Intelligent non-invasive intracranial pressure monitoring and analyzing system for PICU

Abstract

The invention relates to the technical field of intracranial pressure monitoring, in particular to an intelligent non-invasive intracranial pressure monitoring and analysis system for a PICU, which integrates a multi-modal sensing array of a rigid packaging high-frequency ultrasonic probe and a flexible transcranial Doppler probe to realize conformal contact with the head and neck of a child. And synchronously acquiring an optic nerve sheath ultrasonic image and a middle cerebral artery blood flow spectrum. The system ensures data reliability through a signal quality evaluation and self-adaptive enhancement module, and a hybrid deep learning fusion core is used for estimating intracranial pressure from multi-source features with high precision and giving a confidence interval. The system is internally provided with an intracranial environment dynamic simulation and prediction engine, deduces a compensatory reserve state and predicts a future change trend based on a volume-pressure model, and finally outputs a personalized Pareto optimal intervention strategy sequence through a decision module driven by reinforcement learning. According to the PICU intracranial pressure management system and method, full-process support from non-invasive monitoring and accurate prediction to intelligent decision making is achieved, and the technical level and clinical efficacy of PICU intracranial pressure management are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intracranial pressure monitoring technology, specifically to an intelligent non-invasive intracranial pressure monitoring and analysis system for PICU. Background Technology

[0002] Intracranial pressure monitoring is a core tool for assessing the neurological status of critically ill children in the PICU. Sustained elevation of intracranial pressure can severely impair cerebral perfusion, leading to irreversible neurological damage and even death. Therefore, accurate and continuous monitoring is of paramount clinical importance. Currently, the clinical gold standard still relies on invasive monitoring techniques, such as intraventricular or intraparenchymal probe placement. However, these methods carry significant risks, including intracranial infection, hemorrhage, and tissue damage, with particularly pronounced limitations in pediatric patients. Especially for very young children or those with unstable conditions, the feasibility of invasive procedures is low, restricting their widespread use in early intervention and long-term monitoring.

[0003] To overcome the limitations of invasive monitoring, various non-invasive or minimally invasive techniques have been proposed, such as transcranial Doppler ultrasound, optic nerve sheath diameter measurement, and fundus examination. However, existing non-invasive methods often rely on a single signal or parameter, making them susceptible to individual differences, operator experience, and physiological noise interference, resulting in insufficient measurement accuracy and robustness. For example, while relying solely on optic nerve sheath diameter can indirectly reflect intracranial pressure, it is easily affected by intraorbital pressure and measurement location; while using only transcranial Doppler blood flow parameters is difficult to maintain sensitivity under low flow velocity conditions. Furthermore, most systems lack real-time judgment and dynamic compensation capabilities for signal quality, and cannot adapt to changes in the child's position or agitation during monitoring, further limiting their reliability in real clinical settings.

[0004] Current technologies have not yet effectively achieved deep fusion and dynamic interpretation of multimodal signals, nor do they provide quantitative assessments of intracranial compensatory capacity and disease progression trends. Most systems stop at parameter display, failing to integrate child-specific information, provide individualized risk prediction, or form closed-loop clinical decision support. In the special group of pediatrics, due to differences in cranial cavity development and rapid disease progression, general monitoring solutions often fall short of the needs of precision medicine. Therefore, there is an urgent need for a non-invasive intracranial pressure monitoring and analysis system that can adapt to the physiological characteristics of children, integrate multi-source information, and possess intelligent analysis and dynamic optimization capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent non-invasive intracranial pressure monitoring and analysis system for PICU. It collects physiological signals through a multimodal sensor array and uses a hybrid deep learning model and intracranial dynamics simulation to achieve high-precision non-invasive intracranial pressure estimation, trend prediction, and personalized intervention strategy recommendation, thereby improving the accuracy and safety of pediatric intensive care.

