Multi-mode intracranial pressure intelligent early warning method and system

By simultaneously acquiring intracranial pressure waveforms, electroencephalograms, and transcranial Doppler blood flow velocity signals, extracting the time-domain, frequency-domain, and nonlinear characteristics of multimodal physiological data, and assigning dynamic weights, the problem of single-modal signals being unable to capture the multi-system correlation of intracranial pressure changes is solved, thus achieving accurate intracranial pressure prediction and early warning.

CN121003424AInactive Publication Date: 2025-11-25TONGJI HOSPITAL ATTACHED TO TONGJI MEDICAL COLLEGE HUAZHONG SCI TECH
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202511436087.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-10-09
Publication Date
2025-11-25
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In existing technologies, single-modal physiological signals cannot fully capture the multi-system interconnected mechanisms of intracranial pressure changes, resulting in limited dimensions for perceiving abnormal intracranial pressure trends, one-sided early warning basis, difficulty in identifying potential risks in advance, and a high likelihood of false or missed warnings.

Method used

By simultaneously acquiring intracranial pressure waveforms, electroencephalogram (EEG) signals, and transcranial Doppler blood flow velocity signals, preprocessed multimodal physiological datasets are extracted, data are integrated, time-domain, frequency-domain, and nonlinear features are extracted, dynamic weights are assigned, intracranial pressure is predicted, and early warning signals are generated.

Benefits of technology

It enables accurate prediction and timely early warning of changes in intracranial pressure, improving the accuracy and clinical applicability of the warning, ensuring that the warning signals meet actual clinical needs, and providing effective treatment guidance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121003424A_ABST
    Figure CN121003424A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of intelligent early warning, and discloses a multi-mode intracranial pressure intelligent early warning method and system.The method comprises the steps that an intracranial pressure waveform signal, an electroencephalogram signal and a transcranial Doppler blood flow velocity signal are synchronously collected, and a multi-mode physiological data set is obtained; extracting time domain features, frequency domain features and nonlinear features of the preprocessed multi-modal physiological data set to obtain a standardized feature parameter set; distributing dynamic weights for features in the standardized feature parameter set according to feature modal types of the standardized feature parameter set to obtain a fusion feature vector; performing intracranial pressure prediction on the patient according to the fusion feature vector after nonlinear transformation to obtain an intracranial pressure change sequence; according to the deviation condition of the intracranial pressure change sequence and a clinical preset safety threshold value, an intracranial pressure early warning signal is output; the accuracy of intracranial pressure early warning can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent early warning technology, and in particular to a multimodal intelligent early warning method and system for intracranial pressure. Background Technology

[0002] In clinical practice of intracranial pressure monitoring and early warning, traditional techniques often rely on single-modality physiological signals for analysis. Because changes in intracranial pressure are influenced by the synergistic effects of multiple physiological systems, including cerebrovascular hemodynamics and brain electrical activity, single-modality data cannot comprehensively capture the interrelationships between these systems, resulting in limited dimensions for perceiving abnormal trends in intracranial pressure. This limitation leads to incomplete early warning systems, making it difficult to identify potential risks of increased intracranial pressure in advance. Furthermore, it may result in false or missed warnings due to neglecting key physiological signals, failing to meet the core clinical requirement for accurate early warning and providing weak support for subsequent treatment decisions.

[0003] While some existing multimodal correlation technologies attempt to integrate multiple physiological signals, they have significant shortcomings in feature processing, fusion, and prediction. On the one hand, feature weight allocation often adopts a fixed pattern, failing to dynamically adjust according to the actual dispersion of different feature modalities and their correlation with intracranial pressure changes. This weakens the contribution of important features, while irrelevant or low-correlation features interfere with early warning decisions. On the other hand, in the intracranial pressure prediction process, there is often a lack of in-depth mining of temporal dependencies and verification of physiological patterns, making the prediction results susceptible to random signal fluctuations, resulting in numerous abnormal prediction values ​​and a lag in early warning signal output. This makes it difficult to effectively support timely clinical intervention, ultimately affecting the timing of patient treatment and prognosis. Summary of the Invention

[0004] This invention provides a multimodal intelligent early warning method and system for intracranial pressure to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a multimodal intelligent early warning method for intracranial pressure, comprising: S1. Simultaneously acquire the patient's intracranial pressure waveform signal, electroencephalogram signal, and transcranial Doppler blood flow velocity signal to obtain the patient's multimodal physiological dataset; S2. Extract the time-domain features, frequency-domain features, and nonlinear features of the preprocessed multimodal physiological dataset to obtain the standardized feature parameter set of the patient; S3. Assign dynamic weights to the features in the standardized feature parameter set according to the feature modality types of the standardized feature parameter set to obtain the fusion feature vector of the patient; S4. Based on the fused feature vector after nonlinear transformation, intracranial pressure is predicted for the patient to obtain the intracranial pressure change sequence of the patient. S5. Based on the deviation between the intracranial pressure change sequence and the clinically preset safety threshold, output the intracranial pressure warning signal for the patient.

[0006] In a preferred embodiment, the simultaneous acquisition of the patient's intracranial pressure waveform signal, electroencephalogram signal, and transcranial Doppler blood flow velocity signal yields a multimodal physiological dataset of the patient, including: The patient's intracranial pressure waveform signal, electroencephalogram signal, and blood flow velocity signal were collected. The time-aligned intracranial pressure waveform signal, the electroencephalogram signal, and the blood flow velocity signal are integrated to obtain the patient's multimodal physiological dataset; The multimodal physiological dataset of the patient is obtained by removing segments whose signal quality does not meet the preset requirements.

[0007] In a preferred embodiment, the extraction of time-domain features, frequency-domain features, and nonlinear features from the preprocessed multimodal physiological dataset yields a standardized set of feature parameters for the patient, including: The mean, variance, and peak factor of the preprocessed multimodal physiological dataset are aggregated into a set of intracranial pressure time-domain features of the patient. Fast Fourier Transform is performed on the EEG signals in the preprocessed multimodal physiological dataset to obtain the frequency domain feature set of the patient's EEG. The nonlinear dynamic features of the transcranial Doppler blood flow velocity signal in the preprocessed multimodal physiological dataset are extracted to obtain the blood flow nonlinear feature set of the patient. The intracranial pressure time-domain feature set, the electroencephalogram frequency-domain feature set, and the blood flow nonlinear feature set are merged to form the patient's initial feature parameter set; Eliminate the dimensional differences of features in the initial feature parameter set to obtain the standardized feature parameter set of the patient.

