An integrated system for real-time monitoring and early warning of hemodynamics in critically ill ICU patients

Abstract

This invention relates to the field of intensive care technology, specifically to an integrated system for real-time hemodynamic monitoring and early warning of critically ill ICU patients. The system includes: a physiological signal synchronous acquisition and quality enhancement module for acquiring and processing multi-channel physiological waveforms; a core physiological state vector calculation module, based on a first-principles model of the circulatory system, for real-time calculation of the left ventricular maximum contractile elasticity Emax, representing myocardial contractility, and the gain coefficient γk, representing preload responsiveness; a dynamic risk assessment and early warning module, which predicts decompensation risk and triggers early warnings by fusing temporal features through a bidirectional gated circulatory unit neural network; a personalized physiological simulator module, which dynamically calibrates a cardiovascular mechanism model using an unscented Kalman filter to generate a physiological image of the patient; and a system adaptive evolution module, which iteratively optimizes model parameters through a two-layer closed-loop mechanism. This invention achieves closed-loop management from accurate perception and risk prediction to personalized decision support.

Description

Technical Field

[0001] This invention relates to the field of intensive care technology, specifically to an integrated system for real-time monitoring and early warning of hemodynamics in critical care ICU patients. Background Technology

[0002] In the field of critical care medicine, continuous and accurate hemodynamic monitoring of critically ill patients, coupled with timely and effective intervention, is crucial for reducing mortality and improving prognosis. Current clinical practice primarily relies on healthcare professionals to integrate and interpret discrete data and waveform signals generated by various invasive and non-invasive monitoring devices. These data come from diverse sources, vary in frequency, and are often subject to noise interference. This forces healthcare professionals to rely on their professional experience to construct a real-time picture of the patient's circulatory status in their minds, which is not only extremely demanding but also prone to information overload or delayed judgment, potentially leading to missed opportunities for optimal intervention.

[0003] While some existing intelligent monitoring systems can provide alarms and trend displays for basic vital signs, their analysis is often limited to threshold judgments of single or a few explicit parameters, lacking in-depth and continuous calculations of the intrinsic physiological state of the cardiovascular system. Furthermore, traditional early warning mechanisms are often based on static thresholds derived from post-hoc statistics, making it difficult to detect complex and dynamically evolving pre-decompensation states in advance, and often resulting in a trade-off between sensitivity and specificity in early warning systems.

[0004] In terms of decision support, existing auxiliary tools are mostly based on fixed treatment rules or population statistical models, which cannot dynamically simulate and predict the unique pathophysiological state of individual patients. When faced with multiple possible treatment options, clinicians often lack a tool that can quantify and predict the hemodynamic impact of each option on a specific patient, making treatment decisions still largely trial-and-error in nature and difficult to achieve true individualization and optimization.

[0005] Therefore, there is an urgent clinical need for an intelligent system that can deeply integrate multi-source heterogeneous physiological data, reveal the key internal state of the circulatory system in real time, and conduct prospective virtual verification and optimization of treatment strategies based on individualized patient models. This system would form a complete closed loop from accurate perception and risk prediction to decision recommendation, assisting clinicians in achieving earlier, more accurate, and more personalized hemodynamic management. Summary of the Invention

[0006] The purpose of this invention is to provide an integrated system for real-time monitoring and early warning of hemodynamics in critically ill ICU patients. By calculating the core parameters of myocardial contractility and volume responsiveness, combined with machine learning, dynamic risk assessment is performed, and individualized physiological models are used to simulate, predict, and optimize treatment strategies, thus forming a closed-loop management system.