[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: A smart non-invasive intracranial pressure monitoring and analysis system for PICU includes: A multimodal physiological sensing array, consisting of a rigidly packaged high-frequency linear ultrasound probe and a flexiblely packaged transcranial Doppler probe integrated onto an adjustable elastic headband, is used to achieve conformal contact with the child's head and neck, simultaneously acquiring B-mode ultrasound image sequences of the optic nerve sheath and Doppler spectrum signals of blood flow in the middle cerebral artery; the center frequency of the high-frequency linear ultrasound probe is not less than 15MHz, and the transmission frequency of the transcranial Doppler probe is 2MHz; The signal quality assessment and adaptive enhancement module has a built-in lightweight convolutional neural network based on the MobileNetV2 architecture. It performs real-time quality scoring on each frame of ultrasound image and each segment of Doppler spectrum, outputs a quality score Q between 0 and 1, and dynamically adjusts the signal acquisition parameters based on this score, while triggering the corresponding motion artifact compensation algorithm. The multi-level feature extraction and spatiotemporal alignment module is used to extract morphological features, spectral dynamic features and nonlinear features related to intracranial pressure from the enhanced signal, and to uniformly resample the feature sequences extracted at different frequencies onto a 100Hz time grid through cubic spline interpolation to form a feature matrix with a unified timestamp. The hybrid deep learning fusion core employs a dual-branch network structure, with one branch being an encoder based on a multi-head self-attention mechanism and the other branch being a temporal encoder based on a gated recurrent unit. The outputs of the two branches are integrated through a gated fusion unit, ultimately outputting a point estimate of intracranial pressure, an 80% prediction interval based on quantile regression, and a confidence score representing the reciprocal of the prediction interval width. The intracranial environment dynamic simulation and prediction engine takes the intracranial pressure sequence output by the fusion core as the observation value and embeds it into a simplified intracranial volume-pressure dynamic differential equation model. It performs state estimation through extended Kalman filtering, thereby deduce the intracranial compensatory reserve state quantity that cannot be directly measured, and makes multi-step predictions on the future intracranial pressure change trajectory based on the current state. The personalized clinical decision automation module integrates a reinforcement learning strategy network trained based on a proximal strategy optimization algorithm. This network takes the current intracranial environment state, predicted trajectory and individual patient metadata as input, outputs a set of Pareto optimal intervention strategy sequences, and estimates the potential long-term benefits of each strategy. The system employs a self-learning and adaptive update cycle. After each monitoring cycle, the system automatically generates an encrypted data snapshot and decision log for that monitoring period. When the system detects that the mean absolute error between its prediction and the actual clinical outcome of the child continuously exceeds a preset threshold, it will initiate a local incremental learning process. The elastic weight consolidation regularization method is used to fine-tune the parameters of the last few fully connected layers of the hybrid deep learning fusion core, thereby achieving personalized adaptation based on specific child data.

[0007] Furthermore, the specific workflow of the signal quality assessment and adaptive enhancement module is as follows: Quality Score: The lightweight CNN analyzes the input single-frame 256x256 pixel ultrasound image or a 2-second Doppler spectrogram and outputs a quality score Q. us For ultrasound images, the network assesses the sharpness of the optic nerve sheath boundary; for Doppler signals, it assesses the signal-to-noise ratio and the integrity of the spectral envelope. Parameter Adaptation and Motion Artifact Compensation: The system calculates the mean of ultrasound and Doppler mass fractions. ;when When the value is less than 0.7, the system automatically increases the ultrasound transmission power and adjusts the TCD focusing depth, and initiates the following algorithm: For ultrasound image sequences, a motion estimation and compensation algorithm based on the Lucas-Kanade optical flow method is used to stabilize the sequences; For the Doppler spectrum, an improved spectral subtraction algorithm is used: in: It is the power spectral density of the noisy signal. The noise power spectrum is dynamically estimated during the signal silence period using a recursive least squares algorithm. To the mean of the mass fraction The negatively correlated adaptive decay factor, the specific value of which is determined by a pre-calibrated function. =1.5- Sure.