[0008] In a preferred embodiment, the step of assigning dynamic weights to features in the standardized feature parameter set based on the feature modality types of the standardized feature parameter set to obtain the patient's fusion feature vector includes: Based on the feature modality types of the standardized feature parameter set, the standardized feature parameter set is grouped to obtain a feature subset of the standardized feature parameter set; The dispersion of features in the standardized feature parameter set is evaluated based on the variance coefficient of the feature subset, and the feature importance evaluation result of the standardized feature parameter set is obtained. Based on the feature importance evaluation results, initial weight coefficients are assigned to the feature subset; Based on prior clinical knowledge, the initial weight coefficients that are more correlated with changes in intracranial pressure are increased to obtain the dynamic weight coefficient set of the standardized feature parameter set; The standardized feature parameter set is weighted and fused using the dynamic weight coefficient set to obtain the patient's fused feature vector.

[0009] In a preferred embodiment, the initial weighting coefficient is calculated using the following formula: ; In the formula, For the first The initial weight coefficients of each feature subset. For the first The average variance coefficient of each feature subset is the average variance coefficient of all feature subsets.

[0010] In a preferred embodiment, the step of predicting intracranial pressure in the patient based on the fused feature vector after nonlinear transformation to obtain the patient's intracranial pressure change sequence includes: Temporal dependency mining is performed on the fused feature vector after nonlinear transformation to obtain preliminary prediction results for the patient. By smoothing through a sliding window, the random fluctuations of the initial prediction results are reduced, resulting in a smoothed prediction sequence for the patient. Identify the trend pattern of intracranial pressure changes in the smoothed prediction sequence and the patient's historical intracranial pressure data to obtain the trend enhancement prediction result for the patient; By removing abnormal predicted values ​​that do not conform to physiological laws from the trend enhancement prediction results, the intracranial pressure change sequence of the patient is obtained.

[0011] In a preferred embodiment, the step of performing temporal dependency mining on the fused feature vector after nonlinear transformation to obtain the preliminary prediction result for the patient includes: By progressively capturing the dependencies of the fused feature vector after the nonlinear transformation at different time scales, the high-level temporal features of the patient are obtained. The features of key time steps in the high-level temporal features are enhanced to obtain the patient's attention-weighted temporal features; The attention-weighted temporal features are mapped to predicted intracranial pressure values ​​at future time points to obtain preliminary prediction results for the patient.

[0012] In a preferred embodiment, identifying the trend pattern of intracranial pressure changes in the smoothed prediction sequence and the patient's historical intracranial pressure data to obtain the patient's trend enhancement prediction result includes: The smoothed prediction sequence is time-aligned with the patient's historical intracranial pressure data to obtain time-aligned sequence data; The pattern matching result for the patient is obtained by matching the most similar change pattern in the historical data to the smoothed predicted sequence based on the time-aligned sequence data. Based on the pattern matching results, the subsequent change trend of the corresponding segment in the historical data is extracted as a reference pattern to obtain the trend reference template of the patient. The smoothed prediction sequence is weighted and fused with the trend reference template to obtain the trend enhancement prediction result for the patient; The clinical rationality of the trend enhancement prediction results was verified to obtain the trend enhancement prediction results for the patients.

[0013] In a preferred embodiment, the step of outputting an intracranial pressure warning signal for the patient based on the deviation between the intracranial pressure change sequence and a clinically preset safety threshold includes: Identify abnormal data points in the intracranial pressure change sequence that exceed a clinically preset safety threshold to obtain the patient's initial abnormal identification sequence; The duration and intensity of continuous abnormalities in the initial abnormality identification sequence are detected to obtain the abnormality persistence assessment results for the patient. Based on the results of the abnormal persistence assessment, a graded early warning indication signal is generated for the patient; By combining the patient's current physiological status data and historical warning records, the clinical rationality of the graded warning indication signal is verified, and the final warning decision for the patient is obtained.

[0014] To address the above problems, the present invention also provides a multimodal intelligent intracranial pressure early warning system, the system comprising: The physiological data acquisition module is used to simultaneously acquire the patient's intracranial pressure waveform signal, electroencephalogram signal, and transcranial Doppler blood flow velocity signal to obtain the patient's multimodal physiological dataset; The feature extraction module is used to extract the time-domain features, frequency-domain features, and nonlinear features of the preprocessed multimodal physiological dataset to obtain the standardized feature parameter set of the patient. The feature fusion module is used to assign dynamic weights to the features in the standardized feature parameter set according to the feature modality types of the standardized feature parameter set, so as to obtain the fusion feature vector of the patient; The intracranial pressure prediction module is used to predict the intracranial pressure of the patient based on the fused feature vector after nonlinear transformation, and to obtain the intracranial pressure change sequence of the patient. The early warning module is used to output an intracranial pressure early warning signal for the patient based on the deviation between the intracranial pressure change sequence and a clinically preset safety threshold.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a multimodal physiological dataset by simultaneously acquiring intracranial pressure waveform signals, electroencephalogram (EEG) signals, and transcranial Doppler blood flow velocity signals from patients. This dataset comprehensively captures multidimensional physiological information related to intracranial pressure, avoiding information loss due to the limitations of a single signal dimension. Furthermore, by extracting the time-domain, frequency-domain, and nonlinear features of the multimodal data and assigning dynamic weights based on the feature modality types, the correlation between different features and intracranial pressure changes is fully highlighted. This makes the fused feature vector more closely aligned with the physiological mechanisms of intracranial pressure changes, providing accurate and comprehensive feature support for subsequent intracranial pressure prediction and effectively improving the basic reliability of intracranial pressure prediction.

[0016] 2. In the intracranial pressure prediction stage, this invention effectively reduces the interference of random signal fluctuations on the prediction results by mining the temporal dependency relationship and identifying the trend pattern of the fused feature vector after nonlinear transformation, combined with sliding window smoothing. This accurately captures the dynamic trend of intracranial pressure changes and obtains an intracranial pressure change sequence that is more in line with actual physiological laws. In the early warning stage, based on the deviation between the intracranial pressure change sequence and the clinically preset safety threshold, combined with the abnormal persistence assessment and clinical rationality verification, a graded early warning signal is generated. This ensures that the early warning signal is both timely and meets actual clinical needs, significantly improving the accuracy and clinical applicability of intracranial pressure early warning and providing effective guidance for timely clinical intervention. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a multimodal intelligent early warning method for intracranial pressure according to an embodiment of the present invention. Figure 2 This is a functional block diagram of a multimodal intelligent intracranial pressure early warning system provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a multimodal intelligent intracranial pressure warning method. The executing entity of this multimodal intelligent intracranial pressure warning method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the multimodal intelligent intracranial pressure warning method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a multimodal intelligent early warning method for intracranial pressure according to an embodiment of the present invention. In this embodiment, the multimodal intelligent early warning method for intracranial pressure includes: S1. Simultaneously acquire the patient's intracranial pressure waveform signal, electroencephalogram signal, and transcranial Doppler blood flow velocity signal to obtain the patient's multimodal physiological dataset; In this embodiment of the invention, the simultaneous acquisition of the patient's intracranial pressure waveform signal, electroencephalogram signal, and transcranial Doppler blood flow velocity signal yields a multimodal physiological dataset of the patient, including: The patient's intracranial pressure waveform signal, electroencephalogram signal, and blood flow velocity signal were collected. The time-aligned intracranial pressure waveform signal, the electroencephalogram signal, and the blood flow velocity signal are integrated to obtain the patient's multimodal physiological dataset; The multimodal physiological dataset of the patient is obtained by removing segments whose signal quality does not meet the preset requirements.