[0007] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: An integrated system for real-time hemodynamic monitoring and early warning of critically ill ICU patients includes: The physiological signal synchronous acquisition and quality enhancement module is used to acquire and time-calibrate multi-channel physiological waveforms from invasive and non-invasive monitoring devices, and perform adaptive processing on the waveforms based on quantitative quality scoring to output reliable time-series data. The core physiological state vector calculation module is connected to the physiological signal synchronous acquisition and quality enhancement module. It is used to receive the reliable time-series data and, based on the preset first-principles model of the circulatory system, calculate in real time the internal state vector, which includes at least a first parameter representing the state of myocardial contractility and a second parameter representing the state of preload responsiveness. The dynamic risk assessment and early warning module receives the core physiological state vector and basic vital signs, integrates multi-dimensional temporal features through a machine learning model, calculates the composite risk index of hemodynamic decompensation events occurring within a specific time window in the future, and triggers multi-level warnings based on a preset progressive risk threshold matrix. The personalized physiological simulator module includes a cardiovascular system mechanism model with online updatable parameters, configured to dynamically calibrate model parameters using real-time monitoring data to generate a dynamic physiological mirror synchronized with the current patient status. The treatment intervention virtualization and strategy optimization module responds to the output of the dynamic risk assessment and early warning module and calls the individualized physiological simulator module to generate intervention recommendations based on expected effect ranking by simulating the execution of a set of candidate treatment strategies and quantifying and predicting their hemodynamic response. In addition, the system adaptive evolution module is configured to collect actual clinical intervention and physiological response data, compare them with the prediction results of the treatment intervention virtualization and strategy optimization module, and use the comparison results to drive the parameters of the machine learning model and the cardiovascular system mechanism model to perform iterative optimization.

[0008] Furthermore, in the core physiological state vector calculation module, the first parameter is the maximum contractile elasticity of the left ventricle. The following steps will solve the problem: S1: Identify the start point t0 and end point t1 of the cardiac cycle from the arterial blood pressure waveform; S2: Obtain the stroke volume SV of the cardiac cycle based on pulse contour analysis and estimate the instantaneous ventricular volume V(t), using the following formula: Where V ed To estimate the end-diastolic volume, t0 ≤ t ≤ t1; S3: Construct a time-varying elastic model P(t) = E(t)·[V(t)-V0], where P(t) is arterial blood pressure, E(t) is the time-varying elastic function to be determined, and V0 is the theoretical ventricular volume zero point, which is a constant related to the resting state of the myocardium; S4: Substitute the measured V(t) estimated in step S2 into the model and use the nonlinear least squares method to fit and solve E(t), and define the maximum value of E(t) during the cardiac cycle as E max E max =max(E(t)).

[0009] Furthermore, in the core physiological state vector calculation module, the second parameter, namely the capacity responsiveness gain coefficient... The following online estimation algorithm was used to obtain the result: S1: During the mechanical ventilation cycle, extract continuous... indivual( Stroke volume sequence of the cardiac cycle And calculate the variability of stroke volume: in Indicates the first One respiratory cycle; S2: Set validity criteria: The patient's heart rhythm is sinus rhythm and the tidal volume is... (Ideal weight); current weight is only considered if the criterion is met. Only then is it marked as a valid measurement value. S3: Will The input recursive least squares (RLS) filter is used for online estimation, and the update formula is as follows: in, This is the estimated value of the gain coefficient at the current moment. The regression vector includes preload-related indicators, such as pulse pressure. This is the Kalman gain matrix; the coefficients It directly reflects the slope of the rising segment of the patient's current cardiac function curve (FrankStarling curve).

[0010] Furthermore, in the dynamic risk assessment and early warning module, the composite risk index The calculation method is as follows: S1: Constructing multi-dimensional temporal feature vectors in, Heart rate; Mean arterial pressure; , These correspond to the rate of change of the parameters within a preset time window; S2: X Input the hidden states of a pre-trained bidirectional gated recurrent unit (Bi-GRU) neural network. The calculation is as follows: S3: Map the final hidden state to a risk index using a fully connected layer and a Sigmoid activation function. in, For the Sigmoid function, and For weights and biases, This indicates the predicted probability of cardiogenic shock occurring within the next 30 minutes. S4: Warning Triggering Logic Basis With threshold set The comparison results are executed, where .

[0011] Furthermore, the cardiovascular system mechanism model in the personalized physiological simulator module is a four-element lumped parameter model, and its core dynamics are described by the following system of differential-algebraic equations: in, , , These are the pressures in the left ventricle, aorta, and right atrium, respectively. For time-varying left ventricular compliance, satisfy ; , These are the compliance of the aorta and right atrium, respectively. , These are valve outlet resistance and peripheral systemic resistance, respectively. This represents venous return volume. The dynamic calibration is achieved using an unscented Kalman filter (UKF), based on the data monitored at the front end. and central venous pressure As an observation, the time-varying parameters in the model and It is updated online as the primary state estimate.

