Noninvasive hemodynamic detection and decision-making assistance method and device

Abstract

The invention discloses a non-invasive hemodynamic detection and decision-making assistance method and device. The method comprises the steps of obtaining a first data set; preprocessing the first data set to obtain a second data set; an intracardiac prediction model based on a Transform architecture is constructed and trained; inputting the second data set to the trained intracardiac prediction model, and outputting the second data set as a third data set; collecting expert experience knowledge; constructing a relational expert knowledge base by adopting a knowledge representation method of a semantic network; performing trend analysis and correlation analysis on the third data set, the real-time electrocardiogram monitor physiological parameters and the electronic medical record to obtain an analysis result; performing matching reasoning on the analysis result and a treatment scheme of a corresponding case in an expert knowledge base to obtain a reasoning result; and generating personalized decision suggestions based on the reasoning result. According to the method, the treatment scheme decision is generated only based on the electrocardiograph monitor, and an efficient, intelligent and non-invasive diagnosis and treatment assistant is provided for clinical medical staff.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical auxiliary diagnosis, artificial intelligence technology and machine learning, in particular to a non-invasive hemodynamic detection and decision assistance method and device, which is suitable for heart function monitoring and diagnosis and treatment assistance of critical patients. BACKGROUND

[0002] In the diagnosis and treatment of critical patients, vital signs and heart function are important indicators of the cardiovascular system. Reliable measurement methods and recognition of dynamic changes play an important role in disease evolution prevention and symptomatic treatment. Studies suggest that before the clinical symptoms of critical patients appear, they will go through a hemodynamic dysfunction stage. Hemodynamic parameters such as ventricular filling pressure, cardiac output, circulation perfusion, and intrathoracic impedance will be detected before the symptoms appear. In this stage, early adjustment and intervention can effectively reduce cardiovascular mortality and re-hospitalization rate. When there is a lack of advanced hemodynamic monitoring, clinicians' ability to predict important physiological indicators from clinical symptoms is very limited, which may significantly underestimate cardiac output and more physiological indicators, and have a major impact on major hemodynamic decisions. Although hemodynamic monitoring can effectively guide disease monitoring and treatment and improve the clinical outcome of critical patients, the use of invasive hemodynamic monitoring methods currently limits its use.

[0003] The 2022 AHA / ACC / HFSA Heart Failure Management Guidelines also recommend against routine use of invasive hemodynamic monitoring methods for heart failure patients. For non-invasive hemodynamic monitoring methods, especially the introduction of wearable devices and other non-invasive hemodynamic monitoring, the guidelines introduce them, but also point out that there is no uniform standard at present, and more data and further research are needed. Because of its good application prospect, current domestic and foreign research has actively explored it.

[0004] Artificial intelligence machine learning usually uses a large amount of data to build models through algorithms and can identify non-linear relationships and high-order interactions between multiple variables. With the intelligentization of monitoring equipment in the ICU and the informatization of data collection, medical staff are no longer limited to evidence-based medicine and clinical experience to make major clinical medical decisions for critically ill patients. Based on the application of artificial intelligence technology, medical staff can more timely identify clinical problems, better develop clinical decisions, and more effectively improve the clinical service capabilities for critically ill patients.

[0005] In recent years, artificial intelligence has received great attention in medicine, especially in medical image recognition, disease diagnosis, tumor recognition and typing, drug application, etc. SUMMARY

[0006] The application provides a non-invasive hemodynamic detection and decision assistance method and device, which can solve the technical problems of high risk of invasive monitoring, insufficient accuracy of non-invasive monitoring and dependence on professional operation in the prior art, and can provide effective treatment scheme decision.

[0007] In a first aspect, a non-invasive hemodynamic detection and decision assistance method is provided, comprising: obtaining a first data set by collecting real-time electrocardiogram monitor physiological parameters of a heart critical monitoring object; preprocessing the first data set to obtain a second data set; constructing and training an intracardiac prediction model based on a Transformer architecture; inputting the second data set into the trained intracardiac prediction model to output a predicted hemodynamic parameter prediction value as a third data set; collecting expert experience knowledge and organizing it into a semantic network structure knowledge; using a semantic network knowledge representation method, the obtained knowledge is represented in the form of nodes and edges to form the association relationship between the physical parameters and diseases, and then a relational expert knowledge base is constructed; performing trend analysis and correlation analysis on the third data set, real-time electrocardiogram monitor physiological parameters of the heart critical monitoring object and electronic medical records to obtain an analysis result; matching and reasoning the analysis result with the treatment scheme of the corresponding case in the expert knowledge base to obtain a reasoning result; based on the reasoning result, generating a personalized decision suggestion combined with the individual situation of the patient.

[0008] In a possible implementation manner, the electrocardiogram monitor physiological parameters include heart rate HR, blood oxygen saturation SpO2, non-invasive systolic blood pressure NIBP-S, non-invasive diastolic blood pressure NIBP-D, respiratory rate RR, pulse rate PR and central venous pressure CVP.