[0021] Specifically, an invasive intracranial pressure monitor is used, in which a pressure sensor is inserted through a hole drilled in a suitable location in the patient's skull, allowing the sensor to directly contact the intracranial tissue and record changes in intracranial pressure over time in real time, thereby acquiring the patient's intracranial pressure waveform signal; an electroencephalogram (EEG) is used, in which conductive electrodes are attached to the patient's scalp according to standard lead positions, and the electrodes transmit the weak electrical activity generated by brain neurons to the EEG instrument, which amplifies and records the electrical signals, thereby acquiring the patient's electroencephalogram (EEG) signal; a transcranial Doppler ultrasound is used, in which an ultrasound probe is placed in the temporal window or other ultrasound-transmitting window of the patient's head, the probe emits ultrasound waves and receives ultrasound signals reflected back from intracranial blood vessels, and the instrument calculates and records changes in blood flow velocity over time based on the signals, thereby acquiring the patient's blood flow velocity signal.

[0022] Furthermore, the timestamp of any one of the intracranial pressure waveform signal, electroencephalogram (EEG) signal, and blood flow velocity signal is selected as the reference timestamp. The timestamps corresponding to each data point of the other two signals are compared one by one with the reference timestamp. By adjusting the arrangement order of the signal data, the data at the same time point in the three signals can correspond one-to-one. After time alignment, the time-aligned intracranial pressure waveform signal, EEG signal, and blood flow velocity signal are integrated into the same data file in chronological order to form a set containing complete data of the three signals, thus obtaining the patient's multimodal physiological dataset.

[0023] Furthermore, the signal quality preset requirements are first defined, specifically that the signal has no continuous sharp noise interference, the signal amplitude is within the normal physiological fluctuation range of the physiological signal, and the signal data is continuous without obvious data loss or breakage. Then, the multimodal physiological dataset is divided into multiple continuous signal segments according to a fixed time length. Each signal segment is checked one by one to determine whether the intracranial pressure waveform signal, electroencephalogram signal, and blood flow velocity signal in the segment meet the preset requirements. If any signal in a segment does not meet the preset requirements, the segment is removed from the multimodal physiological dataset. After checking and removing all segments, the patient's multimodal physiological dataset is obtained.

[0024] In summary, this acquisition method can comprehensively cover the core physiological dimensions related to intracranial pressure changes. The three types of signals correspond to direct dynamic indicators of intracranial pressure, brain electrical activity status, and cerebral hemodynamic characteristics, respectively. It can completely capture the information of the synergistic effects of multiple systems involved in intracranial pressure fluctuations, avoid the information loss caused by the limitation of a single signal dimension, and provide comprehensive data support for the subsequent accurate extraction of time domain, frequency domain, and nonlinear features.

[0025] In summary, by aligning data over time and removing low-quality segments during the data acquisition process, we can ensure the consistency of data timing and signal quality, reduce the interference of noise and abnormal data on subsequent analysis, and ensure that multimodal data can truly reflect the correlation of patients' physiological states.

[0026] In summary, this dataset lays the foundation for subsequent dynamic weight allocation and fusion feature vector construction. Only based on multimodal data can the weights be further adjusted according to the correlation between different feature modes and intracranial pressure changes, ultimately improving the accuracy of intracranial pressure prediction and early warning, and providing more reliable physiological data for clinical practice.

[0027] S2. Extract the time-domain features, frequency-domain features, and nonlinear features of the preprocessed multimodal physiological dataset to obtain the standardized feature parameter set of the patient; In this embodiment of the invention, the extraction of time-domain features, frequency-domain features, and nonlinear features from the preprocessed multimodal physiological dataset to obtain the standardized feature parameter set of the patient includes: The mean, variance, and peak factor of the preprocessed multimodal physiological dataset are aggregated into a set of intracranial pressure time-domain features of the patient. Fast Fourier Transform is performed on the EEG signals in the preprocessed multimodal physiological dataset to obtain the frequency domain feature set of the patient's EEG. The nonlinear dynamic features of the transcranial Doppler blood flow velocity signal in the preprocessed multimodal physiological dataset are extracted to obtain the blood flow nonlinear feature set of the patient. The intracranial pressure time-domain feature set, the electroencephalogram frequency-domain feature set, and the blood flow nonlinear feature set are merged to form the patient's initial feature parameter set; Eliminate the dimensional differences of features in the initial feature parameter set to obtain the standardized feature parameter set of the patient.

[0028] Specifically, first, all intracranial pressure-related data points in the preprocessed multimodal physiological dataset are acquired. The values ​​of all these data points are summed, and the sum is divided by the total number of data points to obtain the mean. Then, the difference between each intracranial pressure data point and the mean is calculated, and each difference is squared. The sum of all squared results is then divided by the total number of data points to obtain the variance. Next, the maximum value in the intracranial pressure data is found, and the square root (i.e., root mean square) of the sum of the squares of all intracranial pressure data points is calculated. The maximum value is divided by the root mean square to obtain the peak factor. The mean, variance, and peak factor are then organized into the same set to obtain the patient's intracranial pressure time-domain feature set.

[0029] Furthermore, the EEG signals in the preprocessed multimodal physiological dataset are divided into continuous signal segments of equal length according to fixed time intervals. For each EEG signal in the time domain (data on voltage changes over time), it is decomposed into sine and cosine wave components of different frequencies. The frequency value and power value at that frequency are determined for each component. The frequencies and corresponding power values ​​obtained from all signal segments are summarized to form a set containing frequency and power information, thus obtaining the patient's EEG frequency domain feature set.

[0030] Furthermore, continuous data points of transcranial Doppler blood flow velocity signals in the preprocessed multimodal physiological dataset were selected. When calculating the Lyapunov index, two points with a small initial distance were first found, and the distance between the two points was tracked over time. The rate of change of distance at each time step was calculated and the average was accumulated. When calculating the correlation dimension, the data points were constructed into a multidimensional space. A distance threshold was set to count the number of data point pairs less than the threshold. The threshold was changed and the count was repeated. The results were calculated based on the results. These nonlinear dynamic features were organized to obtain the set of nonlinear blood flow features of the patient.