[0012] Furthermore, the treatment intervention virtualization and strategy optimization module performs the following process: S1: Based on the warning category, map from the knowledge base One candidate intervention Each measure Quantified as instantaneous or continuous perturbations to the parameters of the mechanistic model. ; S2: Using the current UKF-calibrated model state as the initial condition, for each Update the parameters to And numerically integrate the system of differential equations to a new quasi-steady state. S3: Calculate each virtual intervention Effect rating Using a linearly weighted utility function: in, , , These represent simulated cardiac output, mean arterial pressure, and pulmonary artery wedge pressure from baseline to quasi-steady state. The change in; The preset positive weighting coefficients, and satisfy the following conditions: S4: Press Sort all values ​​in descending order Output ranking Intervention recommendations for the location and its complete simulated hemodynamic trajectory .

[0013] Furthermore, the iterative optimization of the system's adaptive evolution module is achieved through the following two-layer mechanism: Level 1 (Short-term patient-specific adaptation): During the treatment of a single patient, when an intervention... After execution, record the actual changes in hemodynamic response: Calculate the prediction error vector: use The process noise covariance matrix of the UKF filter in the individualized physiological simulator module is fine-tuned online using a recursive least squares method. and observation noise covariance matrix To improve the accuracy of subsequent predictions for this patient; The second layer (long-term group knowledge evolution): periodically collecting effective "intervention-response" data from all patients. and the corresponding prediction error Using this dataset, the weight parameters of the Bi-GRU neural network were adjusted using gradient descent. Fine-tuning was performed, and the weighting coefficients of the utility function were adjusted. Perform Bayesian optimization to maximize long-term clinical outcome benefits.

[0014] The beneficial effects of this invention are: Traditional clinical assessments of myocardial contractility and volume responsiveness rely on invasive examinations or interventional tests, which cannot provide continuous data. This invention fundamentally changes this situation through a cascaded physiological signal synchronous acquisition and quality enhancement module and a core physiological state vector calculation module. First, the system outputs highly reliable multi-channel time-series data through hardware synchronization and adaptive processing based on quantitative quality scoring, laying the foundation for accurate calculation. Subsequently, the core calculation module performs inverse solving based on the first principles of the circulatory system: for myocardial contractility, the time-varying elastic model P(t) = E(t)·[V(t) - V0] is used. The E(t) curve is derived by nonlinear least squares fitting using the arterial blood pressure waveform P(t) and the estimated instantaneous ventricular volume V(t), and E(t) is used as the vector. max =max(E(t)) is used as a direct and continuous quantitative indicator; for capacity responsiveness, the periodic preload change of mechanical ventilation is used as the system identification input, and the dynamic gain coefficient between stroke volume variability and preload indicator is estimated online through a recursive least squares filter. This directly characterizes the slope of the Frank-Starling curve at the current operating point. This allows clinical monitoring to move beyond reliance on superficial parameters such as blood pressure and heart rate, and instead provide a continuous, non-invasive understanding of myocardial mechanical properties and the sensitivity of the preload-stroke volume relationship.

[0015] This invention constructs a forward-looking dynamic risk assessment system based on multi-dimensional temporal feature fusion and machine learning, realizing a paradigm shift in early warning mechanisms. Existing threshold alarm methods provide delayed responses to single parameter exceedances, making it difficult to identify complex pre-collapse states. The dynamic risk assessment and early warning module of this invention constructs a system that includes... , Basic vital signs and their rate of change , The temporal feature vector X t The study utilizes a bidirectional gated recurrent unit neural network for deep analysis. The Bi-GRU network can simultaneously learn historical and future contextual information, thereby extracting hidden patterns representing risk evolution and ultimately outputting a composite risk index R(t) representing the probability of hemodynamic decompensation events occurring within a specific future timeframe. Multi-level early warnings are triggered based on a progressive risk threshold set Θ, essentially representing risk stratification management based on probability and trends. This allows for earlier clinical intervention, moving from addressing existing crises to managing and blocking emerging risk pathways.

[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. Example 1

[0018] This embodiment describes an integrated system for real-time monitoring and early warning of hemodynamics in critically ill ICU patients. It includes an analog input interface, a digital serial port, and an Ethernet interface for connecting existing bedside invasive and non-invasive monitoring devices. Invasive monitoring devices include, but are not limited to, arterial pressure sensors, central venous pressure catheters, and pulmonary artery catheters. Non-invasive monitoring devices include, but are not limited to, finger-type photoplethysmography (PPG) devices and chest impedance flowmeters. The platform integrates a high-precision synchronous clock source to perform hardware-level time alignment of the raw physiological waveform signals from all the aforementioned devices. This ensures millisecond-level synchronization accuracy in timestamps for arterial blood pressure waveforms, central venous pressure waveforms, electrocardiogram waveforms, and PPG waveforms, providing a rigorous temporal foundation for subsequent multimodal data fusion analysis.