[0009] In a possible implementation manner, the intracardiac prediction model based on the Transformer architecture comprises: an input embedding layer composed of a full connection layer, a GELU activation function and layer normalization, used for mapping static parameters to a multi-dimensional feature space to obtain a multi-dimensional feature vector; a Transformer encoder containing multiple encoder layers, each layer being provided with a multi-head self-attention mechanism, and a feedforward network dimension being used for global relationship modeling of the multi-dimensional feature vector to obtain standardized high-order features; and The regression output layer consists of a fully connected layer, GELU activation, Dropout, layer normalization, and a final fully connected layer. It is used to decode high-order feature representations as predicted values ​​of hemodynamic parameters through fully connected layers, GELU activation, and regularization operations.

[0010] In one possible implementation, the third dataset includes: cardiac output (CO), cardiac index (CI), stroke volume (SV), stroke index (SVI), peripheral vascular resistance (SVR), and peripheral vascular resistance index (SVRI).

[0011] In one possible implementation, the knowledge representation method using semantic networks represents the acquired knowledge in the form of nodes and edges, forming a correlation between vital signs parameters and diseases, including: Using the knowledge representation method of semantic networks, with vital signs parameters, disease names, pathophysiological conditions, and treatment plans as nodes, the relationship between each node is established through edges, thereby forming the association between vital signs parameters and diseases.

[0012] In one possible implementation, the step of matching and reasoning the analysis results with the treatment plans of corresponding cases in the expert knowledge base to obtain the reasoning results includes: The analysis results are compared with the treatment plans of corresponding cases in the expert knowledge base to obtain the first reasoning result; the rule reasoning is to classify the disease based on the vital signs parameters according to the preset rules. The analysis results are compared with the treatment plans of corresponding cases in the expert knowledge base to obtain a second reasoning result; the second reasoning result is obtained by performing a similarity search based on the input parameters to find the closest corresponding case; The first and second inference results are combined to obtain the final inference result.

[0013] In one possible implementation, generating personalized decision recommendations based on the reasoning results and the patient's individual circumstances includes: Based on the reasoning results, clinical decisions are made; Risk stratification based on clinical decision-making; Based on risk classification, a basic solution is obtained; Based on the basic plan, combined with Monte Carlo simulation and the patient's individual circumstances, personalized decision-making suggestions are generated.

[0014] Secondly, it provides non-invasive hemodynamic monitoring and decision support devices, including: The acquisition module is used to acquire a first dataset, which is obtained by collecting real-time electrocardiogram physiological parameters of the cardiac critical care monitoring subjects. The preprocessing module is used to preprocess the first dataset to obtain the second dataset; The first building block is used to build and train a cardiac prediction model based on the Transformer architecture; The output module is used to input the second dataset into the trained intracardiac prediction model and output the predicted hemodynamic parameters as the third dataset. The second construction module is used to collect expert experience and knowledge, organize it into a semantic network structure, and use the knowledge representation method of semantic network to represent the acquired knowledge in the form of nodes and edges, forming the association between vital signs and diseases, and then constructing a relational expert knowledge base. The analysis module is used to perform trend analysis and correlation analysis on the third dataset, the real-time electrocardiogram monitor physiological parameters of the cardiac critical care subjects, and the electronic medical records to obtain the analysis results; The reasoning module is used to match and reason with the treatment plans of corresponding cases in the expert knowledge base to obtain the reasoning results; The generation module is used to generate personalized decision-making suggestions based on the reasoning results and the patient's individual circumstances.

[0015] Thirdly, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the non-invasive hemodynamic detection and decision support method as described in the first aspect.

[0016] Fourthly, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the non-invasive hemodynamic detection and decision support method as described in the first aspect.

[0017] The beneficial effects of this application are as follows: This application utilizes the Transformer architecture to achieve multi-parameter global modeling, enabling high-precision estimation of key intracardiac indicators such as cardiac output and cardiac index. Compared with traditional invasive monitoring methods, this application can complete real-time cardiac function assessment based on non-invasive static indicators, significantly reducing patient risk and clinical costs. At the same time, the Transformer model has strong modeling capabilities and scalability, making it suitable for various medical terminal deployment scenarios.

[0018] Based on non-invasive hemodynamic testing, this application further provides treatment plan decisions that are highly matched to the patient's condition and take into account the individual's specific circumstances, thus providing effective technical support to assist doctors in treatment. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application.

[0020] Figure 1 A flowchart illustrating the method provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the intracardiac prediction model according to an embodiment of this application; Figure 3 The R^2 graph shows the accuracy measurement results of the prediction model in this application embodiment. Figure 4 This is the training and validation loss curve for the prediction of non-invasive hemodynamic parameters in this application embodiment; Figure 5 This is a schematic diagram of the prediction result visualization output interface in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of the non-invasive hemodynamic detection and decision support device provided in the embodiments of this application; Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] The content of this application will be further described below with reference to specific embodiments, but the content of this application is not limited thereto.

[0022] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0023] In the description of the embodiments in this application, terms such as "exemplary," "for example," or "for instance" are used as examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.