[0031] Furthermore, all characteristic parameters, such as the mean, variance, and peak factor in the intracranial pressure time-domain characteristic set, the power values ​​corresponding to each frequency in the EEG frequency-domain characteristic set, and the Lyapunov index and correlation dimension in the blood flow nonlinearity characteristic set, are categorized and organized into the same data table according to characteristic type, ensuring that each parameter has a clear name and value. The resulting set is the patient's initial characteristic parameter set.

[0032] Furthermore, for each feature parameter in the initial feature parameter set, first find the maximum and minimum values ​​of the parameter and calculate the difference between them (range). Then, subtract the minimum value from each value of the parameter and divide the result by the range to obtain a standardized value. After completing this process for all feature parameters, organize the standardized values ​​into a new set to obtain the patient's standardized feature parameter set.

[0033] In summary, the three types of features accurately correspond to the core physiological information of different modal signals: time-domain features (mean, variance, peak factor) can capture the dynamic changes of intracranial pressure waveforms; frequency-domain features (fast Fourier transform results of EEG) can reflect the frequency distribution characteristics of brain electrical activity; and nonlinear features (dynamic characteristics of transcranial Doppler blood flow velocity signals) can reflect the complex fluctuation characteristics of cerebral blood flow. The combination of the three can comprehensively cover the multi-dimensional physiological indicators associated with intracranial pressure changes, avoid information omissions caused by the limitations of a single feature type, and provide rich feature support for subsequent analysis.

[0034] In summary, standardization by eliminating the dimensional differences in the initial feature parameter set can avoid interference caused by different features having different units and numerical ranges, ensure fair comparability of each feature in subsequent dynamic weight allocation, prevent important features from being underestimated or irrelevant features from being overemphasized due to dimensional issues, and guarantee the accuracy of feature evaluation and fusion.

[0035] In summary, this standardized set of feature parameters directly lays a high-quality foundation for subsequent dynamic weight allocation and feature vector fusion construction, ensuring that the subsequent intracranial pressure prediction process can be carried out based on accurate and unified feature data, thereby improving the reliability and accuracy of the overall early warning method.

[0036] S3. Assign dynamic weights to the features in the standardized feature parameter set according to the feature modality types of the standardized feature parameter set to obtain the fusion feature vector of the patient; In this embodiment of the invention, the step of assigning dynamic weights to features in the standardized feature parameter set based on the feature modality types of the standardized feature parameter set to obtain the patient's fusion feature vector includes: Based on the feature modality types of the standardized feature parameter set, the standardized feature parameter set is grouped to obtain a feature subset of the standardized feature parameter set; The dispersion of features in the standardized feature parameter set is evaluated based on the variance coefficient of the feature subset, and the feature importance evaluation result of the standardized feature parameter set is obtained. Based on the feature importance evaluation results, initial weight coefficients are assigned to the feature subset; Based on prior clinical knowledge, the initial weight coefficients that are more correlated with changes in intracranial pressure are increased to obtain the dynamic weight coefficient set of the standardized feature parameter set; The standardized feature parameter set is weighted and fused using the dynamic weight coefficient set to obtain the patient's fused feature vector.

[0037] In this embodiment of the invention, the initial weighting coefficient is calculated using the following formula: ; In the formula, For the first The initial weight coefficients of each feature subset. For the first The average variance coefficient of each feature subset is the average variance coefficient of all feature subsets.

[0038] Specifically, the modal types of features in the standardized feature parameter set are first identified, namely, intracranial pressure time-domain feature mode, EEG frequency-domain feature mode, and blood flow nonlinearity feature mode. Then, all features belonging to the intracranial pressure time-domain feature mode are selected from the standardized feature parameter set and grouped into one group; all features belonging to the EEG frequency-domain feature mode are selected and grouped into another group; and all features belonging to the blood flow nonlinearity feature mode are selected and grouped into a third group. Each group is a feature subset of the standardized feature parameter set.

[0039] Furthermore, the mean of each feature within each feature subset is calculated first, and then the standard deviation of that feature is calculated. The variance coefficient of that feature is obtained by dividing the standard deviation by the mean. The variance coefficients of all features within each feature subset are calculated one by one. The degree of feature dispersion is judged based on the size of the variance coefficient. The larger the variance coefficient, the higher the degree of feature dispersion and the higher the corresponding feature importance. The variance coefficients of each feature and their corresponding importance levels are compiled into a document to obtain the feature importance evaluation results of the standardized feature parameter set.

[0040] Furthermore, the initial weight coefficients are first set to a range of 0 to 1. Referring to the feature importance assessment results, the feature subset with the highest feature importance is assigned the largest initial weight coefficient in the range; the feature subset with medium feature importance is assigned the middle initial weight coefficient in the range; and the feature subset with the lowest feature importance is assigned the smallest initial weight coefficient in the range, ensuring that each feature subset corresponds to a unique initial weight coefficient, thus completing the allocation of the initial weight coefficients.

[0041] Furthermore, prior knowledge summarized by doctors in clinical practice was collected, which clearly showed that the correlation between intracranial pressure time-domain features and intracranial pressure changes was higher than that between EEG frequency-domain features and blood flow nonlinear features. The initial weight coefficients of the feature subsets corresponding to the intracranial pressure time-domain feature mode were found, and fixed values ​​were added to the original initial weight coefficients. The initial weight coefficients of the feature subsets of the EEG frequency-domain feature mode and the blood flow nonlinear feature mode remained unchanged. The weight coefficients of all the adjusted feature subsets were summarized to obtain the dynamic weight coefficient set of the standardized feature parameter set.

[0042] Furthermore, all feature values ​​within each feature subset are extracted from the standardized feature parameter set. Each feature value is multiplied by the dynamic weight coefficient corresponding to the feature subset to which it belongs, resulting in a weighted value for each feature. The weighted values ​​of all features are then arranged sequentially according to their original order in the standardized feature parameter set, forming an ordered numerical sequence, thus obtaining the patient's fusion feature vector.

[0043] Specifically, That is, the first The initial weight coefficients of each feature subset are obtained through calculation using this formula, and the calculation requires... That is, the first The average variance coefficient of the nth feature subset is derived from the calculation of the nth feature subset first. The variance coefficient of each feature within a feature subset (obtained by dividing the standard deviation of that feature by the mean) is then summed. The sum is then divided by the total number of features in that feature subset. The result is... ; That is, the sum of the average variance coefficients of all feature subsets, which is derived from the calculation of each feature subset individually. The sum of all the mean variance coefficients of each feature subset is the total. .

