DC components below 0.1 Hz. A 50 Hz notch filter was used to remove power frequency interference for intracranial pressure and cerebrospinal fluid pressure signals. A 2–100 Hz bandpass filter was used for blood flow velocity signals to retain effective spectral components, and a 0.5–45 Hz bandpass filter was used for EEG signals to remove electromyography artifacts and high-frequency noise. The preprocessed signals were stored with a unified timestamp, providing a standardized data foundation for subsequent phase analysis.

[0048] The key to this synchronous acquisition scheme lies in ensuring the precise correspondence of different physiological parameters in the time domain through a unified clock source and standardized signal conditioning circuits. This provides a reliable data foundation for subsequent phase-locked index calculation and multivariate coupling analysis.

[0049] The phase-locking calculation module is used to extract the instantaneous phase of the four types of signals by performing Hilbert transform on them respectively. It adaptively adjusts the calculation time window of the phase-locking index according to the individual cerebrovascular conduction delay characteristics of the monitored object, and calculates the phase-locking index between intracranial pressure-blood flow, blood flow-cerebrospinal fluid pressure, and cerebrospinal fluid pressure-EEG signal respectively.

[0050] Furthermore, the adaptive adjustment based on the individual cerebrovascular conduction delay characteristics of the monitored object includes: calculating the cross-correlation function of intracranial pressure signal and blood flow velocity signal at different time offsets, determining the cerebrovascular conduction delay time by using the time offset corresponding to the maximum value of the correlation coefficient, and setting an integer multiple of the delay time as the sliding time window length calculated by the phase-locking index.

[0051] Specifically, Hilbert transforms are performed on the four types of preprocessed signals to construct analytic signals and obtain instantaneous phase information. Let the original signal be... Its Hilbert transformation is Then analyze the signal Instantaneous phase This method allows for the extraction of intracranial pressure signal phases. Blood flow velocity signal phase Cerebrospinal fluid pressure signal phase Phase with EEG signals .

[0052] The key to this invention lies in adaptively adjusting the calculation window based on the individual cerebral vascular conduction delay characteristics of the monitored subject. First, the time offset between intracranial pressure and blood flow velocity signals is calculated. The cross-correlation function is as follows:

[0053] ;

[0054] By finding the maximum correlation coefficient Corresponding time offset Determining individual cerebral vascular conduction delay time This delay time reflects the physiological conduction characteristics of an individual's cerebrovascular system and exhibits significant individual differences.

[0055] Individual conduction delay time As a benchmark, the length of the sliding time window for calculating the phase-locking index is set. ,in The factor is an integer multiple, typically ranging from 3 to 5, to ensure the window contains sufficient physiological cycle information. This adaptive adjustment mechanism ensures the window contains sufficient physiological cycle information, giving phase-locked analysis of different monitored objects individualized characteristics and improving the accuracy of state trend judgment.

[0056] Further, the phase-locked index calculation step includes: calculating the phase difference between the instantaneous phases of each pair of signals, converting the phase difference sequence into a phase difference vector in complex exponential form, averaging the phase difference vector over time within the sliding time window to obtain an average complex vector, and calculating the magnitude of the average complex vector as the phase-locked index value of the corresponding signal pair.

[0057] Specifically, intracranial pressure and blood flow were calculated separately. Blood flow-cerebrospinal fluid pressure Cerebrospinal fluid pressure-electroencephalogram (EEG) The phase-locking index between three pairs of signals. For each pair of signals, the instantaneous phase difference is first calculated. The phase difference sequence is converted into a phase difference vector in complex exponential form. In the adaptive time window The phase difference vector is averaged over time, and the magnitude of the averaged complex vector is used as the phase-locking index.

[0058] ;

[0059] Where j is the imaginary unit, used to construct the complex number representation; Indicates within the time window Time averaging operation within the period, This represents the complex modulus. Specifically, the phase-locked index is calculated based on the vector averaging principle in the complex domain. It represents the phase difference between two signals. Converted to a complex point on the unit circle using Euler's formula. The real part is The imaginary part is By averaging these complex points within a sliding time window, the average complex vector is obtained. The magnitude of the average complex vector reflects the consistency of the phase difference within the time window: when the phase difference remains stable, the complex points cluster on the unit circle, and the magnitude of the averaged vector is close to 1; when the phase difference changes randomly, the complex points are scattered, and the magnitude of the averaged vector is close to 0. The value ranges from 0 to 1. The closer the value is to 1, the stronger the phase coupling between the two signals, and the better the synchronicity between the corresponding physiological systems.