[0024] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0025] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0026] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0027] HR, Heart Rate; Peripheral Oxygen Saturation; NIBP, Non-Invasive Blood Pressure; NIBP-S: Systolic blood pressure; NIBP-M: Average Pressure; RR, Respiratory Rate; PR, Pulse Rate; CVP, Central Venous Pressure; CO, Cardiac Output; CI, Cardiac Index; SV, Stroke Volume; SVI, Stroke Volume Index; SVR, Systemic Vascular Resistance; SVRI, Systemic Vascular Resistance Index Transformer, a self-attention model; RNN, Recurrent Neural Network; CNN stands for Convolutional Neural Network. GELU, Gaussian Error Linear Unit; Dropout, randomly discarding items; IQR, Interquartile Range; MSE, Mean Squared Error; MAE, Mean Absolute Error; TEB, Thoracic Electrical Bioimpedance, is a method of thoracic impedance measurement. LiDCO, Lithium Dilution Cardiac Output, is a monitoring tool for monitoring the output of lithium dilution processes. epoch, training round; NaN, Not a Number.

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0029] This application provides a non-invasive hemodynamic detection and decision support method and device, which can solve the technical problems of high risk of invasive monitoring, insufficient accuracy of non-invasive monitoring and reliance on professional operation, and inability to provide treatment plan decision-making in the prior art.

[0030] See Figure 1 The non-invasive hemodynamic detection and decision support method provided in this application includes: S101. Obtain the first dataset, which is obtained by collecting real-time electrocardiogram (ECG) physiological parameters of the cardiac critical care monitoring subjects.

[0031] In one possible implementation, the physiological parameters of the electrocardiogram monitor include: heart rate (HR), oxygen saturation (SpO2), non-invasive systolic blood pressure (NIBP-S), non-invasive diastolic blood pressure (NIBP-D), respiratory rate (RR), pulse rate (PR), and central venous pressure (CVP).

[0032] Specifically, time-series data of heart rate, blood oxygen saturation, non-invasive blood pressure, respiratory rate, pulse rate and central venous pressure are acquired synchronously through an electrocardiogram monitor, and parameters such as cardiac output and systemic vascular resistance at corresponding moments are collected as supervision labels through minimally invasive hemodynamic monitoring equipment (such as a pulse profile analyzer) to construct a static physiological parameter dataset.

[0033] This embodiment provides a data foundation for non-invasive monitoring technology by simultaneously collecting routine monitoring indicators and professional hemodynamic parameters.

[0034] S102. Preprocess the first dataset to obtain the second dataset.

[0035] In one possible implementation, S102 includes: S102a. Perform outlier detection and filtering on the collected physiological parameters, remove negative values ​​and data that exceed the clinically reasonable range, and obtain the filtered parameter data. S102b. Standardize the filtered parameter data to obtain standardized data; specifically, robust standardization is used to standardize the filtered parameter data, using the following formula: ;in, Standardized variables , For variables the median of For variables The four-position spacing; S102c. Perform a logarithmic transformation on the vascular resistance parameters in the standardized data as shown in the following formula to obtain the second dataset:

[0036]

[0037] In the formula, For systemic vascular resistance, The systemic vascular resistance after logarithmic transformation. The systemic vascular resistance index, This is the systemic vascular resistance index after logarithmic transformation.

[0038] The data preprocessing provided in this embodiment first employs clinically reasonable range filtering to ensure data reliability, then robust standardization eliminates dimensional differences, and specifically performs logarithmic transformation on vascular resistance parameters to improve distribution characteristics, constructing a multidimensional feature representation that both preserves the intrinsic correlation of physiological characteristics and meets the model input requirements. S102, through multi-step joint optimization, significantly improves data quality and model adaptability, laying a data foundation for subsequent accurate predictions.

[0039] S103. Construct and train a cardiac prediction model based on the Transformer architecture.

[0040] In one possible implementation, the intracardiac prediction model includes: Input embedding layer, Transformer encoder, and regression output layer; The input embedding layer, consisting of a fully connected layer, a GELU activation function, and a layer normalization, is used to map static parameters to a 256-dimensional feature space to obtain multi-dimensional feature vectors. The Transformer encoder consists of four encoder layers, each with a 16-head self-attention mechanism. The feedforward network dimension is used to model the global relationships between multi-dimensional feature vectors and obtain standardized high-order features. The core calculation formula of the multi-head self-attention mechanism is... ,in, For query matrix Key matrix Sum matrix Multi-head attention; Concat is a concatenation operation. To output the projection matrix; No. The formula for calculating single-head attention is:

[0041] In the formula, For the first Each query weight matrix For the first Each key weight matrix; For the first Individual value weight matrix; The regression output layer consists of a fully connected layer, GELU activation, Dropout, layer normalization, and a final fully connected layer. It is used to decode high-order feature representations as hemodynamic parameters through fully connected layers, GELU activation, and regularization operations.

[0042] The Transformer architecture provided in this embodiment effectively extracts basic features through the input embedding layer, captures complex relationships between parameters using a multi-layer encoder with a self-attention mechanism, and decodes high-order features into clinically usable prediction results through nonlinear transformation. Through multi-level feature transformation and fusion, a high-precision mapping from static physiological parameters to hemodynamic indicators is achieved. The overall architecture significantly improves the prediction accuracy of key physiological indicators while ensuring computational efficiency.

