Dynamic monitoring of intracranial pressure and grading intervention method for severe neurologic patients

Abstract

The application discloses a method for monitoring intracranial pressure of a severe nervous patient dynamically and grading intervention, and relates to the technical field of dynamic intracranial pressure monitoring.The application overcomes the defect of low monitoring precision of a single mode by fusing arterial blood pressure waveform, electroencephalogram signals and head CT image multi-modal data, extracting time domain, frequency domain and nonlinear dynamic characteristics.A three-dimensional feature tensor constructed in combination with a 3D-CNN and an LSTM network effectively captures the spatiotemporal nonlinear coupling relationship of cross-modal data, realizes continuous trend prediction of intracranial pressure in the future 5-30 minutes, and provides time for clinical early intervention.A dynamic early warning threshold calculation formula based on individual brain compliance is adopted, combined with hemodynamics and electroencephalogram features for false positive inhibition, which significantly reduces the false alarm rate caused by motion artifacts and equipment interference.The application improves the precision and robustness of monitoring, realizes the transformation from passive alarm to active prediction, and effectively reduces the risk of complications of a severe nervous patient.

Description

Technical Field

[0001] This invention relates to the field of intracranial pressure dynamic monitoring technology, specifically to a method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients. Background Technology

[0002] In neurocritical care medicine, monitoring and controlling intracranial pressure (ICP) is a core aspect of treating patients with severe traumatic brain injury (TBI), subarachnoid hemorrhage (SAH), intracerebral hemorrhage (ICH), and post-brain tumor surgery. Increased intracranial pressure is a leading cause of brain herniation, cerebral ischemia, and death. Currently, the "gold standard" for monitoring ICP in clinical practice is invasive monitoring techniques, such as external ventricular drainage (EVD) or intraparenchymal microprobe monitoring. However, invasive monitoring carries risks of infection and bleeding, and the equipment is expensive, requiring specialized neurosurgeons for operation, making it difficult to implement in primary care hospitals and unsuitable for all patients requiring long-term monitoring.

[0003] To overcome the limitations of invasive monitoring, existing non-invasive or minimally invasive intracranial pressure monitoring technologies have emerged, mainly including the following categories: Ultrasound-based techniques, such as transcranial Doppler (TCD) for detecting cerebral blood flow velocity or measuring the optic nerve sheath diameter (ONSD), are non-invasive, but their accuracy is greatly affected by the operator's experience and cannot achieve continuous real-time monitoring.

[0004] Techniques based on bioelectrical impedance or tympanic membrane displacement: These methods estimate ICP by detecting changes in skull impedance or sound wave conduction characteristics. However, they are susceptible to interference from scalp edema, middle ear lesions, etc., resulting in poor measurement stability.

[0005] Machine learning models based on single physiological signals: Some studies attempt to estimate ICP using only arterial blood pressure (ABP) waveforms or only electroencephalogram (EEG) signals.

[0006] However, existing ICP monitoring and early warning methods have the following significant shortcomings, failing to meet the urgent needs of neurocritical care clinics for precision, personalization, and early warning: First, single-modal data is not fully utilized, and there is a lack of complementary information from multiple sources.

[0007] Existing non-invasive monitoring methods often rely on a single signal source. For example, regression analysis based solely on ABP waveform characteristics (such as systolic and diastolic blood pressure) ignores the brain perfusion and function information contained in electroencephalogram (EEG); or they only analyze static CT / MRI images without incorporating real-time hemodynamic changes. In reality, changes in intracranial pressure are the result of the combined effects of cerebrovascular pressure, brain tissue compliance, and brain volume. Using data from a single modality alone cannot comprehensively reflect the complex pathophysiological state of the intracranial cavity, leading to a significant decrease in predictive accuracy in patients with large individual variability.

[0008] Second, the feature extraction is coarse and lacks in-depth mining of nonlinear dynamic features.

[0009] Current physiological signal analysis often employs simple time-domain statistics (such as mean and variance). For arterial blood pressure, only mean arterial pressure (MAP) is typically calculated; for electroencephalography (EEG), simple frequency-domain power analysis is commonly used. However, the physiological systems of critically ill neurological patients exhibit highly nonlinear characteristics. For instance, the low-frequency / high-frequency power ratio (LF / HF) in heart rate variability (HRV) reflects the state of autonomic nervous system regulation, while approximate entropy (ApEn), sample entropy (SampEn), and wavelet packet energy entropy in EEG can quantify the complexity and disorder of EEG data. Existing techniques lack a systematic extraction and fusion of these deep nonlinear dynamic characteristics, resulting in the loss of a large number of weak signals associated with precursors of increased intracranial pressure.