[0008] Furthermore, the features extracted by the multi-level feature extraction and spatiotemporal alignment module include: From ultrasound images of the optic nerve sheath, its contour was extracted using a U-Net segmentation network, and the maximum diameter ONSD within one heartbeat cycle was calculated. max Minimum diameter ONSD min and pulsation amplitude ONSD amp ; Extract the pulsatility index PI and spectral entropy H from the TCD spectrum.spectrum And the diastolic velocity decay time constant τ, where τ is obtained by fitting the diastolic velocity curve: Calculate the sample entropy of the TCD flow velocity signal and the ONSD pulsation signal.

[0009] Furthermore, the specific implementation of the dual-branch network structure and gated fusion unit of the hybrid deep learning fusion core is as follows: Multi-head self-attention branch: This branch takes the spatiotemporally aligned feature matrix F(t) as input and processes it through an encoder layer containing four attention heads, where the projection dimension d of the query, key, and value matrices is... model Set to 128; GRU Temporal Branch: This branch receives the same feature matrix and passes through a two-layer bidirectional GRU network with a hidden layer dimension of 64. Gated fusion unit: Let the output of the self-attention branch be The final hidden state of the GRU branch is The fusion gate vector g is calculated as follows: The final fusion characterization is as follows: Output layer: h fusion Intracranial pressure was estimated by regression using a fully connected layer. It also outputs its 80% prediction interval using quantile regression technology. , The confidence score C is then calculated as follows: in It is a constant added to prevent division by zero errors.

[0010] Furthermore, the core of the intracranial environment dynamic simulation and prediction engine is the state-space equation based on the volume-pressure model: Equations of state: ,in From the aforementioned differential equation: in: P represents intracranial pressure, I f (t) represents the rate of cerebrospinal fluid production, set as a constant of 0.35 mL / min. out (t) represents the cerebrospinal fluid outflow rate, modeled as follows: Where P vs=5 mmHg is the sagittal sinus pressure, R = 6 mmHg·mL -1 ·min is the resistance to cerebrospinal fluid outflow, and w(t) is the process noise. It is the compliance of the cranial cavity, where k and P0 are individualized parameters estimated in real time through extended Kalman filtering; Observation equation: Where H = [1, 0, 0], and v t It is observation noise, and its covariance R is determined by the confidence score C output by the fusion core, specifically... ; State Estimation and Prediction: The engine uses an extended Kalman filter to continuously estimate the system's state vector. Based on the current state estimate, the engine performs numerical integration on the differential equation to predict future intracranial pressure and compensatory reserve status.

[0011] Furthermore, the reinforcement learning strategy network architecture in the personalized clinical decision automation module is as follows: State space: Where R(t) is the compensation reserve state, It shows the trend of intracranial pressure changes; Action space: a t Defined as a discrete set of intervention actions; Reward function: Designed as a multi-objective weighted sum: in =15mmHg, =0.85, It is a function that quantifies the risk of intervention measures, where w1, w2, and w3 are weighting coefficients; Policy generation: The policy is learned through offline training using a proximal policy optimization algorithm. In practical applications, for a given state s t The network outputs the Top-K actions with the highest probabilities and their estimated values, forming a Pareto optimal policy set.

[0012] Furthermore, the specific steps of the system's self-learning and adaptive update cycle include: S1: Data Snapshot Generation: For a completed monitoring session, the system packages key data into an encrypted data snapshot; S2: Deviation Detection: The system compares its own predictions with the recorded actual clinical outcomes; if the average absolute error of 5 consecutive monitoring sessions exceeds the threshold of 3 mmHg, a learning loop is triggered. S3: Incremental Learning: The system uses a snapshot of the patient's data at an extremely low learning rate. Only the parameters of the last few fully connected layers in the hybrid deep learning fusion core are fine-tuned; training employs an elastic weight consolidation regularization method with the following loss function: in It is the mean squared error loss on the new data. These are the parameters of the old model, F i It is its Fisher information content, regularization strength Set to 100.