[0060] The time series of the three phase-locked indexes are obtained by continuous calculation using a sliding window method. This provides quantitative indicators for subsequent monitoring and early warning of the status of objects.

[0061] The tensor feature fusion module is used to establish a multi-element coupled oscillator model, extract the trajectory topological invariant features formed by the phases of four types of signals in high-dimensional space, and perform tensor fusion transformation on the phase locking index and topological invariant features to generate a comprehensive trend index.

[0062] Furthermore, the establishment of the multi-element coupled oscillator model includes:

[0063] Using the instantaneous phases of the four types of signals as coordinate axes in a four-dimensional phase space, a four-dimensional phase vector sequence is constructed to form an initial trajectory point set in the four-dimensional space.

[0064] The phase space of the four-dimensional phase vector sequence is reconstructed using delayed embedding technology. The embedding dimension is determined by the pseudo nearest neighbor method, and the delay time parameter is determined by the mutual information method. The initial trajectory point set is reconstructed into a high-dimensional embedding space trajectory, and a multi-element coupled oscillator model describing the phase coupling relationship of four types of signals is established.

[0065] Specifically, the instantaneous phase of the four types of signals Using the coordinate axes of the four-dimensional phase space, construct a sequence of four-dimensional phase vectors:

[0066] ;

[0067] This represents the instantaneous phase of the intracranial pressure signal. The instantaneous phase of the blood flow velocity signal; This refers to the instantaneous phase of the cerebrospinal fluid pressure signal. These are instantaneous phases of the EEG signal; all phases are in radians, and the range is... ; This indicates the transpose operation.

[0068] The phase space reconstruction method maps four types of signals into a set of trajectory points in four-dimensional space. ,in This represents the total number of sampling points.

[0069] The phase space trajectory is reconstructed using delayed embedding techniques, and an embedding vector is constructed:

[0070] ;

[0071] in For the embedding dimension, The time delay parameter is used. The optimal embedding dimension is determined using the pseudo nearest neighbor method. The criterion is the minimum dimension when the proportion of false nearest neighbors is lower than a set threshold. The delay time parameter is determined using a mutual information function. :

[0072] ;

[0073] The mutual information function determines the optimal delay parameter by statistically analyzing the dependency between two time series at different delay times. The calculation process includes: first, dividing the numerical range of each phase series into several equal intervals, statistically analyzing the frequency of the sample occurrence in each interval combination, and calculating the joint probability distribution and marginal probability distribution; then, calculating the information entropy difference according to the mutual information formula. Mutual information is 0 when the two sequences are independent, and reaches a local extremum when the strongest nonlinear dependency exists; the delay time corresponding to the first local minimum of the mutual information function is selected as the optimal embedding parameter to ensure the effectiveness of phase space reconstruction. For time delay The joint probability density under, This represents the marginal probability density.

[0074] Furthermore, the step of generating the comprehensive trend indicator includes:

[0075] The phase-locked indexes are used to form the first eigenvector, and the topological invariant features are used to form the second eigenvector.

[0076] Perform a tensor outer product operation on two eigenvectors to form a second-order tensor matrix, and then perform singular value decomposition on the tensor matrix;

[0077] The largest singular value is selected as the principal component, and combined with the norm information of the corresponding left and right singular vectors, a comprehensive trend index is constructed and output through weighted combination.

[0078] Specifically, based on the reconstructed phase space trajectory, topological invariant features describing the phase coupling relationships of four types of signals are calculated. The topological features of the trajectory are extracted using continuous cohomology theory, and the Betti number sequence is calculated. Characterization The number of dimensional topological holes, of which Indicates the number of connected components, Represents the number of one-dimensional holes, Represents the number of two-dimensional cavities, This represents the number of three-dimensional holes.

[0079] Calculate the correlation dimension of the trajectory:

[0080] ;

[0081] in, The integral function is defined as the correlation function, where r is the Euclidean distance threshold, specifically the threshold for determining the Euclidean distance between any two trajectory points in the phase space, ranging from 0.01 to 0.5 times the data standard deviation. The Lyapunov exponent spectrum of the trajectory is extracted. The chaotic properties of quantification systems, the maximum Lyapunov exponent It reflects the speed of trajectory divergence.