[0010] Third, it lacks deep integration of spatiotemporal dimensions and the ability to predict future trends.

[0011] Existing early warning systems mostly operate on a "threshold alarm" model, meaning they trigger an alarm when the measured or estimated ICP exceeds a fixed value (e.g., 20 mmHg). This model is lagging and does not consider baseline differences among patients. More importantly, existing methods often employ simple linear models or shallow neural networks, which cannot effectively handle the high-dimensional coupling relationships of multimodal data across time, space, and feature dimensions. They typically only output the current ICP estimate and cannot predict ICP trends over a future period (e.g., the next 5-30 minutes), preventing doctors from intervening in advance (e.g., dehydration, blood pressure escalation, or surgery).

[0012] Fourth, the warning threshold is fixed, the false positive rate is high, and there is a lack of dynamic calibration mechanism.

[0013] Clinically used ICP warning thresholds (e.g., 20 mmHg) are based on population statistics and do not consider individual brain compliance. For patients with poor brain compliance, a slight increase in ICP (e.g., 15 mmHg) may already indicate danger; while for patients with good brain compliance, a slightly higher ICP may be well tolerated. In addition, existing monitoring systems are prone to false alarms when faced with motion artifacts or equipment interference; they lack a false-positive suppression mechanism that combines hemodynamics (e.g., synchronous increase in mean arterial pressure) and EEG characteristics (e.g., slowing of wavelets) for cross-validation. Summary of the Invention

[0014] To address the aforementioned technical issues, a method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients is provided. This technical solution resolves the problems raised in the background section.

[0015] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect of the invention, a method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients is provided, comprising: Simultaneously collect multimodal monitoring data from critically ill neurological patients. The multimodal monitoring data includes at least continuous arterial blood pressure waveform data, electroencephalogram signal data, and head computed tomography (CT) image data. The collected multimodal monitoring data is time-stamp aligned and preprocessed to remove noise fragments caused by motion artifacts or equipment interference; Based on the preprocessed data, the time-domain and frequency-domain features of arterial blood pressure waveforms, the nonlinear dynamic features of electroencephalogram signals, and the brain tissue morphological features of head computed tomography images were extracted. The extracted features are mapped to a unified high-dimensional space to construct a three-dimensional feature tensor containing spatial, feature, and temporal dimensions. The three-dimensional feature tensor is input into a preset deep neural network model, and a spatiotemporal feature fusion operation is performed to capture the nonlinear coupling relationship between cross-modal data and output the continuous prediction value of intracranial pressure within a preset future time window. Residual analysis was performed between the continuously predicted intracranial pressure values ​​and the gold standard values ​​of invasive intracranial pressure monitoring collected in real time to calculate the dynamic early warning threshold. When the continuously predicted value of intracranial pressure exceeds the dynamic warning threshold and the duration is longer than the preset duration, a graded warning signal is generated.

[0016] Preferably, the simultaneous acquisition of multimodal monitoring data from critically ill neurological patients specifically includes: Collect continuous arterial blood pressure waveform data at a sampling frequency of 100Hz to 500Hz for a duration of no less than 30 minutes; EEG signal data were acquired at a sampling frequency of 256Hz to 1024Hz, with no fewer than 8 leads. Head computed tomography (CT) images were collected upon patient admission and during changes in the patient's condition, and the raw DICOM format data of the images were extracted. A linear interpolation algorithm was used to resample physiological signals with different sampling rates to a uniform frequency, and the R-wave peak of the EEG signal was used as the time reference point to align the arterial blood pressure waveform with the time axis of the EEG signal.

[0017] Preferably, the extraction of brain tissue morphological features from head computed tomography images specifically includes: U-Net convolutional neural network was used to perform cranial dissection and brain tissue segmentation on head computed tomography images to obtain mask images of brain parenchyma, ventricles and subarachnoid space. Calculate the ratio of brain parenchyma volume to total intracranial volume. and the maximum transverse diameter of the bilateral lateral ventricle bodies. ; Based on Hounsfield units, the mean CT values ​​of brain parenchyma regions were statistically analyzed. ,variance and skewness The specific statistical formula is as follows: ; in, This represents the total number of voxels in the brain parenchyma region. For the first CT value of individual pixels; The mask image and volume ratio , transverse diameter value The density feature vectors are combined to form the image morphological feature set.