[0013] The beneficial effects of this invention are: This invention provides a robust, non-invasive, and precise system for monitoring intracranial pressure in complex clinical environments. It employs a multimodal sensor array integrating a rigidly packaged high-frequency ultrasound probe and a flexible transcranial Doppler probe. This ensures the stability of key acoustic components and achieves conformal contact with the irregular head and neck positions of children, physically guaranteeing the fundamental quality of signal acquisition. Building upon this, a lightweight CNN based on the MobileNetV2 architecture is introduced for real-time signal quality scoring, driving an adaptive enhancement process: when the quality score falls below a threshold, the system automatically adjusts hardware parameters such as ultrasound power and depth of focus, and simultaneously initiates targeted software algorithms—using the Lucas-Kanade optical flow method for motion compensation of the ultrasound image sequence and an improved spectral subtraction algorithm based on recursive least squares noise estimation for the Doppler spectrum. This effectively suppresses motion artifacts and noise interference caused by pediatric agitation and probe displacement at the signal source, ensuring the continuous acquisition of high-quality physiological signals even in the highly disruptive environment of the PICU, laying a solid foundation for the reliability of subsequent analysis.

[0014] This invention achieves a comprehensive, high-information assessment of intracranial status through deep fusion of multi-source heterogeneous features and physiological mechanism modeling. At the feature level, the system does not simply use raw signals, but extracts clinically significant morphological pulsation features, including ONSD, from optic nerve sheath ultrasound images using a U-Net network. max ONSD min ONSD ampSimultaneously, dynamic features such as pulsatility index, spectral entropy, and diastolic flow velocity decay time constant are calculated from the TCD spectrum, and nonlinear features such as sample entropy are further extracted, thus constructing a multi-level feature system covering morphology, dynamics, and complexity. The system uses a hybrid deep learning fusion core to precisely align and deeply fuse the features collected at different frequencies along the time axis: the multi-head self-attention branch excels at capturing instantaneous correlations between different features, while the GRU temporal branch effectively models the dynamic trajectory of intracranial pressure evolution over time; the two are adaptively integrated through a gated fusion unit, ultimately outputting not only point estimates but also prediction intervals and confidence scores through quantile regression. This ensures that each intracranial pressure estimate carries a quantifiable measure of uncertainty, greatly improving the precision and reliability of clinical decision-making.

[0015] This invention introduces a dynamic simulation and prediction engine for the intracranial environment. This engine uses the intracranial pressure sequence output by a deep learning module as an observation, embedding it into a simplified but physiologically meaningful intracranial volume-pressure differential equation model. It then employs the classic and efficient extended Kalman filter (EPF) state estimation algorithm to calculate key physiological parameters that cannot be directly measured—intracranial compliance characteristics and compensatory reserve status—in real time. Based on this dynamic model, the system can make multi-step predictions of intracranial pressure changes over a future period, achieving a leap from current state perception to future risk warning. The integrated reinforcement learning strategy network takes the current state, predicted trajectory, and individual patient metadata as input. Within a multi-objective reward function framework that comprehensively considers ICP control objectives, compensatory reserve risks, and intervention costs, it learns and outputs a set of Pareto-optimal intervention strategy sequences through a proximal strategy optimization algorithm. This provides clinicians with decision support that is not only based on the current situation but also considers future trends and individual differences, and is a risk-benefit trade-off, significantly improving the predictability and accuracy of intracranial hypertension management in the PICU.

[0016] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 As exemplified in this embodiment, an intelligent non-invasive intracranial pressure monitoring and analysis system for PICUs is provided. The modules are interconnected through a data bus and control logic to form a closed loop from physical signal acquisition to intelligent decision support.

[0019] In this embodiment, the multimodal physiological sensing array consists of a rigidly packaged high-frequency linear ultrasound probe (center frequency ≥15MHz) and a flexiblely packaged transcranial Doppler probe (emission frequency 2MHz) co-molded onto an elastic urethane headband using medical-grade silicone. The headband uses Velcro and elastic adjustment buckles to accommodate different head circumferences from newborns to adolescents.