[0019] After completing basic processing such as anti-aliasing filtering and power frequency interference suppression, the core function of the physiological signal synchronous acquisition and quality enhancement module is to perform real-time signal quality assessment and enhancement. The built-in signal quality assessment unit jointly evaluates input waveform segments based on a pre-set physiological rationality rule base and a trained waveform classification machine learning model. The rule base includes multiple criteria such as signal amplitude range checking, motion artifact identification, and signal-to-noise ratio calculation. The machine learning model is a convolutional neural network structure specifically designed for binary classification of waveform segment reliability. The two work together to output a quantified quality score for each waveform segment. Only when the score exceeds a preset reliability threshold is the waveform segment marked as valid and sent to the downstream processing pipeline. For segments with scores below the threshold, the module immediately initiates an adaptive repair algorithm. This algorithm extracts the typical cardiac cycle morphology from the patient's recent high-quality waveform as a template, performs template-based interpolation and reconstruction on low-quality segments, or intelligently switches to a backup signal channel when multiple signals are available to maintain data stream continuity, thereby ensuring that the output to subsequent modules is always a set of highly reliable synchronous time-series data.

[0020] In the core physiological state vector solution module, the first parameter, the maximum contractile elasticity of the left ventricle, E maxThe solution is based on the theory of time-varying elasticity of the heart. The module continuously receives the arterial blood pressure waveform P(t) from upstream and the stroke volume SV estimated by pulse contour analysis. For each identified valid cardiac cycle, the module first accurately locates the start point t0 and the end point t1 of the cycle. Subsequently, the module uses a linear decay model to estimate the instantaneous volume change V(t) of the left ventricle during the cycle, and the calculation formula is as follows: , t0≤t≤t1 Among them, V ed The estimated left ventricular end-diastolic volume is initially calculated using an empirical formula based on the patient's height, weight, and gender, and can be subsequently calibrated using echocardiographic measurements. After obtaining the instantaneous arterial blood pressure P(t) and the estimated instantaneous ventricular volume V(t), the module constructs and solves the time-varying elastic equation P(t) = E(t)·[V(t) - V0]. Here, E(t) is the time-varying elastic function of the left ventricle, and V0 is the theoretical ventricular volume zero, an individualized constant related to the myocardial resting state. The solution process employs a nonlinear least squares fitting algorithm, iteratively calculating the time-varying elastic function curve E(t) that best fits the pressure-volume data of the current cycle. Finally, the maximum value extracted from this curve is defined as the E(t) for that cardiac cycle. max The parameters, as core indicators characterizing myocardial contractility, are stored in the state vector, namely: E max =max(E(t)); For the second parameter, capacity reactivity gain coefficient The online estimation is based on the principle of periodic preload changes caused by mechanical ventilation. The module first extracts the stroke volume sequence SV(j) for N consecutive (N≥3) cardiac cycles within a mechanical ventilation cycle and calculates the stroke volume variability SVV(k): Where k represents the k-th respiratory cycle. To ensure the physiological validity of the data, the system performs strict validity checks, including the following conditions: the patient's heart rhythm must be sinus rhythm, and the tidal volume Vt of the ventilator must be [value missing]. T satisfy (Based on ideal body weight), the arterial waveform profile must remain stable. Only when all conditions are met is the calculated SVV(k) marked as a valid measurement. The data is then fed into the parameter update stage. A recursive least squares filter is used as the core estimation algorithm, and the update formula is: in, The volume responsiveness gain coefficient estimated for the current k-th respiratory cycle; This is an estimate from the previous period; K(k) is the regression vector, containing preload-related monitoring indicators, such as respiratory variability in central venous pressure; K(k) is the Kalman gain matrix, recursively calculated and updated according to the algorithm. The filter dynamically updates K(k) based on the error between the latest valid measurement and the current state estimate, and recursively calculates new gain coefficient estimates. This coefficient directly quantifies the slope of the rising segment of the patient's current cardiac function curve (Frank-Starling curve), and its value becomes a direct basis for judging whether the patient is responding to volume overload.