[0018] Preferably, the extraction of time-domain and frequency-domain features of arterial blood pressure waveforms and nonlinear dynamic features of electroencephalogram signals specifically includes: For arterial blood pressure waveforms, calculate the values ​​of systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse pressure, heart rate variability, and low-frequency / high-frequency power ratio; For EEG signals, calculate the 90% spectral edge frequency, burst suppression ratio, approximate entropy, and sample entropy; Wavelet packet decomposition was performed on the EEG signal to extract the wavelet packet energy entropy of nodes from the 3rd to the 5th layer. The extracted physiological signal features and image morphological features are subjected to Z-score normalization, and the specific calculation formula is as follows: ; in, These are the standardized eigenvalues. These are the original eigenvalues. The mean of this feature for the training set, The standard deviation of this feature in the training set.

[0019] Preferably, the construction of the three-dimensional feature tensor, which includes spatial, feature, and temporal dimensions, specifically includes: The standardized arterial blood pressure feature vector With EEG feature vectors Concatenate along the feature dimension to form a one-dimensional feature vector sequence. ; Morphological feature vectors extracted from head computed tomography images The replication expands into a sequence matrix with a duration equal to that of the physiological signal. ; Based on the time window length The sliding step size is The above sequence data is sliced ​​into multiple data segments; The physiological features, EEG features, and imaging features within each data segment are stacked spatially to generate a shape of... 3D feature tensor The specific construction logic satisfies: ; in, The number of feature channels, This is a spatial feature dimension, typically 1 or the number of brain regions. For time step, and = , These are channel, spatial, and temporal indexes, respectively.

[0020] Preferably, the spatiotemporal feature fusion operation specifically includes: The three-dimensional feature tensor is input into the three-dimensional convolutional neural network layer, and local spatiotemporal features are extracted using a 3×3×3 convolutional kernel. The convolutional layer is followed by a batch normalization layer and a ReLU activation function. The output of the 3D convolution is flattened and then fed into a long short-term memory network layer to capture long-term temporal dependencies. A channel attention mechanism module is inserted between the 3D convolutional network and the long short-term memory network, and the weight coefficients of each feature channel are calculated through global average pooling and fully connected layers. The specific calculation formula is as follows: ; in, For the input feature map, This is a global average pooling operation. and This is the weight matrix of the fully connected layer. It is the ReLU activation function. Use the Sigmoid activation function; The input and output of the 3D convolution are added together using a residual connection structure to prevent gradient vanishing.

[0021] Preferably, the output of continuously predicted intracranial pressure values ​​within a future preset time window specifically includes: Set the prediction time window to the future. minutes, time step is minute; The output layer of the neural network model uses a fully connected layer and a linear activation function, and the output dimension is a vector of intracranial pressure values ​​equal to the length of the prediction time window. ; During the model training phase, the loss function Using the mean square error loss function With smoothing L1 loss function The weighted combination is calculated using the following formula: ; ; ; in, and Weighting coefficients =0.7, , This represents the gold standard value for true intracranial pressure. These are the model's predicted values. This represents the number of samples.

[0022] Preferably, the calculation of the dynamic early warning threshold specifically includes: Obtain patient past The average value of invasive intracranial pressure monitoring data within one hour was used as the baseline. ; Calculate baseline Standard deviation ; Set individualized early warning coefficient ,in The value range is 1.5≤ ≤3.0, and Values ​​vary with the patient's brain compliance index The decrease in linearly increases the value, and the specific relationship is as follows: ; Among them, among them, =1.5, =3.0, This is the reference value for normal brain compliance; The formula for calculating the dynamic warning threshold (Threshold) is as follows: ; When the predicted value is greater than and less than A Level 1 warning is triggered when the predicted value is greater than or equal to... A level-two warning is triggered at this time, among which =5mmHg.

[0023] Preferably, before generating a graded warning signal when the continuously predicted value of intracranial pressure exceeds the dynamic warning threshold and the duration is greater than a preset duration, a false positive suppression step is also included: Morphological filtering is performed on the continuous intracranial pressure prediction values ​​output by the neural network to remove spurious noise with a width smaller than the preset pixel width. Calculate the slope of the predicted value curve near the warning threshold. The specific calculation formula is as follows: ; in, The predicted value at the current moment. The sampling interval; If the slope If the value exceeds the preset threshold for a sharp increase, it is determined to be an acute event; By combining the instantaneous changes in arterial blood pressure waveform, if the increase in intracranial pressure is accompanied by a synchronous increase in mean arterial pressure (MAP) and the appearance of slow waves on the electroencephalogram (EEG), the warning signal is confirmed to be effective; if only the intracranial pressure increases while the hemodynamics and EEG characteristics remain stable, the issuance of the warning signal is suppressed.