[0020] A high-frequency linear ultrasound probe is used to acquire high-resolution B-mode images of the optic nerve sheath at the back of the eye, with an axial resolution of <0.1 mm. Its rigid encapsulation ensures the stability of the acoustic lens, while the surrounding flexible material cushions the contact pressure with the child's skin. The transcranial Doppler probe uses phased array technology, and its flexible encapsulation allows it to better fit the temporal window area for continuous monitoring of blood flow velocity in the middle cerebral artery. Data from both probes is transmitted to the signal processing unit via an integrated shielded cable.

[0021] In this embodiment, the signal quality assessment and adaptive enhancement module includes a pre-trained lightweight convolutional neural network (CNN), such as a variant of MobileNetV2.

[0022] For ultrasound images, the network input is a single frame of 256x256 pixel grayscale image. The output is the quality score Q. us During the training phase, three expert physicians labeled tens of thousands of images with 0 / 1 (1 for clear, 0 for blurry) based on the sharpness of the optic nerve sheath boundary (continuous boundary, no ghosting). The network was trained using cross-entropy loss, and the final output probability value is Q. us .

[0023] For the Doppler spectrum, the network input is a 2-second spectrum with a frequency range of 0-200Hz. The output quality score is Q. tcd The training data was also labeled by experts based on signal-to-noise ratio and spectral envelope integrity.

[0024] Adaptive enhancement process: The system calculates Q at a frequency of 1Hz. us and Q tcd mean .when When the value is less than 0.7, the system control module automatically executes: 1. Increase the transmission power of the ultrasound probe (up to 20%) and fine-tune the focusing depth.

[0025] 2. Activate the motion artifact compensation algorithm: For images: the Lucas-Kanade optical flow method is used to calculate the global motion vector between consecutive frames, and reverse translation and rotation are performed to compensate.

[0026] For the spectrum: Perform the spectral subtraction algorithm: The noise spectrum is estimated and updated during the signal silence period using the Recursive Least Squares (RLS) algorithm. Adaptive attenuation factor. The calculation formula is: =1.5- This means that the lower the quality, the more noise is reduced.

[0027] In this embodiment, the multi-level feature extraction and spatiotemporal alignment module processes two signals in parallel: Optic nerve sheath morphological feature extraction: Using a U-Net segmentation network pre-trained on a medical image dataset and fine-tuned on domain-specific data, semantic segmentation is performed on each frame of ultrasound image to extract the contour of the optic nerve sheath. Subsequently, calculations are performed: ONSD max ONSD min The maximum and minimum diameter of the optic nerve sheath during a complete cardiac cycle (determined by synchronized ECG signals or TCD spectrum).

[0028] ONSD amp =ONSD max -ONSD min : Pulse amplitude.

[0029] The extraction frequency is approximately 30Hz (depending on the image frame rate).

[0030] TCD spectral dynamics feature extraction: Calculate the following every 0.5 seconds for the preprocessed TCD spectral signal: Pulsatility Index (PI): The diastolic velocity decay time constant (τ) is obtained by fitting the diastolic velocity curve using a nonlinear least squares method; the model is as follows: The τ obtained from the fitting is the feature.

[0031] Spectral entropy (H): Where P k It is the normalized power of the current spectrum frame at the k-th frequency bin.

[0032] The extraction frequency is 2Hz.

[0033] Spatiotemporal alignment: All features (ONSD-related features, PI, τ, H) are uniformly resampled to a 100Hz temporal grid. For ONSD features, due to their low native frequency, cubic spline interpolation is used for upsampling to form a uniform feature matrix. , where N is the total number of features and T is the number of time points.

[0034] The fusion core described in this embodiment is implemented using the PyTorch or TensorFlow framework, including: Multi-head self-attention branch: The feature matrix F(t) is first normalized through a layer, and then input into an encoder layer containing four attention heads. The query, key, and value matrices are obtained through linear transformations, and their dimension d... model Set to 128. This branch aims to discover deep cross-feature associations, such as high PI values ​​accompanied by a rapid increase in ONSD.