[0021] The dynamic risk assessment and early warning module is implemented using a pre-trained deep time-series model. It was trained using a dataset of critically ill patients containing a large number of labeled events before deployment. The module constructs a multi-dimensional feature vector X(t) in real time, with the following expression: Where HR(t) is the real-time heart rate, and MAP(t) is the real-time mean arterial pressure. and Parameter E max and The rate of change within a preset time window (e.g., 120 seconds). This feature vector is input into a bidirectional gated recurrent neural network. Through its forward and backward recurrent structures, it simultaneously captures the evolution patterns of the feature in past and future contexts, thereby extracting a hidden state representation h(t) that can predict the risk of decompensation: The hidden states are processed by a fully connected layer and a sigmoid activation function, and finally mapped to a composite risk index R(t) between 0 and 1, representing the probability predicted by the system that the patient will experience a hemodynamic decompensation event within a specific time period in the future: in, For the Sigmoid function, and This represents the weight matrix and bias vector of the network output layer. The system will use the real-time calculated risk index R(t) and a preset set of multi-level risk thresholds. Comparison, among which Based on the comparison results, different levels of alerts are triggered. The alert information is released in real time through the human-machine interface in visual and auditory form, and the complete context data of the alert event is automatically archived into the medical record system.

[0022] The core of the personalized physiological simulator module is a lumped-parameter mechanism model of the cardiovascular system with adjustable parameters. This embodiment uses a four-element model, describing the system dynamics through the following system of differential-algebraic equations: Among them, P lv P ao P ra These represent left ventricular pressure, aortic pressure, and right atrial pressure, respectively; C lv (t) represents the time-varying left ventricular compliance, which satisfies the relationship C with the aforementioned time-varying elastic function E(t). lv (t) = 1 / E(t); C ao With C ra These are aortic compliance and right atrial compliance, respectively; R v With R sys These are valve outlet resistance and systemic peripheral resistance, respectively; Q in This represents venous return. In the model, the time-varying compliance parameter of the left ventricle is directly driven by the reciprocal of the time-varying elastic function obtained from the aforementioned core solution module, thus embedding the patient's actual cardiac contractile characteristics into the simulation model. To ensure that the general model remains consistent with the real-time state of a specific patient, the module employs an unscented Kalman filter for dynamic parameter calibration. During calibration, multiple state variables of the model and key parameters to be estimated are combined to form an extended state vector. The mean arterial pressure (MAP) and central venous pressure (CVP) obtained from real-time monitoring are used as observation input filters. The filter continuously compares the observed values ​​predicted by the model with the actual monitored values, and recursively updates the optimal estimate and its uncertainty covariance of the extended state vector accordingly. Through this process, key parameters in the model, such as systemic circulatory resistance R... sys Aortic compliance C ao The simulation output of the entire mechanism model can be adjusted online to closely track the patient's current actual hemodynamic state, forming a dynamically updated physiological image of the patient.

[0023] The treatment intervention virtualization and strategy optimization module operates based on the aforementioned calibrated individualized physiological simulator. Internally, the module maintains a treatment measure knowledge base, storing the quantitative mapping relationships between common hemodynamic interventions and model parameter perturbations. Upon receiving a trigger signal from the early warning module, the module extracts a set of candidate treatment measures from the knowledge base that match the current risk level. For each candidate measure The module will change the corresponding parameter perturbation amount The current state is applied to a personalized physiological simulator, and the complete trajectory of the cardiovascular system state over time after the virtual intervention is simulated using numerical integration methods until a new quasi-steady state is reached. Subsequently, the module quantifies the effect of each virtual intervention using a preset utility function, U=, whose expression is: in, , , These represent virtual intervention A. i Under these conditions, simulated cardiac output, mean arterial pressure, and pulmonary artery wedge pressure change from baseline to quasi-steady state. The change in; The preset positive weighting coefficients satisfy the normalization condition. The initial values ​​for all weighting coefficients are based on clinical consensus. Finally, the module assigns all candidate measures to their utility scores (U). i The top K intervention recommendations are sorted in descending order, and the simulated hemodynamic prediction trajectory for each recommendation is also provided. This information is provided for clinical decision-making reference.