[0044] Furthermore, the significance of this formula is that by using the first... The average variance coefficient of each feature subset The sum of the average variance coefficients of all feature subsets Performing a division operation yields the result of the first division operation. Initial weight coefficients for each feature subset This result can reflect the first The proportion of the average dispersion of a feature subset to the total average dispersion of all feature subsets is directly related to the magnitude of the initial weight coefficient of that feature subset.

[0045] Furthermore, when the first The average variance coefficient of each feature subset When the variance coefficient of all other feature subsets increases, and the average variance coefficient of all feature subsets remains unchanged, the sum of the average variance coefficients of all feature subsets is... Will follow It increases with the increase of, but The proportion in the total will increase, leading to the first Initial weight coefficients for each feature subset Increase; when the first When the average variance coefficient (CV_avg,i) of one feature subset decreases, and the average variance coefficients of all other feature subsets remain unchanged, the sum of the average variance coefficients of all feature subsets is... Will follow The decrease is due to the decrease. The proportion in the total will decrease, leading to the first Initial weight coefficients for each feature subset Decrease; when the average variance coefficient of other feature subsets increases, and the th The average variance coefficient of each feature subset When kept constant, the sum of the average variance coefficients of all feature subsets It will increase. The proportion in the total will decrease, leading to the first Initial weight coefficients for each feature subset Decrease; when the average variance coefficient of other feature subsets decreases, and the th The average variance coefficient of each feature subset When kept constant, the sum of the average variance coefficients of all feature subsets It will decrease. The proportion in the total will increase, leading to the first Initial weight coefficients for each feature subset Increase.

[0046] In summary, after grouping by feature modality, the variance coefficient is used to assess the feature dispersion to determine its importance, ensuring that the initial weight allocation is accurately matched with the actual contribution of the feature. This avoids the feature value mismatch caused by a "one-size-fits-all" approach with fixed weights, and ensures that features with high dispersion and close correlation with intracranial pressure receive reasonable initial weights, thus solidifying the objective basis of weight allocation.

[0047] In summary, by combining prior clinical knowledge with weights that are more correlated with changes in intracranial pressure, the weight allocation is based on data patterns and aligns with clinical understanding of factors influencing intracranial pressure. This effectively avoids the problem of disconnect from clinical practice that may arise from purely data-driven approaches, and further enhances the rationality of the weights.

[0048] In summary, the fusion feature vector formed by weighted fusion can efficiently integrate the advantages of multimodal features such as time domain, frequency domain, and nonlinearity, fully retain the core information related to intracranial pressure in each modality, avoid the limitations of single-modal features, provide accurate and comprehensive feature support for subsequent intracranial pressure prediction, and help improve the accuracy and reliability of the prediction process.

[0049] S4. Based on the fused feature vector after nonlinear transformation, intracranial pressure is predicted for the patient to obtain the intracranial pressure change sequence of the patient. In this embodiment of the invention, the step of predicting intracranial pressure of the patient based on the fused feature vector after nonlinear transformation to obtain the intracranial pressure change sequence of the patient includes: Temporal dependency mining is performed on the fused feature vector after nonlinear transformation to obtain preliminary prediction results for the patient. By smoothing through a sliding window, the random fluctuations of the initial prediction results are reduced, resulting in a smoothed prediction sequence for the patient. Identify the trend pattern of intracranial pressure changes in the smoothed prediction sequence and the patient's historical intracranial pressure data to obtain the trend enhancement prediction result for the patient; By removing abnormal predicted values ​​that do not conform to physiological laws from the trend enhancement prediction results, the intracranial pressure change sequence of the patient is obtained.

[0050] In this embodiment of the invention, the step of performing temporal dependency mining on the fused feature vector after nonlinear transformation to obtain the preliminary prediction result for the patient includes: By progressively capturing the dependencies of the fused feature vector after the nonlinear transformation at different time scales, the high-level temporal features of the patient are obtained. The features of key time steps in the high-level temporal features are enhanced to obtain the patient's attention-weighted temporal features; The attention-weighted temporal features are mapped to predicted intracranial pressure values ​​at future time points to obtain preliminary prediction results for the patient.

[0051] In this embodiment of the invention, identifying the trend pattern of intracranial pressure changes in the smoothed prediction sequence and the patient's historical intracranial pressure data to obtain the patient's trend enhancement prediction result includes: The smoothed prediction sequence is time-aligned with the patient's historical intracranial pressure data to obtain time-aligned sequence data; The pattern matching result for the patient is obtained by matching the most similar change pattern in the historical data to the smoothed predicted sequence based on the time-aligned sequence data. Based on the pattern matching results, the subsequent change trend of the corresponding segment in the historical data is extracted as a reference pattern to obtain the trend reference template of the patient. The smoothed prediction sequence is weighted and fused with the trend reference template to obtain the trend enhancement prediction result for the patient; The clinical rationality of the trend enhancement prediction results was verified to obtain the trend enhancement prediction results for the patients.

[0052] Specifically, the fused feature vectors after nonlinear transformation are obtained, and they are arranged sequentially according to the time order corresponding to the feature vectors. The numerical change relationship between the feature vectors at two adjacent time points is analyzed, and the numerical correlation between feature vectors at intervals of one or two time points is also analyzed. Based on the numerical dependence relationship between the feature vectors at these time points, the intracranial pressure-related predicted values ​​corresponding to each subsequent time point are inferred. All the inferred predicted values ​​are organized in chronological order to obtain the preliminary prediction results for the patient.

[0053] Furthermore, a fixed-length sliding window is set, with the window length being three consecutive preliminary prediction result values. Starting from the first preliminary prediction result value, the three values ​​within the window are added together, and the sum is divided by 3 to obtain the first smoothed value. Then, the window is moved one position backward so that it covers the original second, third, and fourth preliminary prediction result values. The above calculation process of adding and dividing by 3 is repeated to obtain the second smoothed value. The window is continuously moved and calculated until the window covers all preliminary prediction result values. All the calculated smoothed values ​​are arranged in the order of calculation to obtain the patient's smoothed prediction sequence.

[0054] Furthermore, historical intracranial pressure data of patients were collected, and the trend patterns of rising, falling, and stable intracranial pressure values ​​were analyzed in chronological order. The smoothed prediction sequence was divided into multiple consecutive numerical segments according to time. The direction and magnitude of change of values ​​in each segment were compared with which trend pattern in the historical intracranial pressure data. The predicted values ​​in the corresponding segments were adjusted according to the consistent trend pattern to make the numerical changes more consistent with the historical intracranial pressure change pattern. All the adjusted predicted values ​​were sorted out to obtain the trend enhancement prediction results of patients.