[0035] GRU temporal branch: A two-layer bidirectional GRU is used with a hidden layer dimension of 64. This branch is used to learn the evolutionary pattern of intracranial pressure over time, such as a slow increase or a sharp fluctuation trend.

[0036] Gated fusion unit: The output of the self-attention branch (take the CLS tag bit or the vector after average pooling) is concatenated with the final hidden state of the GRU branch, and a gated vector g is generated through a fully connected layer activated by Sigmoid.

[0037] Output layer: fused representation h fusion Through two parallel fully connected layers: A single layer outputs a single estimate of intracranial pressure. (Unit: mmHg).

[0038] Another layer uses quantile regression to output the values ​​corresponding to the 10th and 90th quantiles. and This forms the 80% prediction interval. The confidence score C is calculated as follows: in It is a small constant added to prevent division by zero errors.

[0039] In this embodiment, the intracranial environment dynamic simulation and prediction engine is implemented in C++ or Python on an embedded system or server, including: State-space model: State vector: , where P t It's intracranial pressure, k. t and P 0，t It is an individualized cranial cavity compliance parameter.

[0040] State equations (after discretization): .in From the aforementioned differential equation: It is obtained by discretization using the Euler method or the Runge-Kutta method. t It is process noise.

[0041] Observation equation: Where H = [1, 0, 0], and the observation noise v t The covariance R is set to That is, the lower the confidence score, the greater the observation noise.

[0042] Extended Kalman Filter (EKF) Procedure: Prediction step: Based on the state estimate of the previous time step Covariance P t-1|t-1 Predict the current state Covariance P t-1|t-1 .

[0043] Update step: When new observations are observed When the output of the fusion core arrives, calculate the Kalman gain K. t And use it to correct the predicted state to obtain the optimal state estimate. .

[0044] Predicted output: The engine estimates the output every 5 minutes based on the current state. A forward simulation of the differential equation is performed for 15 minutes, outputting the predicted trajectory of future intracranial pressure and the compensatory reserve index R(t).

[0045] In this embodiment, the modular policy network for automating personalized clinical decision-making is trained offline using Ray or a similar reinforcement learning framework, including: Offline training: Environment: A simulation environment was built using historical PICU patient data (de-identified).

[0046] Status: s t Including normalized R(t), the first derivative of ICP, and age, weight, and Glasgow Coma Scale score.

[0047] Reward function: Among them, Cost(a t The value was 0 for observation, 2 for mannitol 1 g / kg, and 5 for hyperventilation (due to its potential risks).

[0048] Training algorithm: The Proximal Policy Optimization (PPO) algorithm is used to train the policy until it converges.

[0049] Online applications: for real-time status s t The trained policy network The system outputs the probability distribution of all possible actions. It selects the top-3 actions with the highest probabilities and displays their estimated Q-values ​​(long-term returns), forming a Pareto-optimal strategy set, which is presented in list form on the UI for doctors' decision-making reference.

[0050] In this embodiment, the system self-learning and adaptation update cycle is triggered in the background when the system is idle (such as after monitoring ends), including: Data snapshot: Stored feature sequences, ICP estimates, EKF status, clinical intervention records, and final discharge diagnoses are stored in encrypted form.

[0051] Deviation detection: The system calculates the mean absolute error (MAE) between the predicted ICP and the invasive monitoring data (if subsequent invasive monitoring was performed) or the gold standard assessment based on the physician's comprehensive clinical judgment during this monitoring period. If the MAE continuously exceeds 3 mmHg, incremental learning is triggered.

[0052] Incremental learning: The system loads the pre-trained fusion core model and freezes all parameters except the last two fully connected layers. Then, using a snapshot of the patient's data, a learning rate of 10 is applied. -5 Fine-tuning is performed over a maximum of 10 epochs. During training, an Elastic Weight Consolidation (EWC) regularization term is added to the loss function. .in, These are the parameters of the original model, F i These are the diagonal elements of the Fisher information matrix calculated on the original training set. Set it to 100. This effectively protects existing knowledge from being forgotten while also adapting to the characteristics of new patients.