[0024] The system's adaptive evolution module implements a two-layer optimization architecture to achieve system self-iteration. The first layer of optimization performs short-term adaptive learning for a single patient. When a treatment measure A recommended by the system... j After actual execution, the module records the actual changes in the patient's hemodynamic response before and after the execution. By comparing the actual response with the simulated predicted response. Calculate the prediction error vector The error is input into a recursive least squares algorithm to fine-tune the process noise covariance matrix Q and observation noise covariance matrix R of the unscented Kalman filter used in the patient-individualized physiological simulator online. This fine-tuning allows the filter to better adapt to the physiological noise characteristics of the patient, thereby improving accuracy in subsequent predictions. The second layer of optimization performs long-term knowledge evolution based on all patient data. The system periodically collects effective "intervention-response" data generated by all patients during treatment. and the corresponding prediction error This results in a continuously growing clinical dataset. Using this dataset, the weight parameters of the bidirectional gated recurrent neural network used in the dynamic risk assessment module are adjusted using gradient descent. Fine-tuning was performed to improve its risk identification capabilities. Simultaneously, a Bayesian optimization algorithm was used to adjust the weight coefficients of the utility function in the treatment strategy optimization module. Global adjustments are made to optimize the system, aiming to maximize the correlation between the system's recommended treatment strategies and improved long-term clinical outcomes for patients. Through a two-layer closed-loop mechanism, the system can continuously evolve with ongoing use, improving its overall performance and clinical applicability.

[0025] In summary, this invention proposes an integrated system for real-time hemodynamic monitoring and early warning in critically ill ICU patients, comprising: a physiological signal synchronous acquisition and quality enhancement module for acquiring and processing multi-channel physiological waveforms; a core physiological state vector calculation module, based on a first-principles model of the circulatory system, for real-time calculation of the left ventricular maximum contractile elasticity Emax, representing myocardial contractility, and the gain coefficient γk, representing preload responsiveness; a dynamic risk assessment and early warning module, which predicts decompensation risk and triggers early warning by fusing temporal features through a bidirectional gated circulatory unit neural network; a personalized physiological simulator module, which dynamically calibrates the cardiovascular mechanism model using an unscented Kalman filter to generate a physiological mirror image of the patient; a treatment intervention virtualization and strategy optimization module, which simulates the effects of candidate treatment strategies on the mirror image and provides optimization suggestions; and a system adaptive evolution module, which iteratively optimizes model parameters through a two-layer closed-loop mechanism. This invention achieves closed-loop management from accurate perception and risk prediction to personalized decision support.

[0026] 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. A smart hemodynamic monitoring and decision support system for critically ill patients, characterized in that, include: The physiological signal synchronous acquisition and quality enhancement module is used to acquire and time-calibrate multi-channel physiological waveforms from invasive and non-invasive monitoring devices, and perform adaptive processing on the waveforms based on quantitative quality scoring to output reliable time-series data. The core physiological state vector calculation module is connected to the physiological signal synchronous acquisition and quality enhancement module. It is used to receive the reliable time-series data and, based on the preset first-principles model of the circulatory system, calculate in real time the internal state vector, which includes at least a first parameter representing the state of myocardial contractility and a second parameter representing the state of preload responsiveness. The dynamic risk assessment and early warning module receives the core physiological state vector and basic vital signs, integrates multi-dimensional temporal features through a machine learning model, calculates the composite risk index of hemodynamic decompensation events occurring within a specific time window in the future, and triggers multi-level warnings based on a preset progressive risk threshold matrix. The personalized physiological simulator module, including a cardiovascular system mechanism model with online updatable parameters, is configured to dynamically calibrate model parameters using real-time monitoring data to generate a dynamic physiological mirror synchronized with the current patient status. The treatment intervention virtualization and strategy optimization module responds to the output of the dynamic risk assessment and early warning module and calls the individualized physiological simulator module to generate intervention recommendations based on expected effect ranking by simulating the execution of a set of candidate treatment strategies and quantifying and predicting their hemodynamic response. In addition, the system adaptive evolution module is configured to collect actual clinical intervention and physiological response data, compare them with the prediction results of the treatment intervention virtualization and strategy optimization module, and use the comparison results to drive the parameters of the machine learning model and the cardiovascular system mechanism model to perform iterative optimization.

2. The system as described in claim 1, characterized in that: In the core physiological state vector solution module, the first parameter is the maximum contractile elasticity of the left ventricle. The following steps will solve the problem: S1: Identify the start point t0 and end point t1 of the cardiac cycle from the arterial blood pressure waveform; S2: Obtain the stroke volume SV of the cardiac cycle based on pulse contour analysis and estimate the instantaneous ventricular volume V(t), using the following formula: where V ed to estimate end diastolic volume, t0≤t≤t1; S3: Construct a time-varying elastic model P(t) = E(t)·[V(t)-V0], where P(t) is arterial blood pressure, E(t) is the time-varying elastic function to be determined, and V0 is the theoretical ventricular volume zero point, which is a constant related to the resting state of the myocardium; S4: Substitute the measured V(t) estimated in step S2 into the model to solve E(t) using a non-linear least squares method, and define the maximum value of E(t) in the cardiac cycle as E max , i.e. E max = max(E(t)).