[0024] In a second aspect of the invention, a system for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients is also provided, comprising: The acquisition module is used to simultaneously acquire multimodal monitoring data of critically ill neurological patients. The multimodal monitoring data includes at least continuous arterial blood pressure waveform data, electroencephalogram signal data, and head computed tomography (CT) image data. The data processing module is used to perform timestamp alignment and preprocessing on the collected multimodal monitoring data to remove noise fragments caused by motion artifacts or equipment interference. The extraction module is used to extract the time-domain and frequency-domain features of arterial blood pressure waveforms, the nonlinear dynamic features of electroencephalogram signals, and the brain tissue morphological features of head computed tomography images based on the preprocessed data. The construction module is used to map the extracted features to a unified high-dimensional space and construct a three-dimensional feature tensor containing spatial dimension, feature dimension and time dimension. The output module is used to input three-dimensional feature tensors into a preset deep neural network model, perform spatiotemporal feature fusion operations, capture the nonlinear coupling relationship between cross-modal data, and output continuous predicted values ​​of intracranial pressure within a preset future time window. The calculation module is used to perform residual analysis based on the continuous predicted value of intracranial pressure and the gold standard value of invasive intracranial pressure monitoring collected in real time, and to calculate the dynamic early warning threshold. The early warning module is used to generate a graded early warning signal when the continuously predicted value of intracranial pressure exceeds the dynamic early warning threshold and the duration is greater than a preset duration.

[0025] Compared with existing technologies, this invention provides a method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients, which has the following beneficial effects: This invention overcomes the low accuracy of single-modal monitoring by fusing multimodal data from arterial blood pressure waveforms, electroencephalogram (EEG) signals, and head CT images to extract time-domain, frequency-domain, and nonlinear dynamic features. The constructed three-dimensional feature tensor, combined with 3D-CNN and LSTM networks, effectively captures the spatiotemporal nonlinear coupling relationships across modal data, enabling continuous prediction of intracranial pressure trends over the next 5-30 minutes, thus providing valuable time for early clinical intervention. A dynamic warning threshold calculation formula based on individualized brain compliance is employed, along with hemodynamic and EEG features for false positive suppression, significantly reducing the false alarm rate caused by motion artifacts and equipment interference. Furthermore, a sliding window online update mechanism is introduced, using nearly 24 hours of gold-standard data to fine-tune model parameters, allowing the system to adapt to the dynamic evolution of the patient's condition. This invention improves the accuracy and robustness of monitoring, achieving a shift from passive alarm to proactive prediction, effectively reducing the risk of complications in critically ill neurological patients. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the method flow of S101-S107 in this invention; Figure 2 This is a schematic diagram of the method flow for S201-S204 in this invention; Figure 3 This is a schematic diagram of the method flow for S301-S304 in this invention; Figure 4 This is a schematic diagram of the method flow for S401-S404 in this invention; Figure 5 This is a schematic diagram of the method flow for S501-S504 in this invention. Detailed Implementation

[0027] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0028] Example 1 Please refer to Figure 1 As shown, in a first aspect of the present invention, a method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients is provided, comprising: S101. Simultaneously collect multimodal monitoring data of critically ill neurological patients. The multimodal monitoring data shall include at least continuous arterial blood pressure waveform data, electroencephalogram signal data and head computed tomography image data. S102. Perform timestamp alignment and preprocessing on the collected multimodal monitoring data to remove noise fragments caused by motion artifacts or equipment interference. S103. Based on the preprocessed data, extract the time-domain and frequency-domain features of arterial blood pressure waveform, the nonlinear dynamic features of electroencephalogram signals, and the brain tissue morphological features of head computed tomography images, respectively. S104. Map the extracted features to a unified high-dimensional space to construct a three-dimensional feature tensor containing spatial dimension, feature dimension and time dimension. S105. Input the three-dimensional feature tensor into the preset deep neural network model, perform spatiotemporal feature fusion operation, capture the nonlinear coupling relationship between cross-modal data, and output the continuous predicted value of intracranial pressure within the preset time window in the future. S106. Perform residual analysis based on the continuous predicted value of intracranial pressure and the gold standard value of invasive intracranial pressure monitoring collected in real time, and calculate the dynamic early warning threshold. S107. When the continuous predicted value of intracranial pressure exceeds the dynamic warning threshold and the duration is longer than the preset duration, a graded warning signal is generated.