[0043] In one embodiment, all four layers of the Transformer encoder include the following components: The self-attention sublayer takes multi-dimensional feature vectors as input, calculates the global dependencies between these feature vectors, and outputs higher-order features. The calculation formula for the self-attention sublayer is as follows:

[0044] In the formula, The attention mechanism for query matrix Q, key matrix K, and sum matrix V, Let be the dimension of the key vector, and softmax be the normalization exponential function. For query-key matrix multiplication; The feedforward sublayer, consisting of two fully connected layers and a GELU activation function, is used to mine deep patterns in high-order features through dimensional expansion and nonlinear transformation, and output deep features. The residual connection and layer normalization component is used to perform element-wise addition between the output of the self-attention sublayer and the input features, and a second residual connection between the output of the feedforward sublayer and the intermediate features. After double-layer normalization, the standardized high-order features are output, as shown in the formula:

[0045] In the formula, For input The final output result, For input The sub-layer, LayerNorm, is the layer normalization. This is the discard method.

[0046] In one embodiment, the calculation formula for the self-attention mechanism of the self-attention sublayer is as follows:

[0047] In the formula: Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. is the dimension of the key vector.

[0048] In one possible implementation, the training method of the intracardiac prediction model includes: training the constructed intracardiac prediction model using a dynamic weighted loss function and an adaptive learning rate scheduling strategy to obtain the trained intracardiac prediction model, specifically including the following steps: S1031. Based on the pairing relationship between static physiological parameters and hemodynamic parameter labels, the static physiological parameter dataset is divided into training set, validation set, and test set; specifically, a dataset containing static parameters and hemodynamic parameter labels is constructed and divided into training set, validation set, and test set in an 8:1:1 ratio. S1032. Based on the training set data, the model parameters are optimized using a dynamic weighted relative error loss function, where the weight allocation corresponds to different output parameters, to obtain a preliminary optimized model; S1033. Based on the change in validation set loss, implement a dynamic learning rate decay strategy for the initially optimized model. When the validation loss does not decrease continuously, decay the initial learning rate. At the same time, obtain the adjusted model through a pruning operation constrained by the global gradient norm. S1034. Based on the validation set monitoring results, trigger the early stopping mechanism for the adjusted model. When the validation loss does not improve continuously, terminate the training process and obtain the optimal model at the time of training termination. S1035. Input test set data into the optimal model to evaluate its prediction performance and obtain a trained cardiac prediction model that meets the prediction performance standards.

[0049] This embodiment employs a scientific data partitioning strategy at the data organization level to ensure the sufficiency of training data and the independence of evaluation, laying a reliable foundation for the model. In terms of optimization strategy, an innovative differentiated weighting mechanism is introduced, prioritizing predictive indicators according to clinical importance, significantly improving the prediction accuracy of key parameters. During training, dynamic learning rate adjustment and gradient pruning provide dual protection, maintaining the stability of parameter updates while avoiding the risk of numerical oscillations. An intelligent early stopping mechanism monitors model performance in real time, automatically terminating training when performance is optimal, effectively preventing overfitting. The final testing and validation phase comprehensively evaluates the model's generalization ability, ensuring that the prediction results meet clinical application standards. The entire solution, through the organic coordination of data partitioning, optimization strategies, process control, and performance evaluation, enables the model to achieve optimal predictive performance while maintaining stability.

[0050] In one embodiment, the model training method involves dividing the static physiological parameter dataset into a training set, a validation set, and a test set based on the pairing relationship between static physiological parameters and hemodynamic parameter labels. In one embodiment, in S1032, the dynamically weighted relative error loss function for:

[0051] In the formula, The total number of samples, For the first The weighting coefficients for the hemodynamic parameters, w=[0.5,0.5,0.5,0.5,0.75,0.75], correspond to CO, CI, SV, SVI, SVR, and SVRI, respectively. For the first In the nth sample True labeled values ​​of hemodynamic parameters For the first In the nth sample Model-predicted values ​​of several hemodynamic parameters It is a very small constant.

[0052] In a more specific embodiment, the training method of the intracardiac prediction model includes: training the constructed intracardiac prediction model using a dynamic weighted loss function and an adaptive learning rate scheduling strategy to obtain the trained intracardiac prediction model, specifically including the following steps: Based on the pairing relationship between static physiological parameters (such as heart rate and blood oxygen) and hemodynamic parameter labels (such as cardiac output and vascular resistance), the static physiological parameter dataset is divided into training set, validation set and test set to ensure the balance and representativeness of data distribution. The model parameters are optimized using a weighted relative error loss function based on the training set data. The weights are dynamically adjusted according to the importance of the output parameters (e.g., the weight of SVR is higher than that of CO) to prioritize the prediction accuracy of key indicators and obtain a preliminary optimized model. The initial learning rate is set to 0.001. A dynamic decay strategy is implemented based on the change in the loss on the validation set. If the loss does not decrease for 5 consecutive epochs, the rate is decayed by 50%, gradually converging to a steady state. Gradient explosion is prevented by limiting the global gradient norm (with a threshold set to 1.0), thus improving training stability.

[0053] Training is terminated when the validation loss fails to improve for 15 consecutive training epochs to avoid overfitting, and the model with the lowest validation loss is saved as the optimal model.

[0054] Record validation set performance metrics (such as mean squared error and correlation coefficient) at the end of each training epoch to monitor the model's generalization ability. Input test set data into the optimal model to evaluate its predictive performance, ensuring that key indicators (such as the prediction error of CO ≤10%) meet clinical requirements, and finally obtain the trained intracardiac prediction model.

[0055] S104. Input the second dataset into the trained intracardiac prediction model, and output the predicted hemodynamic parameters as the third dataset.