3. The system as described in claim 1, characterized in that, In the core physiological state vector calculation module, the second parameter, namely the capacity responsiveness gain coefficient... The following online estimation algorithm was used to obtain the result: S1: During the mechanical ventilation cycle, extract continuous... indivual( Stroke volume sequence of the cardiac cycle And calculate the variability of stroke volume: in Indicates the first One respiratory cycle; S2: Set validity criteria: The patient's heart rhythm is sinus rhythm and the tidal volume is... (Ideal weight); current weight is only considered if the criterion is met. Only then is it marked as a valid measurement value. S3: Will The input recursive least squares (RLS) filter is used for online estimation, and the update formula is as follows: in, This is the estimated value of the gain coefficient at the current moment. The regression vector includes preload-related indicators, such as pulse pressure. Here is the Kalman gain matrix; coefficients It directly reflects the slope of the rising segment of the patient's current cardiac function curve.

4. The system as described in claim 1, characterized in that: In the dynamic risk assessment and early warning module, the composite risk index The calculation method is as follows: S1: Constructing multi-dimensional temporal feature vectors in, Heart rate; Mean arterial pressure; , These correspond to the rate of change of the parameters within a preset time window; S2: X Input to a pre-trained bidirectional gated recurrent unit (Bi-GRU) neural network, hidden state The calculation is as follows: S3: Map the final hidden state to a risk index using a fully connected layer and a Sigmoid activation function. in, For the Sigmoid function, and For weights and biases, This indicates the predicted probability of cardiogenic shock occurring within the next 30 minutes. S4: Warning Triggering Logic Basis With threshold set The comparison results are executed, where .

5. The system as described in claim 1, characterized in that, The cardiovascular system mechanism model in the individualized physiological simulator module is a four-element lumped parameter model, and its core dynamics are described by the following system of differential algebraic equations: in, , , These are the pressures in the left ventricle, aorta, and right atrium, respectively. For time-varying left ventricular compliance, satisfy ; , These are the compliance of the aorta and right atrium, respectively. , These are valve outlet resistance and peripheral systemic resistance, respectively. This refers to venous return volume; the dynamic calibration is achieved through an unscented Kalman filter (UKF), based on the front-end monitoring data. and central venous pressure As an observation, the time-varying parameters in the model and It is updated online as the primary state estimate.

6. The system as described in claim 5, characterized in that, The treatment intervention virtualization and strategy optimization module executes the following process: S1: Based on the warning category, map from the knowledge base One candidate intervention Each measure Quantified as instantaneous or continuous perturbations to the parameters of the mechanistic model. ; S2: Using the current UKF-calibrated model state as the initial condition, for each Update the parameters to And numerically integrate the system of differential equations to a new quasi-steady state. S3: Calculate each virtual intervention Effect rating Using a linearly weighted utility function: in, , , These represent simulated cardiac output, mean arterial pressure, and pulmonary artery wedge pressure from baseline to quasi-steady state. The change in; The preset positive weighting coefficients, and satisfy the following conditions: S4: Press Sort all values ​​in descending order Output ranking Intervention recommendations for the location and its complete simulated hemodynamic trajectory .

7. The system as described in claim 1, characterized in that, The iterative optimization of the system's adaptive evolution module is achieved through the following two-layer mechanism: First level: In the treatment of a single patient, when an intervention... After execution, record the actual changes in hemodynamic response: Calculate the prediction error vector: use The process noise covariance matrix of the UKF filter in the individualized physiological simulator module is fine-tuned online using a recursive least squares method. and observation noise covariance matrix To improve the accuracy of subsequent predictions for patients; The second layer: periodically collecting effective "intervention-response" data from all patients. and the corresponding prediction error Using this dataset, the weight parameters of the Bi-GRU neural network were adjusted using gradient descent. Fine-tuning was performed, and the weighting coefficients of the utility function were adjusted. Perform Bayesian optimization to maximize long-term clinical outcome benefits.

Images