[0082] The three phase-locking indices are combined to form the first eigenvector:

[0083] ;

[0084] The topological invariant features are combined to form the second eigenvector:

[0085] ;

[0086] Perform a tensor outer product operation on two eigenvectors to form a second-order tensor matrix. ,in This indicates a tensor outer product operation, which yields a 3×4 dimensional tensor matrix.

[0087] For tensor matrices Implement singular value decomposition ,in It is a left singular matrix. It is a right singular matrix. It is a singular value diagonal matrix.

[0088] Selecting the maximum singular value As the principal component of the comprehensive trend indicator, the comprehensive trend indicator is constructed by combining the norm information of the left and right singular vectors:

[0089] ;

[0090] in The left singular vector corresponding to the maximum singular value. For the corresponding right singular vector, and For normalized weight coefficients, satisfying The optimal weight ratio is determined through statistical analysis of the training data.

[0091] This comprehensive trend index integrates phase coupling strength information and topological features, and achieves the optimal combination of multimodal features through tensor decomposition. The changes in the values ​​reflect the overall trend of coordination among multiple physiological systems, providing a more sensitive and reliable early warning signal for the status of the monitored object than a single indicator.

[0092] The dual-layer early warning module is used to establish an individualized four-dimensional phase entropy baseline model. It adopts a dual-layer early warning mechanism: when all three phase locking indices show continuous sign reversal of the second derivative, it is directly identified as an abnormal state event; when any two reversals occur, it needs to be confirmed by a secondary early warning through a comprehensive trend indicator that deviates from the phase entropy baseline threshold.

[0093] Furthermore, the establishment of the individualized four-dimensional phase entropy baseline model includes: collecting phase data of four types of signals during a continuous period of stable condition of the monitored object, extracting four instantaneous phase values ​​corresponding to each moment to form a four-dimensional phase vector, calculating the joint probability density distribution of the four-dimensional phase vector sequence, calculating the phase entropy value based on the joint probability distribution, and using the mean and standard deviation of the phase entropy value during the stable period as the individualized baseline parameters of the monitored object.

[0094] Specifically, data is collected during a continuous period of stable disease in the monitored subjects. Phase data of four types of signals within a typical 24–48 hour period are used to construct a four-dimensional phase vector sequence in the S300. The four-dimensional phase space is divided into a uniform grid. The frequency of occurrence of the phase vector in each grid region is counted, and the joint probability density distribution of the four-dimensional phase vector sequence is calculated. .

[0095] Calculate the phase entropy value based on the joint probability distribution:

[0096] ;

[0097] in Represents a four-dimensional grid index; For the joint probability density, maximum likelihood estimation is used, and Laplace smoothing is applied to the empty grid. The smoothing parameter is... .

[0098] The mean value of the phase entropy during the stable period of the monitored object and standard deviation Establish baseline thresholds as individualized baseline parameters. This baseline pattern reflects the phase coordination characteristics of an individual's brain physiological system under steady-state conditions, exhibiting significant individual variability.

[0099] Furthermore, the determination of the continuous sign flipping of the second derivative includes:

[0100] A first-order difference operation is performed on each phase-locking index to obtain a first-order derivative sequence, and a second-order difference operation is performed on the first-order derivative sequence to obtain a second-order derivative sequence.

[0101] The sign change of three consecutive sampling points in the second derivative sequence is detected. When the signs of the three consecutive points change alternately between positive and negative, it is determined that the phase-locked index has undergone continuous sign flipping of the second derivative.

[0102] Specifically, for each phase-locking index The first derivative sequence is obtained by performing first-order difference operations. Then, perform a second-order difference operation on the first-order derivative sequence to obtain the second-order derivative sequence. Detecting three consecutive sampling points in a second derivative sequence. The sign change occurs when the following condition is met:

[0103] ;

[0104] The phase-locked index was determined to have experienced a continuous sign flip of its second derivative.

[0105] Establish a tiered early warning judgment logic and set a flip counter. Count the number of phase-locked exponents that simultaneously exhibit continuous sign flipping of the second derivative. When At that time, that is, the three phase-locked indexes If all events reverse, a Level 1 warning is triggered, indicating an abnormal status event. At that time, the Level 2 early warning verification mechanism is activated to calculate the current comprehensive trend indicators. Deviation from baseline threshold .

[0106] When satisfied And the duration exceeds the preset threshold When the condition is confirmed as an abnormal event, this two-tiered mechanism combines the high sensitivity of the primary warning with the high specificity of the secondary warning, ensuring timely identification of acute conditions while reducing the false positive warning rate.