[0029] Those skilled in the art will understand that this invention overcomes the low accuracy of single-modality monitoring by fusing multimodal data from arterial blood pressure waveforms, electroencephalogram (EEG) signals, and head CT images to extract time-domain, frequency-domain, and nonlinear dynamic features. The constructed three-dimensional feature tensor, combined with 3D-CNN and LSTM networks, effectively captures the spatiotemporal nonlinear coupling relationship of cross-modal data, enabling continuous prediction of intracranial pressure trends over the next 5-30 minutes, thus providing time for early clinical intervention. A dynamic early warning threshold calculation formula based on individualized brain compliance is employed, combined with hemodynamic and EEG features for false positive suppression, significantly reducing the false alarm rate caused by motion artifacts and equipment interference. Furthermore, a sliding window online update mechanism is introduced, using nearly 24 hours of gold standard data to fine-tune model parameters, allowing the system to adapt to the dynamic evolution of the patient's condition. This invention improves the accuracy and robustness of monitoring, realizing a shift from passive alarm to active prediction, effectively reducing the risk of complications in critically ill neurological patients.

[0030] Please refer to Figure 2 As shown, multimodal monitoring data were collected simultaneously from critically ill neurological patients, specifically including: S201. Collect continuous arterial blood pressure waveform data at a sampling frequency of 100Hz to 500Hz for a duration of not less than 30 minutes. S202. Collect EEG signal data at a sampling frequency of 256Hz to 1024Hz, with no less than 8 leads collected. S203. Collect head computed tomography (CT) images of the patient upon admission and when the condition changes, and extract the raw DICOM format data of the images. S204. Using a linear interpolation algorithm, physiological signals with different sampling rates are resampled to a uniform frequency, and the R-wave peak of the EEG signal is used as the time reference point to align the arterial blood pressure waveform with the time axis of the EEG signal.

[0031] Please refer to Figure 3 As shown, the morphological features of brain tissue extracted from head computed tomography (CT) images specifically include: S301. Use U-Net convolutional neural network to perform cranial dissection and brain tissue segmentation on head computed tomography images to obtain mask images of brain parenchyma, ventricles and subarachnoid space. S302. Calculate the ratio of brain parenchyma volume to total intracranial volume. and the maximum transverse diameter of the bilateral lateral ventricle bodies. ; S303. Based on Hounsfield units, statistically analyze the mean CT values ​​of brain parenchyma regions. ,variance and skewness The specific statistical formula is as follows: ; in, This represents the total number of voxels in the brain parenchyma region. For the first CT value of individual pixels; S304, Mix the mask image and volume ratio. , transverse diameter value The density feature vectors are combined to form the image morphological feature set.

[0032] Please refer to Figure 4 As shown, the time-domain and frequency-domain features of arterial blood pressure waveforms and the nonlinear dynamic features of electroencephalogram (EEG) signals are extracted, specifically including: S401. For arterial blood pressure waveform, calculate the values ​​of systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse pressure, heart rate variability and low-frequency / high-frequency power ratio. S402. For EEG signals, calculate the 90% spectral edge frequency, burst suppression ratio, approximate entropy, and sample entropy. S403. Perform wavelet packet decomposition on the EEG signal and extract the wavelet packet energy entropy of the nodes from the 3rd to the 5th layer. S404. The extracted physiological signal features and image morphological features are subjected to Z-score normalization. The specific calculation formula is as follows: ; in, These are the standardized eigenvalues. These are the original eigenvalues. The mean of this feature for the training set, The standard deviation of this feature in the training set.

[0033] Please refer to Figure 5 As shown, constructing a three-dimensional feature tensor that includes spatial, feature, and temporal dimensions specifically involves: S501, Standardize the arterial blood pressure feature vector With EEG feature vectors Concatenate along the feature dimension to form a one-dimensional feature vector sequence. ; S502, Morphological feature vectors extracted from head computed tomography images The replication expands into a sequence matrix with a duration equal to that of the physiological signal. ; S503, with time window length as The sliding step size is The above sequence data is sliced ​​into multiple data segments; S504. Stack the physiological features, EEG features, and imaging features within each data segment in a spatial dimension to generate a shape as follows: 3D feature tensor The specific construction logic satisfies: ; in, The number of feature channels, This is a spatial feature dimension, typically 1 or the number of brain regions. For time step, and = , These are channel, spatial, and temporal indexes, respectively.