[0053] In summary, this invention proposes an intelligent non-invasive intracranial pressure monitoring and analysis system for PICU use. It integrates a multimodal sensor array of a rigid-encapsulated high-frequency ultrasound probe and a flexible transcranial Doppler probe, achieving conformal contact with the child's head and neck area and simultaneously acquiring ultrasound images of the optic nerve sheath and the blood flow spectrum of the middle cerebral artery. The system ensures data reliability through signal quality assessment and adaptive enhancement modules, and utilizes a hybrid deep learning fusion core to accurately estimate intracranial pressure from multi-source features and provide confidence intervals. Furthermore, the system incorporates a built-in intracranial environment dynamic simulation and prediction engine, extrapolating the compensatory reserve state and predicting future trends based on a volume-pressure model. Finally, a reinforcement learning-driven decision-making module outputs a personalized Pareto-optimal intervention strategy sequence. This invention achieves end-to-end support from non-invasive monitoring and accurate prediction to intelligent decision-making, significantly improving the technical level and clinical efficacy of intracranial pressure management in the PICU.

[0054] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An intelligent non-invasive intracranial pressure monitoring and analysis system for PICU, characterized in that, The system comprises: a multi-modal physiological sensor array, including an ultrasonic probe and a transcranial Doppler probe integrated on an elastic headband, for realizing conformal contact with the head and neck of a child, and synchronously collecting B-mode ultrasonic image sequences of the optic nerve sheath and blood flow Doppler spectrum signals of the middle cerebral artery; a signal quality assessment and adaptive enhancement module, which is internally provided with a convolutional neural network, for performing real-time quality scoring on the collected ultrasonic images and Doppler spectrum, and dynamically adjusting signal acquisition parameters and triggering motion artifact compensation algorithms based on the quality score; a multi-level feature extraction and spatio-temporal alignment module, for extracting morphological features, spectral dynamics features and nonlinear features related to intracranial pressure from the quality-enhanced signals, and uniformly resampling the feature sequences extracted at different frequencies to a uniform time grid to form a feature matrix; a hybrid deep learning fusion core, which adopts a double-branch network structure including an encoder branch based on a multi-head self-attention mechanism and a time series encoder branch based on a gated recurrent unit, and integrates the outputs of the double branches through a gated fusion unit to finally output point estimates, prediction intervals and confidence scores of intracranial pressure; an intracranial environment dynamic simulation and prediction engine, for embedding the intracranial pressure sequence output by the fusion core as an observation value into a cranial cavity volume-pressure dynamics differential equation model, performing state estimation through an extended Kalman filter, calculating cranial compensatory reserve state variables, and performing multi-step prediction of future intracranial pressure change trajectories based on the current state; an individualized clinical decision automation module, which integrates a reinforcement learning strategy network trained based on a reinforcement learning algorithm, takes the current intracranial environment state, prediction trajectory and individual metadata of the child as input, and outputs a Pareto-optimal intervention strategy sequence; a system self-learning and adaptive updating cycle, for generating an encrypted data snapshot and a decision log at the end of a monitoring period, and starting an incremental learning process to fine-tune the parameters of the hybrid deep learning fusion core when the error between the system prediction and the actual clinical outcome of the child continuously exceeds a preset threshold, to realize individualized adaptation.

2. The system of claim 1, wherein, In the signal quality assessment and adaptive enhancement module: The lightweight convolutional neural network is based on the MobileNetV2 architecture, and is used to analyze the input single-frame ultrasonic image or Doppler spectrum graph with a time length of 2 seconds, and output a quality score between 0 and 1. The module calculates the mean of the ultrasonic and Doppler quality scores, and when the mean is lower than a first threshold, automatically increases the ultrasonic transmission power and adjusts the transcranial Doppler focusing depth, and starts the motion artifact compensation algorithm. For ultrasonic image sequences, a motion estimation and compensation algorithm based on optical flow method is used; for Doppler spectrum, an improved spectral subtraction algorithm is used. wherein is the power spectral density of the noise-containing signal, is the noise power spectrum dynamically estimated by a recursive least squares algorithm, is an adaptive attenuation factor negatively correlated with the mass fraction mean.