[0055] Furthermore, the normal physiological range of intracranial pressure and the maximum reasonable range of intracranial pressure changes within a short period of time are determined. Each predicted value in the trend enhancement prediction results is checked one by one. If a predicted value exceeds the normal physiological range, or if the difference between the predicted value and the adjacent predicted value exceeds the maximum reasonable range of change, the predicted value is removed from the trend enhancement prediction results. The remaining predicted values ​​are arranged in the original time sequence to obtain the patient's intracranial pressure change sequence.

[0056] Specifically, the fusion feature vectors after nonlinear transformation are obtained, and three time scales (short, medium, and long) are determined. For the short time scale, fusion feature vectors from three consecutive time points are selected, and the correlation between the numerical changes of the last vector and the first two vectors is analyzed. For the medium time scale, five fusion feature vectors at one time point interval are selected, and the numerical dependence of the interval position vector and the vectors before and after it is analyzed. For the long time scale, seven fusion feature vectors at two time points interval are selected, and the numerical relationship between long-distance vectors is analyzed. The features corresponding to the dependencies captured under the three time scales are compiled and summarized to obtain the patient's high-level temporal features.

[0057] Furthermore, by comparing the feature values ​​of each time step in the high-level temporal features with the feature values ​​of two adjacent time steps, the magnitude of the change in values ​​is calculated. When the magnitude of the change exceeds 10% of the average value of the surrounding values, the time step is identified as a critical time step. The feature values ​​of the critical time steps are multiplied by a fixed coefficient of 1.2 for enhancement, while the feature values ​​of non-critical time steps remain unchanged. All features, both enhanced and unenhanced, are arranged in the original time step order to obtain the patient's attention-weighted temporal features.

[0058] Furthermore, three consecutive future time points need to be predicted. For the first future time point, the attention-weighted temporal feature values ​​of the most recent five time steps are selected, these values ​​are added together and divided by 5 to obtain the average value, which is used as the intracranial pressure prediction value for the first future time point. In the same way, the attention-weighted temporal feature values ​​of the corresponding most recent five time steps are selected and averaged to obtain the intracranial pressure prediction values ​​for the second and third future time points. The prediction values ​​of the three future time points are arranged in chronological order to obtain the preliminary prediction results for the patient.

[0059] Specifically, the timestamp corresponding to each data point in the smoothed prediction sequence is selected as the reference timestamp. The timestamp corresponding to each intracranial pressure value is extracted from the patient's historical intracranial pressure data. Each timestamp of the historical intracranial pressure data is compared with the reference timestamp one by one. The arrangement order of the historical intracranial pressure data is adjusted so that the timestamp of each value in the historical intracranial pressure data is completely consistent with the timestamp of the corresponding value in the smoothed prediction sequence. This forms a set of smoothed prediction sequences and historical intracranial pressure data that contain time matching, resulting in time-aligned sequence data.

[0060] Furthermore, the smoothed prediction sequence in the time-aligned sequence data is divided into continuous target segments with a fixed time length. At the same time, the patient's historical intracranial pressure data in the time-aligned sequence data is divided into multiple historical segments with the same time length. The magnitude of change (the sum of the differences between each subsequent value and the previous value) and the direction of change (positive difference means increase, negative difference means decrease) of the values ​​in each target segment are calculated. The magnitude and direction of change of each target segment are compared with all historical segments one by one. The historical segment with the smallest difference in magnitude and the completely consistent direction of change is found. This historical segment is the matching result. The matching historical segments corresponding to all target segments are sorted out to obtain the patient's pattern matching result.

[0061] Furthermore, the end time point of each matching historical segment in the pattern matching results is determined in the patient's historical intracranial pressure data. Starting from this end time point, subsequent time segments with the same length as the smoothed prediction sequence are selected. All intracranial pressure values ​​and their changing trends (duration and amplitude of increase, decrease, or stabilization) within these subsequent time segments are extracted. The values ​​and changing trends of the subsequent time segments corresponding to each matching historical segment are organized into structured data. Each structured data is a reference pattern. All reference patterns are summarized to obtain the patient's trend reference template.

[0062] Furthermore, the weight of the smoothed prediction sequence is set to 0.6, and the weight of the corresponding reference pattern in the trend reference template is set to 0.4. The value of the first time point is taken from the smoothed prediction sequence and multiplied by 0.6. At the same time, the value of the reference pattern corresponding to the first time point in the trend reference template is multiplied by 0.4. The two products are added together to obtain the first fusion value. The fusion values ​​of all corresponding time points in the smoothed prediction sequence and the trend reference template are calculated in the same way. All fusion values ​​are arranged in chronological order to obtain the trend enhancement prediction result of the patient.

[0063] Furthermore, the normal physiological range of intracranial pressure and the maximum reasonable range of intracranial pressure variation within 1 hour, determined in clinical diagnosis and treatment, were collected. Each predicted value in the trend enhancement prediction results was checked one by one to determine whether the value was within the normal physiological range. At the same time, the difference between the value and the previous predicted value was calculated, and it was determined whether the difference was within the maximum reasonable range of variation within 1 hour. If both judgments were satisfied, the value was retained; if either judgment was not satisfied, the value was removed. All values ​​that satisfied the judgments were retained and arranged in the original time order to obtain the trend enhancement prediction results for the patients.

[0064] In summary, by mining the temporal dependencies of the fused feature vector, we can gradually capture the feature associations at different time scales, and apply attention weighting to the features at key time steps to accurately extract high-level temporal features related to intracranial pressure changes. This makes the preliminary prediction results more consistent with the dynamic changes in intracranial pressure, and improves the pertinence and accuracy of the prediction.

[0065] In summary, sliding window smoothing can effectively reduce random fluctuations in preliminary prediction results, filter signal noise interference, make the obtained smooth prediction sequence more stable, better reflect the true trend of intracranial pressure changes, and avoid prediction bias caused by noise.

[0066] In summary, by matching the smoothed predicted sequence with the patient's historical intracranial pressure data and extracting subsequent trends from the historical data as a reference template and weighting and fusing them, combined with clinical rationality verification, the trend fit of the prediction results can be further enhanced, ensuring that it conforms to physiological laws. On the other hand, by removing abnormal predicted values ​​that do not conform to physiological laws, the authenticity and reliability of the intracranial pressure change sequence can be guaranteed, laying a high-quality data foundation for the subsequent output of early warning signals based on the deviation between the sequence and the safety threshold, and helping to improve the overall accuracy of early warning.

[0067] S5. Based on the deviation between the intracranial pressure change sequence and the clinically preset safety threshold, output the intracranial pressure warning signal for the patient.