[0034] The spatiotemporal feature fusion operation specifically includes: The three-dimensional feature tensor is input into the three-dimensional convolutional neural network layer, and local spatiotemporal features are extracted using a 3×3×3 convolutional kernel. The convolutional layer is followed by a batch normalization layer and a ReLU activation function. The output of the 3D convolution is flattened and then fed into a long short-term memory network layer to capture long-term temporal dependencies. A channel attention mechanism module is inserted between the 3D convolutional network and the long short-term memory network, and the weight coefficients of each feature channel are calculated through global average pooling and fully connected layers. The specific calculation formula is as follows: ; in, For the input feature map, This is a global average pooling operation. and This is the weight matrix of the fully connected layer. It is the ReLU activation function. Use the Sigmoid activation function; The input and output of the 3D convolution are added together using a residual connection structure to prevent gradient vanishing.

[0035] Output continuous predicted values ​​of intracranial pressure within a future preset time window, specifically including: Set the prediction time window to the future. minutes, time step is minute; The output layer of the neural network model uses a fully connected layer and a linear activation function, and the output dimension is a vector of intracranial pressure values ​​equal to the length of the prediction time window. ; During the model training phase, the loss function Using the mean square error loss function With smoothing L1 loss function The weighted combination is calculated using the following formula: ; ; ; in, and Weighting coefficients =0.7, , This represents the gold standard value for true intracranial pressure. These are the model's predicted values. This represents the number of samples.

[0036] The calculation of dynamic early warning thresholds specifically includes: Obtain patient past The average value of invasive intracranial pressure monitoring data within one hour was used as the baseline. ; Calculate baseline Standard deviation ; Set individualized early warning coefficient ,in The value range is 1.5≤ ≤3.0, and Values ​​vary with the patient's brain compliance index The decrease in linearly increases, and the specific relationship is as follows: ; Among them, among them, =1.5, =3.0, This is the reference value for normal brain compliance; The formula for calculating the dynamic warning threshold (Threshold) is as follows: ; When the predicted value is greater than and less than A Level 1 warning is triggered when the predicted value is greater than or equal to... A level-two warning is triggered at this time, among which =5mmHg.

[0037] Before generating a graded warning signal, a false positive suppression step is included when the continuously predicted value of intracranial pressure exceeds the dynamic warning threshold and the duration is longer than the preset duration. Morphological filtering is performed on the continuous intracranial pressure prediction values ​​output by the neural network to remove spurious noise with a width smaller than the preset pixel width. Calculate the slope of the predicted value curve near the warning threshold. The specific calculation formula is as follows: ; in, The predicted value at the current moment. The sampling interval; If the slope If the value exceeds the preset threshold for a sharp increase, it is determined to be an acute event; By combining the instantaneous changes in arterial blood pressure waveform, if the increase in intracranial pressure is accompanied by a synchronous increase in mean arterial pressure (MAP) and the appearance of slow waves on the electroencephalogram (EEG), the warning signal is confirmed to be effective; if only the intracranial pressure increases while the hemodynamics and EEG characteristics remain stable, the issuance of the warning signal is suppressed.

[0038] In a second aspect of the invention, a system for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients is also provided, comprising: The acquisition module is used to simultaneously acquire multimodal monitoring data of critically ill neurological patients. The multimodal monitoring data includes at least continuous arterial blood pressure waveform data, electroencephalogram signal data, and head computed tomography (CT) image data. The data processing module is used to perform timestamp alignment and preprocessing on the collected multimodal monitoring data to remove noise fragments caused by motion artifacts or equipment interference. The extraction module is used to extract the time-domain and frequency-domain features of arterial blood pressure waveforms, the nonlinear dynamic features of electroencephalogram signals, and the brain tissue morphological features of head computed tomography images based on the preprocessed data. The construction module is used to map the extracted features to a unified high-dimensional space and construct a three-dimensional feature tensor containing spatial, feature, and temporal dimensions. The output module is used to input the three-dimensional feature tensor into the preset deep neural network model, perform spatiotemporal feature fusion operation, capture the nonlinear coupling relationship between cross-modal data, and output the continuous predicted value of intracranial pressure within the preset time window in the future. The calculation module is used to perform residual analysis based on the continuous predicted intracranial pressure value and the gold standard value of invasive intracranial pressure monitoring collected in real time, and to calculate the dynamic early warning threshold. The early warning module is used to generate a graded early warning signal when the continuously predicted value of intracranial pressure exceeds the dynamic early warning threshold and the duration is longer than the preset duration.