[0056] In one possible implementation, S104 includes: inputting a second dataset into the trained intracardiac prediction model, obtaining predicted values ​​of hemodynamic parameters, wherein the predicted values ​​of hemodynamic parameters include 6 hemodynamic parameters (CO, CI, SV, SVI, SVR, SVRI), and the output 6 hemodynamic parameters need to be comprehensively judged by clinicians in conjunction with other examinations and cannot be directly used as a basis for diagnosis.

[0057] S105. Collect expert experience and knowledge, organize it into a semantic network structure; use the knowledge representation method of semantic network to represent the acquired knowledge in the form of nodes and edges, form the correlation between vital signs and diseases, and then construct a relational expert knowledge base.

[0058] In one possible implementation, S105 includes: Using a semantic network knowledge representation method, nodes such as vital signs parameters, multi-lead electrocardiograms, disease names, pathophysiological conditions, treatment plans, and electronic medical records are used to establish relationships between nodes through edges, thereby forming the association between vital signs parameters and diseases.

[0059] Furthermore, the relationships between vital signs and diseases are stored in a database, employing a combination of relational databases (such as MySQL) and graph databases (such as Neo4j). Relational databases are used to store structured knowledge, such as charts and tables; graph databases are used to store semantic network structures, facilitating knowledge retrieval and reasoning.

[0060] In one possible implementation, the expert knowledge base is dynamically updated through a federated learning framework, integrating clinical guidelines, medical literature, and expert experience to establish a model of the relationship between diseases and hemodynamic parameters: each hospital deploys a lightweight client locally, encrypting and uploading anonymous decision logs; a central server aggregates the data and generates a new version of the decision rule set monthly; the update process complies with HIPAA / GDPR privacy standards.

[0061] S106. Perform trend analysis and correlation analysis on the third dataset, the real-time electrocardiogram monitor physiological parameters of the cardiac critical care subjects, and the electronic medical records to obtain the analysis results.

[0062] In one possible implementation, trend analysis in S106 includes: The directionality, stability and periodicity of physiological parameters changing over time are quantified and analyzed to identify the evolution patterns of patients' vital signs and discover potential abnormalities or key clinical events. Moving statistical calculations are performed on continuous monitoring data through time windows (such as short-term 60 seconds, medium-term 5–15 minutes, and long-term 1–6 hours), including indicators such as mean, standard deviation, extreme values, skewness, and kurtosis. Linear fitting is used to obtain the slope of change, which is used to determine the trend of the indicator, such as whether it is rising, falling, or stable. A change point detection algorithm is used to identify abrupt changes in a signal, thereby capturing acute physiological changes.

[0063] For signals with periodic fluctuations, the Time Series Decomposition (STL) method is used to separate trend, periodic and random components, eliminate interference such as circadian rhythms, and extract the true trend of change.

[0064] Furthermore, time-series forecasting models can be introduced to predict the short-term future, revealing potential anomalies through the deviation between predicted and actual values.

[0065] Ultimately, the results of trend analysis include: the trend direction (increasing, decreasing, or stabilizing) of each parameter, the rate of change, the time point of abrupt change, periodic characteristics, and future predicted trajectories, providing quantitative evidence for subsequent correlation analysis, risk assessment, and clinical decision-making.

[0066] In one possible implementation, the correlation analysis in S106 includes: The multidimensional physiological parameters (such as heart rate, blood pressure, respiratory rate, blood oxygen saturation, cardiac output, etc.) obtained through trend analysis are time-synchronized and normalized. Statistical and time-series methods were used to calculate the linear and nonlinear correlations among the parameters.

[0067] Specifically, linear relationships are assessed using the Pearson correlation coefficient, Spearman rank correlation coefficient is used to measure the consistency of monotonic trends, and mutual information methods are combined to capture potential nonlinear dependencies.

[0068] Furthermore, to identify the time-delay effect between parameters, cross-correlation analysis is used to calculate the correlation strength under different lag times, thereby revealing the correlation of the parameters' temporal characteristics. For dynamic data, a sliding window is used to calculate the time-varying correlation coefficient, obtaining the curve of correlation evolution over time. The obtained correlation matrix is ​​mapped onto a semantic network to update the edge weights between the vital signs parameters, providing a quantitative basis for subsequent risk prediction and decision-making reasoning.

[0069] For example, analyzing the trends of heart rate and blood pressure over time, as well as the correlations between different vital sign parameters, can provide more comprehensive information for decision-making.

[0070] S107. Match the analysis results with the treatment plans of corresponding cases in the expert knowledge base to obtain the reasoning results.

[0071] In one possible implementation, S107 includes: S107a. Perform rule-based reasoning with the analysis results and the treatment plans of corresponding cases in the expert knowledge base to obtain the first reasoning result.

[0072] Furthermore, the rule-based reasoning involves classifying diseases based on pre-defined rules and vital sign parameters.

[0073] Specifically, the vital sign parameters from the analysis results (trend analysis and correlation analysis results) are input into the rule engine and matched with pre-defined rules in the expert knowledge base. These rules are based on disease classification and treatment decision-making logic, combined with physiological thresholds, trends, and dynamic trajectories, and implemented using a hybrid reasoning architecture: including symbolic reasoning and neural network reasoning. The symbolic reasoning part (Drools engine) performs logical matching based on rule conditions, while the neural network reasoning part (LSTM trajectory prediction model) determines trends and predicts states for time-series signals. The system performs matching calculations on the real-time input every second and outputs the first reasoning result, including: disease type judgment, corresponding pathophysiological state, rule matching confidence, and recommended preliminary treatment direction.