[0068] In this embodiment of the invention, the step of outputting an intracranial pressure warning signal for the patient based on the deviation between the intracranial pressure change sequence and a clinically preset safety threshold includes: Identify abnormal data points in the intracranial pressure change sequence that exceed a clinically preset safety threshold to obtain the patient's initial abnormal identification sequence; The duration and intensity of continuous abnormalities in the initial abnormality identification sequence are detected to obtain the abnormality persistence assessment results for the patient. Based on the results of the abnormal persistence assessment, a graded early warning indication signal is generated for the patient; By combining the patient's current physiological status data and historical warning records, the clinical rationality of the graded warning indication signal is verified, and the final warning decision for the patient is obtained.

[0069] Specifically, a clinically preset safety threshold is defined. This threshold is a fixed safe range of intracranial pressure values ​​in clinical diagnosis and treatment. Data points are extracted one by one from the intracranial pressure change sequence. The value of each data point is compared with the clinically preset safety threshold. If the value of the data point exceeds the threshold, it is marked as 1; if it does not exceed the threshold, it is marked as 0. All the marked results are arranged in the original time order of the intracranial pressure change sequence to obtain the patient's initial abnormal identification sequence.

[0070] Furthermore, segments marked with consecutive 1s are selected from the initial abnormality identification sequence. Each data point corresponds to a fixed time unit. The duration of each consecutive abnormal segment is obtained by counting the number of consecutive 1s and multiplying by the time unit. At the same time, the corresponding data points of the consecutive abnormal segments in the intracranial pressure change sequence are mapped to the data points of the corresponding consecutive abnormal segments. The direction of change of these data point values ​​is observed to determine whether they are continuously increasing, continuously decreasing, or remaining stable. The duration of each consecutive abnormal segment and the corresponding abnormal intensity change trend are compiled into a document to obtain the patient's abnormality persistence assessment results.

[0071] Furthermore, a graded early warning standard is set: a level 1 warning is defined as a continuous abnormality lasting 1 to 3 time units with a stable abnormality intensity; a level 2 warning is defined as a continuous abnormality lasting 3 to 5 time units with a continuously increasing abnormality intensity; and a level 3 warning is defined as a continuous abnormality lasting more than 5 time units with a continuously increasing abnormality intensity. By comparing the duration and intensity trends of each continuous abnormality segment in the abnormality persistence assessment results, the corresponding warning level is determined, and a signal with a warning level identifier is generated. All signals are then aggregated to obtain the patient's graded early warning indication signal.

[0072] Furthermore, the patient's current physiological status data, including indicators such as heart rate and blood pressure, are collected. At the same time, the patient's historical warning records are retrieved to view records in the past that are similar to the current abnormality persistence assessment results and the corresponding clinical treatment conclusions. It is determined whether the current physiological status data is related to intracranial pressure abnormality. If both heart rate and blood pressure are within the normal range, the correlation is low. If the heart rate is increased and blood pressure is elevated, the correlation is high. Then, it is determined whether the level of the current graded warning indicator signal matches the final treatment conclusion of similar historical records. If the correlation is high and the level matches, the current graded warning indicator signal is retained. If the correlation is low or the level does not match, the warning level is adjusted. The final warning information determined is the patient's final warning decision.

[0073] In summary, this process uses "sequences" as the analysis object. By identifying abnormal data points in the sequence that exceed the safety threshold, and further detecting the time length and intensity change trend of continuous anomalies, it can avoid misjudgments caused by a single abnormal data point, accurately assess the persistence and severity of anomalies, reduce false alarms caused by accidental fluctuations or missed alarms caused by ignoring continuous anomalies, and improve the accuracy of early warning judgment.

[0074] In summary, the generation of graded early warning signals based on the results of abnormal persistence assessment can match different early warning levels according to the severity of the abnormality, enabling clinical medical staff to quickly identify intervention priorities and take different targeted measures. This avoids the problems of resource misallocation or untimely intervention under a unified early warning model, and improves the efficiency of clinical response.

[0075] In summary, combining the patient's current physiological status data with historical warning records to verify clinical rationality before issuing warning signals ensures that the warning results align with the patient's individual physiological characteristics and past disease patterns. This avoids situations where relying solely on general thresholds leads to discrepancies with individual realities, ensuring the clinical applicability and reliability of the warning signals. It provides scientific guidance for timely and precise clinical intervention, helping to safeguard the patient's treatment timing and prognosis.

[0076] like Figure 2 The diagram shown is a functional block diagram of a multimodal intracranial pressure intelligent early warning system provided in an embodiment of the present invention.

[0077] The multimodal intracranial pressure intelligent early warning system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the multimodal intracranial pressure intelligent early warning system 100 may include a physiological data acquisition module 101, a feature extraction module 102, a feature fusion module 103, an intracranial pressure prediction module 104, and an early warning module 105. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0078] In this embodiment, the functions of each module / unit are as follows: The physiological data acquisition module 101 is used to simultaneously acquire the patient's intracranial pressure waveform signal, electroencephalogram signal and transcranial Doppler blood flow velocity signal to obtain the patient's multimodal physiological dataset; The feature extraction module 102 is used to extract the time-domain features, frequency-domain features, and nonlinear features of the preprocessed multimodal physiological dataset to obtain the standardized feature parameter set of the patient. The feature fusion module 103 is used to assign dynamic weights to the features in the standardized feature parameter set according to the feature modality types of the standardized feature parameter set, so as to obtain the fusion feature vector of the patient. The intracranial pressure prediction module 104 is used to predict the intracranial pressure of the patient based on the fused feature vector after nonlinear transformation, and obtain the intracranial pressure change sequence of the patient. The early warning module 105 is used to output an intracranial pressure early warning signal for the patient based on the deviation between the intracranial pressure change sequence and the clinically preset safety threshold.

[0079] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0080] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0081] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0082] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0083] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multimodal intelligent early warning method for intracranial pressure, characterized in that, The method includes: S1. Simultaneously acquire the patient's intracranial pressure waveform signal, electroencephalogram signal, and transcranial Doppler blood flow velocity signal to obtain the patient's multimodal physiological dataset; S2. Extract the time-domain features, frequency-domain features, and nonlinear features of the preprocessed multimodal physiological dataset to obtain the standardized feature parameter set of the patient; S3. Assign dynamic weights to the features in the standardized feature parameter set according to the feature modality types of the standardized feature parameter set to obtain the fusion feature vector of the patient; S4. Based on the fused feature vector after nonlinear transformation, intracranial pressure is predicted for the patient to obtain the intracranial pressure change sequence of the patient. S5. Based on the deviation between the intracranial pressure change sequence and the clinically preset safety threshold, output the intracranial pressure warning signal for the patient.