[0107] The trend prediction output module is used to analyze the occurrence patterns and evolution characteristics of abnormal state events. It establishes a state trend prediction model based on the support vector regression model and outputs the current state level, trend level, and expected evolution trajectory.

[0108] Furthermore, the occurrence patterns and evolutionary characteristics of the analyzed abnormal events include:

[0109] The magnitude and rate of change of the phase-locking index before the occurrence of an abnormal statistical event;

[0110] Calculate the evolution trajectory parameters of topological invariant features during the occurrence of risk events, including the gradient of the Betty number change, the evolution rate of the correlation dimension, and the fluctuation amplitude of the maximum Lyapunov exponent;

[0111] Extract the volatility characteristics of the comprehensive trend indicators before and after the risk event is triggered, including volatility variance, kurtosis coefficient and skewness coefficient;

[0112] The risk event feature vector is composed of the phase-locked index change magnitude and rate of change, topological invariant evolution trajectory parameters, and comprehensive trend indicator fluctuation characteristics.

[0113] Specifically, the characteristics of phase-locking index changes before the occurrence of statistical anomalies are analyzed, and the magnitude and rate of change of each phase-locking index are calculated:

[0114] ;

[0115] ;

[0116] in .

[0117] Calculate the evolutionary trajectory parameters of topological invariant features during the occurrence of risk events, including the gradient of Betti number changes. , Evolution speed of correlation dimension and the maximum fluctuation range of the Lyapunov index .

[0118] Analysis of comprehensive trend indicators Calculate the volatility variance based on the volatility characteristics before and after the risk event is triggered. Kurtosis coefficient Skewing coefficient :

[0119] ;

[0120] ;

[0121] ;

[0122] in, The average value of a comprehensive trend indicator over a specific time window; This represents the mathematical expectation operation. The above statistical parameters and feature parameters are combined to form a risk event feature vector:

[0123] .

[0124] Furthermore, the state trend prediction model based on the support vector regression model includes:

[0125] The risk event feature vector is used as the input feature of the support vector regression model, and the state level label and trend level label are used as the output label.

[0126] The radial basis function is used as the kernel function to train the support vector regression model. The regularization parameters and kernel function parameters of the model are optimized by cross-validation, and a mapping relationship from feature vectors to state prediction results is established.

[0127] Specifically, the risk event feature vector As input features to the support vector regression model, a nonlinear mapping relationship is established from the feature space to the prediction result:

[0128] ;

[0129] in , It is a Lagrange multiplier. For the number of support vectors, For kernel function, This is a bias term.

[0130] Radial basis functions are used as kernel functions. ,in For kernel function parameters, Let represent the Euclidean norm. The regularization parameter of the model is optimized using cross-validation. and kernel function parameters A grid search algorithm is used in the parameter space. Finding the optimal parameter combination This makes the mean squared error of cross-validation... Reaching the minimum value:

[0131] ;

[0132] Where Q is the cross-validation fold number. For the first Sample size For predicted values, This is a real label.

[0133] Establish status level labels This indicates five stages: acute phase, subacute phase, recovery phase, stable phase, and worsening phase, with trend level labels. The risk levels are categorized into four levels: no risk, low risk, medium risk, and high risk. A multi-output support vector regression framework is used to train separate state level predictors. and risk level predictor .

[0134] Based on the current feature vector Based on historical evolution patterns, predict future time windows. The evolution of the internal composite trend indicator:

[0135] ;

[0136] in To predict the time step, it outputs three prediction results: current state level, trend level, and expected evolution trajectory, providing a quantitative assessment of state and trend for clinical decision-making.

[0137] This embodiment achieves deep fusion analysis of multi-source physiological signals through the organic combination of the above five steps. First, the adaptive time window adjustment based on individual cerebrovascular conduction delay characteristics overcomes the limitations of traditional fixed parameter settings. Second, tensor fusion transformation of topological invariant features and phase-locked index achieves deep fusion of multimodal features. Finally, the established two-layer early warning mechanism combined with an individualized four-dimensional phase entropy baseline solves the technical challenge of balancing sensitivity and specificity in traditional early warning methods. This method can identify the risk status of monitored subjects in advance, providing prospective support for clinical decision-making.