[0074] S107b. Perform case reasoning by comparing the analysis results with the treatment plans of corresponding cases in the expert knowledge base to obtain a second reasoning result; the second reasoning result is to perform a similarity search based on the input parameters to find the closest corresponding case.

[0075] Furthermore, based on the same input parameters, a similarity search is performed in the case graph of the knowledge base. By calculating the similarity between the input features and the features of historical case nodes, several most similar typical cases are selected, and their treatment plans and outcome information are extracted. The system integrates the similarity score, treatment response rate, and prognostic results to output a second inference result, including: a set of similar cases, corresponding treatment plans, and their historical success probability distribution.

[0076] S107c: Combine the first reasoning result and the second reasoning result to obtain the final reasoning result.

[0077] Furthermore, during the fusion process, the deterministic logic of rule-based reasoning and the empirical statistical information of case-based reasoning are integrated. These two types of results are combined using a weighted strategy or a Bayesian fusion model to output disease classification confidence, optimal treatment plan, and its predictive effectiveness. The final reasoning results are then fed into the clinical decision-making module to generate individualized treatment recommendations and risk stratification.

[0078] Furthermore, the rule-based reasoning performs matching and reasoning on the input data according to pre-set rules.

[0079] Specifically, the established rules include: when the cardiac index (CI) is consistently below 2.2 L / (min·m²) and the mean arterial pressure (MAP) is below 65 mmHg, it is identified as a low cardiac output state, and volume resuscitation is recommended as a priority; when the systemic vascular resistance index (SVRI) is greater than 1800 dyn·s / cm 5 / m 2Furthermore, when elevated blood pressure is accompanied by an increased heart rate, it indicates a high-resistance condition, and the use of vasodilators is recommended. When the heart rate increases by more than 20% from baseline and blood pressure shows a significant downward trend, it suggests possible septic shock, and the source of infection should be assessed first. Through the logical matching of the above rules, the system can quickly and automatically infer from vital signs parameters to disease categories and preliminary treatment directions.

[0080] S108. Based on the reasoning results and combined with the patient's individual circumstances, generate personalized decision-making suggestions.

[0081] In one possible implementation, S108 includes: S108a. Based on the reasoning results, obtain clinical decisions.

[0082] Furthermore, in S108a, the treatment plan set in the expert knowledge base is matched with the reasoning results, and the expected efficacy and adaptability scores of each candidate plan are calculated. The plan with the highest confidence and consistent with the patient's pathophysiological state is selected as the initial clinical decision. This clinical decision serves as the input basis for subsequent risk assessment and individualized optimization.

[0083] S108b, Risk stratification based on clinical decision-making.

[0084] Furthermore, in S108b, a risk score is comprehensively calculated based on key physiological parameters (including cardiac index CI, systemic vascular resistance index SVRI, mean arterial pressure MAP, lactate level, and heart rate HR) in the clinical decision-making and reasoning results, as well as the patient's individual characteristics (including age, underlying diseases, organ function status, etc.); and risk stratification is performed based on the risk score.

[0085] Specifically, the risk scoring model uses a logistic regression model or a neural network probability assessment method to quantitatively assess the risk of short-term deterioration, shock progression, or treatment failure in patients.

[0086] S108c. Based on risk stratification, a basic treatment plan is obtained. Patients are classified into warning levels I to IV according to their risk score ranges, where a risk probability ≥0.9 is Level I, 0.7–0.9 is Level II, 0.5–0.7 is Level III, and ≤0.5 is Level IV. The risk stratification results are used to guide the intensity of subsequent interventions and the selection of the basic treatment plan.

[0087] Furthermore, in S108c, based on the risk level and the type of disease in the clinical decision-making process, a matching basic treatment protocol template is retrieved from the expert knowledge base. The basic protocol includes treatment goals, drug selection, dosage range, and monitoring frequency. High-risk levels (Level I and II) correspond to intensive intervention protocols, such as rapid volume resuscitation and combined vasoactive drug therapy; medium- and low-risk levels (Level III and IV) correspond to maintenance therapy and dynamic monitoring protocols. Through a dual index of risk level and disease type, the corresponding basic protocol is automatically selected as input for subsequent Monte Carlo simulations.

[0088] S108d generates personalized decision-making suggestions based on the basic plan, combined with Monte Carlo simulation and the patient's individual circumstances.

[0089] Furthermore, S108d includes: within the framework of the basic protocol, combining the patient's individual parameters (including weight, cardiac reserve, previous medication response, etc.) and generating candidate protocols based on an expert knowledge base; using Monte Carlo simulation to quantitatively evaluate and optimize the candidate protocols, thereby generating the final individualized decision recommendation.

[0090] Specifically, S108d includes: (1) Probability distribution modeling: Under the framework of the basic plan, candidate plans are generated based on the expert knowledge base by combining the patient's individual parameters (including weight, cardiac function reserve, previous drug response, etc.); and by combining the patient's real-time physiological parameters, historical treatment response data and population pharmacokinetic / pharmacodynamic model, key physiological parameters and treatment-related characteristics are modeled as probability distributions to quantify the uncertainty of physiological system and treatment response.