2. The multimodal intelligent early warning method for intracranial pressure as described in claim 1, characterized in that, The simultaneous acquisition of intracranial pressure waveform signals, electroencephalogram (EEG) signals, and transcranial Doppler blood flow velocity signals from the patient yields a multimodal physiological dataset of the patient, including: The patient's intracranial pressure waveform signal, electroencephalogram signal, and blood flow velocity signal were collected. The time-aligned intracranial pressure waveform signal, the electroencephalogram signal, and the blood flow velocity signal are integrated to obtain the patient's multimodal physiological dataset; The multimodal physiological dataset of the patient is obtained by removing segments whose signal quality does not meet the preset requirements.

3. The multimodal intelligent early warning method for intracranial pressure as described in claim 1, characterized in that, The extraction of time-domain features, frequency-domain features, and nonlinear features from the preprocessed multimodal physiological dataset yields a standardized set of feature parameters for the patient, including: The mean, variance, and peak factor of the preprocessed multimodal physiological dataset are aggregated into a set of intracranial pressure time-domain features of the patient. Fast Fourier Transform is performed on the EEG signals in the preprocessed multimodal physiological dataset to obtain the frequency domain feature set of the patient's EEG. The nonlinear dynamic features of the transcranial Doppler blood flow velocity signal in the preprocessed multimodal physiological dataset are extracted to obtain the blood flow nonlinear feature set of the patient. The intracranial pressure time-domain feature set, the electroencephalogram frequency-domain feature set, and the blood flow nonlinear feature set are merged to form the patient's initial feature parameter set; Eliminate the dimensional differences of features in the initial feature parameter set to obtain the standardized feature parameter set of the patient.

4. The multimodal intelligent early warning method for intracranial pressure as described in claim 1, characterized in that, The step of assigning dynamic weights to features in the standardized feature parameter set based on the feature modality types of the standardized feature parameter set to obtain the patient's fusion feature vector includes: Based on the feature modality types of the standardized feature parameter set, the standardized feature parameter set is grouped to obtain a feature subset of the standardized feature parameter set; The dispersion of features in the standardized feature parameter set is evaluated based on the variance coefficient of the feature subset, and the feature importance evaluation result of the standardized feature parameter set is obtained. Based on the feature importance evaluation results, initial weight coefficients are assigned to the feature subset; Based on prior clinical knowledge, the initial weight coefficients that are more correlated with changes in intracranial pressure are increased to obtain the dynamic weight coefficient set of the standardized feature parameter set; The standardized feature parameter set is weighted and fused using the dynamic weight coefficient set to obtain the patient's fused feature vector.

5. The multimodal intelligent early warning method for intracranial pressure as described in claim 4, characterized in that, The formula for calculating the initial weighting coefficients is as follows: ; In the formula, For the first The initial weight coefficients of each feature subset. For the first The average variance coefficient of each feature subset is the average variance coefficient of all feature subsets.

6. The multimodal intelligent early warning method for intracranial pressure as described in claim 1, characterized in that, The step of predicting intracranial pressure in the patient based on the fused feature vector after nonlinear transformation, to obtain the patient's intracranial pressure change sequence, includes: Temporal dependency mining is performed on the fused feature vector after nonlinear transformation to obtain preliminary prediction results for the patient. By smoothing through a sliding window, the random fluctuations of the initial prediction results are reduced, resulting in a smoothed prediction sequence for the patient. Identify the trend pattern of intracranial pressure changes in the smoothed prediction sequence and the patient's historical intracranial pressure data to obtain the trend enhancement prediction result for the patient; By removing abnormal predicted values ​​that do not conform to physiological laws from the trend enhancement prediction results, the intracranial pressure change sequence of the patient is obtained.

7. The multimodal intelligent early warning method for intracranial pressure as described in claim 6, characterized in that, The step of mining the temporal dependency of the fused feature vector after nonlinear transformation to obtain the preliminary prediction results for the patient includes: By progressively capturing the dependencies of the fused feature vector after the nonlinear transformation at different time scales, the high-level temporal features of the patient are obtained. The features of key time steps in the high-level temporal features are enhanced to obtain the patient's attention-weighted temporal features; The attention-weighted temporal features are mapped to predicted intracranial pressure values ​​at future time points to obtain preliminary prediction results for the patient.

8. The multimodal intelligent early warning method for intracranial pressure as described in claim 7, characterized in that, The process of identifying the trend pattern of intracranial pressure changes in the smoothed prediction sequence and the patient's historical intracranial pressure data to obtain the patient's trend enhancement prediction result includes: The smoothed prediction sequence is time-aligned with the patient's historical intracranial pressure data to obtain time-aligned sequence data; The pattern matching result for the patient is obtained by matching the most similar change pattern in the historical data to the smoothed predicted sequence based on the time-aligned sequence data. Based on the pattern matching results, the subsequent change trend of the corresponding segment in the historical data is extracted as a reference pattern to obtain the trend reference template of the patient. The smoothed prediction sequence is weighted and fused with the trend reference template to obtain the trend enhancement prediction result for the patient; The clinical rationality of the trend enhancement prediction results was verified to obtain the trend enhancement prediction results for the patients.

9. The multimodal intelligent early warning method for intracranial pressure as described in claim 1, characterized in that, The step of outputting an intracranial pressure warning signal for the patient based on the deviation between the intracranial pressure change sequence and a clinically preset safety threshold includes: Identify abnormal data points in the intracranial pressure change sequence that exceed a clinically preset safety threshold to obtain the patient's initial abnormal identification sequence; The duration and intensity of continuous abnormalities in the initial abnormality identification sequence are detected to obtain the abnormality persistence assessment results for the patient. Based on the results of the abnormal persistence assessment, a graded early warning indication signal is generated for the patient; By combining the patient's current physiological status data and historical warning records, the clinical rationality of the graded warning indication signal is verified, and the final warning decision for the patient is obtained.

10. A multimodal intelligent early warning system for intracranial pressure, characterized in that, The system includes: The physiological data acquisition module is used to simultaneously acquire the patient's intracranial pressure waveform signal, electroencephalogram signal, and transcranial Doppler blood flow velocity signal to obtain the patient's multimodal physiological dataset; The feature extraction module is used to extract the time-domain features, frequency-domain features, and nonlinear features of the preprocessed multimodal physiological dataset to obtain the standardized feature parameter set of the patient. The feature fusion module is used to assign dynamic weights to the features in the standardized feature parameter set according to the feature modality types of the standardized feature parameter set, so as to obtain the fusion feature vector of the patient; The intracranial pressure prediction module is used to predict the intracranial pressure of the patient based on the fused feature vector after nonlinear transformation, and to obtain the intracranial pressure change sequence of the patient. The early warning module is used to output an intracranial pressure early warning signal for the patient based on the deviation between the intracranial pressure change sequence and a clinically preset safety threshold.

Citation Information

Cited By

  • Dynamic multi-parameter fusion monitoring method and system for intracranial pressure of patient

    CN121647634A

  • Intracranial pressure monitor with trend prediction function

    CN121943252A