3. The system of claim 1, wherein, The features extracted by the multi-level feature extraction and spatio-temporal alignment module include: From the optic nerve sheath ultrasound images, the contour thereof is extracted by a segmentation network, and the maximum diameter ONSD max , the minimum diameter ONSD min and the pulsatility amplitude ONSD amp in a heartbeat cycle are calculated; From the transcranial Doppler spectrum, the pulsatility index PI, the spectral entropy H are extracted spectrum and the diastolic flow velocity decay time constant τ by fitting the diastolic flow velocity curve are obtained. The sample entropy of the transcranial Doppler flow velocity signal and the optic nerve sheath diameter pulsation signal is also calculated.

4. The system of claim 1, wherein, In the hybrid deep learning fusion core: The multi-head self-attention branch takes the spatio-temporally aligned feature matrix as input, and processes it through an encoder layer containing multiple attention heads; The gating recurrent unit timing branch receives the same feature matrix, and is processed by a multi-layer bidirectional gating recurrent unit network; The gating fusion unit calculates a gating vector and obtains a final fusion representation based on this where h attn is the output of the self-attention branch, h gru is the final hidden state of the gating recurrent unit branch; The output layer regresses the intracranial pressure estimate h fusion The intracranial pressure estimate h is regressed by a fully connected layer And its prediction interval is output by quantile regression technique, the confidence score C is calculated as: wherein is a constant.

5. The system of claim 1, wherein, The core of the intracranial environment dynamic simulation and prediction engine is a state space equation based on a volume-pressure model: The equation of state is where is described by the differential equation: where P is the intracranial pressure, I f (t) represents the cerebrospinal fluid production rate, I out (t) represents the cerebrospinal fluid outflow rate, is the compliance of the cranial cavity, where k and P0are individualized parameters estimated in real time by an extended Kalman filter, w(t) is the process noise; The observation equation is where H is the observation matrix, v t is the observation noise, the covariance R of which is determined by the confidence score C output by the fusion core, specifically ; The engine uses an extended Kalman filter to continuously estimate the system's state vector and performs a numerical integration based on the current state estimate to achieve a prediction.

6. The system of claim 1, wherein, The reinforcement learning strategy network architecture in the individualized clinical decision automation modeling module is as follows: State space s t including intracranial pressure estimate , reserve status R(t), intracranial pressure change trend , and patient age, weight, and Glasgow coma score ; Action space a t is a discrete set of intervention actions; Reward function Designed as a multi-objective weighted sum: wherein is the target intracranial pressure, is the compensatory reserve threshold, is a function quantifying the risk of the intervention, w1, w2, w3 are weight coefficients; The policy network is trained by a proximal policy optimization algorithm, for a given state s t , outputs the Top-K actions with the highest probabilities and their estimated values, forming a Pareto optimal policy set.

7. The system of claim 1, wherein, The specific steps of the system self-learning and adaptive updating cycle include: S1: Data snapshot generation: for a completed monitoring session, key data is packaged into an encrypted data snapshot; S2: Compare the system prediction with the recorded actual clinical outcome, if the average absolute error of multiple consecutive monitoring sessions exceeds the preset threshold, trigger the learning cycle; S3: Use the data snapshot of the child to fine-tune only the parameters of the last few fully connected layers in the mixed deep learning fusion core with a low learning rate; the training adopts an elastic weight consolidation regularization method, and the loss function is: where is the loss on new data, is the old model parameters, F i is its Fisher information, is the regularization strength.

8. The system of claim 1, wherein, The center frequency of the high-frequency linear ultrasonic probe is not less than 15 MHz, and the transmission frequency of the transcranial Doppler probe is 2 MHz.

Images