[0138] Example 2: A neurosurgery department in a hospital found that traditional single-parameter monitoring methods had a lag in providing early warning of the condition of subjects with subarachnoid hemorrhage. To improve the predictability and accuracy of intensive care, the department introduced a method such as... Figure 2 The present invention provides a subarachnoid hemorrhage disease progression trend modeling system that integrates multi-source data, and performs retrospective analysis on monitoring data of previously qualified monitoring subjects.

[0139] This validation employed a retrospective data analysis method. The study selected 25 cases with complete multimodal monitoring records from the subarachnoid hemorrhage monitoring subjects admitted to the Department of Neurosurgery of this hospital between January 2022 and December 2023.

[0140] Typical Case: A 56-year-old female subject was admitted to the hospital due to subarachnoid hemorrhage caused by a ruptured anterior communicating artery aneurysm. During her hospitalization, she underwent intracranial pressure monitoring, transcranial Doppler blood flow monitoring, cerebrospinal fluid pressure monitoring, and continuous electroencephalography (EEG) monitoring.

[0141] First, individualized parameters were established. The system determined that the cerebral vascular conduction delay for this monitored subject was 134ms, set an adaptive window of 536ms, and established a phase entropy baseline. =2.28, threshold range [1.52, 3.04]. The system continuously monitored, and the phase lock index began to fluctuate on the 3rd day. It decreased from 0.78 to 0.65. The PLV index decreased from 0.74 to 0.58. At 3:15 AM on the fourth day, the second derivatives of both PLV indices reversed, and the CTI dropped from 2.31 to 1.28, deviating from the baseline by 2.7 times the standard deviation, prompting the system to issue a warning. The model predicted the current stage to be "worsening" and at a "medium risk" level, with symptom changes expected within 6–10 hours. In reality, the monitored subject developed clinical signs such as decreased level of consciousness 11.25 hours later.

[0142] A typical case received an early warning 11.25 hours in advance. Of the 25 monitored subjects, 18 experienced changes in their condition, with an average risk assessment time of 9.2 hours. The individual transmission delay ranged from 97 to 189 ms, with an average of 138 ± 24 ms, demonstrating significant individual variability. Individualized time window settings more accurately reflect the physiological characteristics of the monitored subjects. The dual-layer early warning mechanism effectively identified abnormal events, and the predicted time window aligned with the actual trend of condition evolution.

[0143] This embodiment, through retrospective data verification, demonstrates the technical advantages of the method of the present invention, including timely early warning, precise individualized adaptation, multimodal intelligent fusion, and forward-looking decision support. This method can identify risk conditions on average in advance, with a prediction accuracy exceeding 90%, providing an effective technical solution for precise monitoring and individualized treatment of subarachnoid hemorrhage patients, and has significant clinical application value.

[0144] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A subarachnoid hemorrhage disease course trend modeling system that fuses multi-source data, the system comprising: The method comprises the following steps: A multi-source signal acquisition module is used to synchronously acquire intracranial pressure signals, blood flow velocity signals, cerebrospinal fluid pressure signals and electroencephalogram signals of a subarachnoid hemorrhage monitoring object; A phase locking calculation module is used to perform Hilbert transform on the four types of acquired signals to extract respective instantaneous phases, adaptively adjust the calculation time window of the phase locking index according to the individual cerebral vascular transmission delay characteristics of the monitoring object, and calculate the phase locking indexes between the intracranial pressure-blood flow, blood flow-cerebrospinal fluid pressure and cerebrospinal fluid pressure-electroencephalogram signals, respectively; A tensor feature fusion module is used to establish a multi-element coupled oscillator model, extract the trajectory topological invariant features formed by the phases of the four types of signals in a high-dimensional space, perform tensor fusion transformation on the phase locking indexes and the topological invariant features, and generate a comprehensive trend index; A double-layer early warning module is used to establish an individualized four-dimensional phase entropy baseline mode, and a double-layer early warning mechanism is adopted: when the second-order derivatives of all three phase locking indexes continuously change in sign, a state abnormal event is directly identified; when any two change in sign, a secondary early warning confirmation is required through the comprehensive trend index deviating from the phase entropy baseline threshold; A trend prediction output module is used to analyze the occurrence mode and evolution characteristics of the state abnormal event, establish a state trend prediction model based on a support vector regression model, and output the current state level, trend level and expected evolution trajectory. The adaptive adjustment according to the individual cerebral vascular transmission delay characteristics of the monitoring object comprises the following steps: calculating the cross-correlation functions of the intracranial pressure signals and the blood flow velocity signals at different time offsets, determining the cerebral vascular transmission delay time corresponding to the time offset of the maximum correlation coefficient, and setting an integer multiple of the delay time as the sliding time window length for calculating the phase locking index; The establishment of the multi-element coupled oscillator model comprises the following steps: The instantaneous phases of the four types of signals are taken as the coordinate axes of a four-dimensional phase space to construct a four-dimensional phase vector sequence and form an initial trajectory point set in the four-dimensional space; The delay embedding technique is used to reconstruct the four-dimensional phase vector sequence in the phase space, the embedding dimension is determined by the nearest neighbor method, the delay time parameter is determined by the mutual information method, the initial trajectory point set is reconstructed into a high-dimensional embedded space trajectory, and a multi-element coupled oscillator model describing the coupling relationship of the phases of the four types of signals is established; The establishment of the individualized four-dimensional phase entropy baseline mode comprises the following steps: acquiring the phase data of the four types of signals of the monitoring object in a continuous time period during a stable period, extracting four instantaneous phase values corresponding to each time to form a four-dimensional phase vector, calculating the joint probability density distribution of the four-dimensional phase vector sequence, calculating the phase entropy value according to the joint probability distribution, and taking the mean value and standard deviation of the stable period phase entropy value as the individualized baseline parameters of the monitoring object.