[0091] Optionally, the key physiological parameters include: cardiac output (CO), cardiac index (CI), systemic vascular resistance (SVR), systemic vascular resistance index (SVRI), stroke volume (SV), and stroke index (SVI).

[0092] (2) Random sampling and path simulation: Multiple Monte Carlo random samplings are performed on the candidate treatment plans, and a set of parameter values ​​are randomly selected and input into a pre-trained intracardiac prediction model based on the Transformer architecture. The model is used to simulate the hemodynamic trajectory and physiological changes in a specific time period after the treatment plan is implemented, and the simulation results are obtained. The simulation results include the predicted values ​​of key hemodynamic parameters at each time step, such as cardiac output (CO), cardiac index (CI), and systemic vascular resistance (SVR).

[0093] It should be noted that thousands of independent randomized simulations were performed on the same candidate treatment to ensure that the outcome space that may result from the uncertainty of all parameters was fully explored, thereby obtaining a stable statistical distribution of various possible outcomes under the treatment.

[0094] (3) Based on the predefined clinical target thresholds and adverse event criteria, the entire time trajectory generated by this simulation is scanned and analyzed to determine the outcome of this simulation and record key indicators.

[0095] (4) Comprehensive utility assessment and optimal treatment plan: Based on all simulation results, calculate the expected outcome index distribution for each candidate treatment plan, including the probability of key physiological parameters being within the clinical target range and the probability of adverse events, as well as the probability of drug accumulation or specific reaction risk based on individual patient conditions (such as liver and kidney function).

[0096] It should be noted that the adverse events mentioned include hypotension, tachycardia, and insufficient organ perfusion.

[0097] (5) Based on the distribution of expected outcome indicators and dynamically adjust the weight of each indicator according to the risk classification determined in S108b, conduct a comprehensive utility evaluation for each candidate solution, and select the candidate solution with the highest score as an individualized decision recommendation.

[0098] Specifically, the calculation formula for the comprehensive utility assessment is as follows:

[0099] in: The final overall utility score for candidate treatment options; The probability of achieving the objective of this scheme is directly calculated from the aforementioned Monte Carlo simulation results; The probability of adverse events occurring under this scheme is directly calculated from the aforementioned Monte Carlo simulation results; The individual tolerability risk score for this regimen is assessed and quantified by the system based on the patient's individual circumstances (such as liver and kidney function and cardiac reserve mentioned in S108).

[0100] , , This is a dynamic weighting coefficient, the value of which is set and adjusted according to the risk classification (CH) determined in step S108b. For example, the system will automatically assign a weight to high-risk patients. A higher weighting for adverse events reflects a prioritization of safety in clinical decision-making. All weighting coefficients satisfy the normalization condition. The design goal of this function is to maximize The value is determined to achieve the optimal balance between efficacy and safety.

[0101] The personalized final decision recommendations include treatment suggestions, dynamic monitoring indicators, and reassessment timelines, providing clinicians with quantitative and interpretable decision support.

[0102] The following description, in conjunction with more detailed embodiments, provides further details.

[0103] Non-invasive hemodynamic monitoring and decision support methods include the following steps: Step 1: Data Acquisition and Preprocessing The data acquisition includes obtaining physiological data for patients who require simultaneous electrocardiogram (ECG), central venous pressure (CVP), and minimally invasive hemodynamic monitoring due to their condition. The corresponding indicators are obtained through the following devices: ECG monitor (blood pressure, heart rate, respiration, oxygen saturation, CVP), and pulse contour method minimally invasive hemodynamic monitoring (cardiac output, cardiac index, stroke volume, stroke index, systemic vascular resistance, systemic vascular resistance index). The preprocessing includes outlier filtering, standardization, and logarithmic transformation. Outlier filtering removes records containing null values ​​(NaN) or zero values ​​to ensure data integrity, and sets clinically reasonable thresholds to filter out abnormal data. Standardization uses a robust standardization algorithm, with the following formula: In the formula, IQR is the interquartile range; the logarithmic transformation is performed on SVR and SVRI to alleviate the skewness of the numerical distribution.

[0104] Step 2: Construct a cardiac prediction model based on the Transformer architecture. The Transformer architecture includes: an input embedding layer, a Transformer encoder, and a regression output layer. The input embedding layer maps 7-dimensional static parameters to a 256-dimensional feature space through a fully connected layer and adds sinusoidal positional encoding. The Transformer encoder contains 4 encoder layers, each with a 16-head self-attention mechanism, and the feedforward network has a dimension of 512. The regression output layer directly outputs cardiac output (CO), cardiac index (CI), stroke volume (SV), stroke index (SVI), systemic vascular resistance (SVR), and systemic vascular resistance index (SVRI) from the fully connected layer (256→512→6).

[0105] Step 3: Model Training The dataset used for model training consists of multi-center clinical data (over 8,000,000 records), divided into training, validation, and test sets in an 8:1:1 ratio. The loss function is the dynamically weighted average absolute error (MAE), with weights set as follows: w = [0.5, 0.5, 0.5, 0.5, 0.75, 0.75], corresponding to outputs CO, CI, SV, SVI, SVR, and SVRI; the loss function is defined as: ,in ; The optimization strategy is an initial learning rate of 0.001, gradient pruning (threshold 1.0), and an early stopping mechanism (terminating the process after 15 rounds of verification that the loss has not improved).