[0039] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients, characterized in that, include: Simultaneously collect multimodal monitoring data from critically ill neurological patients. The multimodal monitoring data includes at least continuous arterial blood pressure waveform data, electroencephalogram signal data, and head computed tomography (CT) image data. The collected multimodal monitoring data is time-stamp aligned and preprocessed to remove noise fragments caused by motion artifacts or equipment interference; Based on the preprocessed data, the time-domain and frequency-domain features of arterial blood pressure waveforms, the nonlinear dynamic features of electroencephalogram signals, and the brain tissue morphological features of head computed tomography images were extracted. The extracted features are mapped to a unified high-dimensional space to construct a three-dimensional feature tensor containing spatial, feature, and temporal dimensions. The three-dimensional feature tensor is input into a preset deep neural network model, and a spatiotemporal feature fusion operation is performed to capture the nonlinear coupling relationship between cross-modal data and output the continuous prediction value of intracranial pressure within a preset future time window. Residual analysis was performed between the continuously predicted intracranial pressure values ​​and the gold standard values ​​of invasive intracranial pressure monitoring collected in real time to calculate the dynamic early warning threshold. When the continuously predicted value of intracranial pressure exceeds the dynamic warning threshold and the duration is longer than the preset duration, a graded warning signal is generated.

2. The method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients according to claim 1, characterized in that, The simultaneous acquisition of multimodal monitoring data from critically ill neurological patients specifically includes: Collect continuous arterial blood pressure waveform data at a sampling frequency of 100Hz to 500Hz for a duration of no less than 30 minutes; EEG signal data were acquired at a sampling frequency of 256Hz to 1024Hz, with no fewer than 8 leads. Head computed tomography (CT) images were collected upon patient admission and during changes in the patient's condition, and the raw DICOM format data of the images were extracted. A linear interpolation algorithm was used to resample physiological signals with different sampling rates to a uniform frequency, and the R-wave peak of the EEG signal was used as the time reference point to align the arterial blood pressure waveform with the time axis of the EEG signal.

3. The method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients according to claim 2, characterized in that, The extraction of brain tissue morphological features from head computed tomography images specifically includes: U-Net convolutional neural network was used to perform cranial dissection and brain tissue segmentation on head computed tomography images to obtain mask images of brain parenchyma, ventricles and subarachnoid space. Calculate the ratio of brain parenchyma volume to total intracranial volume. and the maximum transverse diameter of the bilateral lateral ventricle bodies. ; Based on Hounsfield units, the mean CT values ​​of brain parenchyma regions were statistically analyzed. ,variance and skewness The specific statistical formula is as follows: ； in, This represents the total number of voxels in the brain parenchyma region. For the first CT value of individual pixels; The mask image and volume ratio , transverse diameter value The density feature vectors are combined to form the image morphological feature set.

4. The method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients according to claim 3, characterized in that, The extraction of time-domain and frequency-domain features of arterial blood pressure waveforms and nonlinear dynamic features of electroencephalogram signals specifically includes: For arterial blood pressure waveforms, calculate the values ​​of systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse pressure, heart rate variability, and low-frequency / high-frequency power ratio; For EEG signals, calculate the 90% spectral edge frequency, burst suppression ratio, approximate entropy, and sample entropy; Wavelet packet decomposition was performed on the EEG signal to extract the wavelet packet energy entropy of nodes from the 3rd to the 5th layer. The extracted physiological signal features and image morphological features are subjected to Z-score normalization, and the specific calculation formula is as follows: ； in, These are the standardized eigenvalues. These are the original eigenvalues. The mean of this feature for the training set, The standard deviation of this feature in the training set.

5. The method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients according to claim 4, characterized in that, The construction of the three-dimensional feature tensor, which includes spatial, feature, and temporal dimensions, specifically includes: The standardized arterial blood pressure feature vector With EEG feature vectors Concatenate along the feature dimension to form a one-dimensional feature vector sequence. ; Morphological feature vectors extracted from head computed tomography images The replication expands into a sequence matrix with a duration equal to that of the physiological signal. ; Based on the time window length The sliding step size is The above sequence data is sliced ​​into multiple data segments; The physiological features, EEG features, and imaging features within each data segment are stacked spatially to generate a shape of... 3D feature tensor The specific construction logic satisfies: ； in, The number of feature channels, This is a spatial feature dimension, typically 1 or the number of brain regions. For time step, and = , These are channel, spatial, and temporal indexes, respectively.