2. The system of claim 1, wherein the system is configured to: The synchronous acquisition specifically comprises the following steps: implanting a pressure sensor in the skull of the monitoring object to acquire the intracranial pressure signals, using a transcranial Doppler ultrasound probe to detect the middle cerebral artery to acquire the blood flow velocity signals, inserting a pressure catheter through lumbar puncture to acquire the cerebrospinal fluid pressure signals, and using multi-lead electroencephalogram electrodes to adhere to the scalp to acquire the electroencephalogram signals; the four types of sensors are connected to a unified multi-channel data acquisition system for synchronous digital processing.

3. The system of claim 1, wherein the system is configured to: The phase-locked index calculation step comprises: calculating the phase difference between each pair of signal instantaneous phases, converting the phase difference sequence into a phase difference vector in complex exponential form, time-averaging the phase difference vector in a sliding time window to obtain an average complex vector, and calculating the modulus of the average complex vector as the phase-locked index value of the corresponding signal pair.

4. The system of claim 1, wherein the system is configured to: The generating comprehensive trend indicator step comprises: The phase-locked index is formed into a first feature vector, and the topological invariant feature is formed into a second feature vector; A tensor outer product operation is performed on the two feature vectors to form a second-order tensor matrix, and a singular value decomposition operation is performed on the tensor matrix; The largest singular value is selected as the principal component, and the norm information of the corresponding left singular vector and right singular vector is combined to construct a comprehensive trend indicator output through weighted combination.

5. The system of claim 1, wherein the system is configured to: The determination of the continuous sign flip of the second derivative comprises: A first-order difference operation is performed on each phase-locked index to obtain a first-order derivative sequence, and a second-order difference operation is performed on the first-order derivative sequence to obtain a second-order derivative sequence; The signs of three consecutive sampling points in the second-order derivative sequence are detected, and when the signs of the three consecutive points appear to alternate between positive and negative, it is determined that the phase-locked index has a continuous sign flip of the second derivative.

6. The system of claim 1, wherein the system is configured to: The analysis of the occurrence mode and evolution characteristics of the state abnormal event comprises: The amplitude and rate of change of the phase-locked index before the occurrence of the state abnormal event are counted; The evolution trajectory parameters of the topological invariant feature during the occurrence of the risk event are calculated, including the change gradient of the Betti number, the evolution speed of the correlation dimension, and the fluctuation amplitude of the maximum Lyapunov exponent; The fluctuation characteristics of the comprehensive trend indicator before and after the triggering of the risk event are extracted, including the fluctuation variance, kurtosis coefficient, and skewness coefficient; The amplitude and rate of change of the phase-locked index, the evolution trajectory parameters of the topological invariant, and the fluctuation characteristics of the comprehensive trend indicator are combined into a risk event feature vector.

7. The system of claim 1, wherein the system is configured to: The establishment of the state trend prediction model based on the support vector regression model comprises: The risk event feature vector is used as the input feature of the support vector regression model, and the state level label and trend level label are used as the output label; Radial basis function is used as the kernel function to train the support vector regression model, the regularization parameter and kernel function parameter of the model are optimized through cross-validation method, and the mapping relationship from the feature vector to the state prediction result is established.

Images