6. The method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients according to claim 5, characterized in that, The spatiotemporal feature fusion operation specifically includes: The three-dimensional feature tensor is input into the three-dimensional convolutional neural network layer, and the local spatiotemporal features are extracted using a 3×3×3 convolutional kernel. The convolutional layer is followed by a batch normalization layer and a ReLU activation function. The output of the 3D convolution is flattened and then fed into a long short-term memory network layer to capture long-term temporal dependencies. A channel attention mechanism module is inserted between the 3D convolutional network and the long short-term memory network, and the weight coefficients of each feature channel are calculated through global average pooling and fully connected layers. The specific calculation formula is as follows: ； in, For the input feature map, This is a global average pooling operation. and This is the weight matrix of the fully connected layer. It is the ReLU activation function. Use the Sigmoid activation function; The input and output of the 3D convolution are added together using a residual connection structure to prevent gradient vanishing.

7. The method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients according to claim 6, characterized in that, The output of continuous predicted values ​​of intracranial pressure within a future preset time window specifically includes: Set the prediction time window to the future. minutes, time step is minute; The output layer of the neural network model uses a fully connected layer and a linear activation function, and the output dimension is a vector of intracranial pressure values ​​equal to the length of the prediction time window. ； During the model training phase, the loss function Using the mean square error loss function With smoothing L1 loss function The weighted combination is calculated using the following formula: ； ； ； in, and Weighting coefficients =0.7, , This represents the gold standard value for true intracranial pressure. These are the model's predicted values. This represents the number of samples.

8. The method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients according to claim 7, characterized in that, The calculation of the dynamic early warning threshold specifically includes: Obtain patient past The average value of invasive intracranial pressure monitoring data within one hour was used as the baseline. ; Calculate baseline Standard deviation ; Set individualized early warning coefficient ,in The value range is 1.5≤ ≤3.0, and Values ​​vary with the patient's brain compliance index The decrease in linearly increases the value, and the specific relationship is as follows: ； Among them, among them, =1.5, =3.0, This is the reference value for normal brain compliance; The formula for calculating the dynamic warning threshold (Threshold) is as follows: ； When the predicted value is greater than and less than A Level 1 warning is triggered when the predicted value is greater than or equal to... A level-two warning is triggered at this time, among which =5mmHg.

9. The method for dynamic monitoring and graded intervention of intracranial pressure in critically ill neurological patients according to claim 8, characterized in that, Before generating a graded warning signal when the continuously predicted value of intracranial pressure exceeds the dynamic warning threshold and the duration is greater than the preset duration, a false positive suppression step is also included: Morphological filtering is performed on the continuous intracranial pressure prediction values ​​output by the neural network to remove spurious noise with a width smaller than the preset pixel width. Calculate the slope of the predicted value curve near the warning threshold. The specific calculation formula is as follows: ； in, The predicted value at the current moment. The sampling interval; If the slope If the value exceeds the preset threshold for a sharp increase, it is determined to be an acute event; By combining the instantaneous changes in arterial blood pressure waveform, if the increase in intracranial pressure is accompanied by a synchronous increase in mean arterial pressure (MAP) and the appearance of slow waves on the electroencephalogram (EEG), the warning signal is confirmed to be effective; if only the intracranial pressure increases while the hemodynamic and EEG characteristics remain stable, the issuance of the warning signal is suppressed.

10. A dynamic monitoring and grading intervention system for intracranial pressure in critically ill neurological patients, used to implement the method for dynamic monitoring and grading intervention for intracranial pressure in critically ill neurological patients as described in any one of claims 1-9, characterized in that, include: The acquisition module is used to simultaneously acquire multimodal monitoring data of critically ill neurological patients. The multimodal monitoring data includes at least continuous arterial blood pressure waveform data, electroencephalogram signal data, and head computed tomography (CT) image data. The data processing module is used to perform timestamp alignment and preprocessing on the collected multimodal monitoring data to remove noise fragments caused by motion artifacts or equipment interference. The extraction module is used to extract the time-domain and frequency-domain features of arterial blood pressure waveforms, the nonlinear dynamic features of electroencephalogram signals, and the brain tissue morphological features of head computed tomography images based on the preprocessed data. The construction module is used to map the extracted features to a unified high-dimensional space and construct a three-dimensional feature tensor containing spatial dimension, feature dimension and time dimension. The output module is used to input three-dimensional feature tensors into a preset deep neural network model, perform spatiotemporal feature fusion operations, capture the nonlinear coupling relationship between cross-modal data, and output continuous predicted values ​​of intracranial pressure within a preset future time window. The calculation module is used to perform residual analysis based on the continuous predicted value of intracranial pressure and the gold standard value of invasive intracranial pressure monitoring collected in real time, and to calculate the dynamic early warning threshold. The early warning module is used to generate a graded early warning signal when the continuously predicted value of intracranial pressure exceeds the dynamic early warning threshold and the duration is greater than a preset duration.

Images