[0106] Step 4: Real-time monitoring and verification.

[0107] Deployment: Integrate the trained cardiac prediction model into a portable monitor to support real-time prediction at 1Hz; Test Results: On the independent test set (2,400 cases), the mean relative accuracy of predictions compared to invasive measurements was: CO: 0.8036, CI: 0.8410, SV: 0.8216, SVI: 0.8062, SVR: 0.8555, SVRI: 0.8603. This significantly outperformed the random forest model (CO: 0.4907, CI: 0.4411, SV: 0.3216, SVI: 0.3062, SVR: 0.4654, SVRI: 0.4562). Step 5: Collect expert experience and knowledge, organize it into a semantic network structure; use the knowledge representation method of semantic network to represent the acquired knowledge in the form of nodes and edges, form the association between vital signs and diseases, and then build a relational expert knowledge base; Step 6: Perform trend analysis and correlation analysis on the third dataset, the real-time electrocardiogram monitor physiological parameters of the cardiac intensive care subjects, and the electronic medical records to obtain the analysis results; Step 7: Match the analysis results with the treatment plans of corresponding cases in the expert knowledge base to obtain the reasoning results; Step 8: Based on the reasoning results and the patient's individual circumstances, generate personalized decision-making suggestions.

[0108] Gradient clipping refers to scaling the gradient during backpropagation if the norm of the gradient (i.e., the length of the gradient vector) exceeds a set threshold, so that it does not exceed that threshold.

[0109] In the specific implementation process, step 1 uses data collected clinically at a hospital. The dataset provides both invasive and non-invasive indicators. This dataset mainly includes static physiological parameters of subjects monitored intracardiacly via medical sensors, including heart rate, blood oxygen saturation, non-invasive systolic blood pressure, non-invasive mean blood pressure, respiratory rate, pulse rate, and central venous pressure, as well as hemodynamic parameter detection results, including cardiac output, cardiac index, stroke volume, stroke index, systemic vascular resistance, and systemic vascular resistance index. A total of 63 patient samples are included. In this embodiment, 50 patient samples are randomly selected to train the model, 6 patient samples to validate the model, and 7 patient samples to test the model results.

[0110] To verify the model's effectiveness on different patients' physiological conditions, data from patients who required simultaneous electrocardiogram monitoring, central venous pressure monitoring, and minimally invasive hemodynamic monitoring due to their condition were selected. For example... Figure 5 As shown, the data is obtained from monitoring with an electrocardiogram (blood pressure, heart rate, respiration, oxygen saturation, central venous pressure) and minimally invasive hemodynamic monitoring using the pulse contour method (cardiac output, cardiac index, stroke volume, stroke index, systemic vascular resistance, systemic vascular resistance index). Specific detection methods include blood pressure monitoring in a resting state, with the middle of the cuff placed at the level of the right atrium (midpoint of the sternum) on the patient's upper arm. An appropriate cuff size is used, ensuring the cuff surrounds 75%–100% of the arm, with a width of 37%–50% of the arm circumference. The cuff is smoothly wrapped around the middle of the patient's upper arm, with a tightness that allows for the insertion of one finger. The lower edge is 2–3 cm from the elbow crease. Blood pressure monitoring is set to automatic measurement every 1 hour. For heart rate and respiration monitoring, the standard electrode positions are: RA is right upper quadrant, located at the first intercostal space along the right midclavicular line; RL is right lower quadrant, located at the level of the xiphoid process along the right midclavicular line; V or C is the precordial lead, located at the fourth intercostal space along the left sternal border; LA is left upper quadrant, located at the level of the left midclavicular line. The first intercostal space; LL is the lower left, located at the level of the xiphoid process on the left midclavicular line. The oxygen saturation sensor is correctly placed on the patient's finger, toe, or ear. Finger oxygen saturation measurement should ensure good blood circulation at the fingertip, and should be performed on a fingertip without nail bed staining or disease. Central venous pressure (CVP): Measure CVP in the supine position at rest. Set the CVP waveform to the optimal scale and use the main lumen of the central venous catheter to monitor CVP. It is recommended to fix the sensor at the level of the 4th intercostal space at the midaxillary line, at the upper 1 / 3 level of the vertical distance between the anteroposterior and posterior diameters of the chest, or at the level 5cm below the sternal angle. The same surface landmark should be used for the same patient, and the pressure monitoring catheter should be kept patent. For synchronous continuous pulse contour method minimally invasive hemodynamic monitoring, the radial artery (the contralateral arm to which the cuff pressure is measured) is selected. The pressure sensor is placed at the level of the right atrium, i.e., the cardiac zero-point surface reference point (level of the 4th intercostal space at the midaxillary line), for zeroing, and the pressure monitoring catheter should be kept patent. In step 2, specific hyperparameter tuning and optimization methods such as grid search and Bayesian optimization were used to explore the combined effects of the number of encoder layers (2-6) and the number of attention heads (4-16) to determine the comprehensive optimal hyperparameter combination.

[0111] The actual results obtained in step S4 are as follows: The mean squared error (MSE) is expressed as follows: Cardiac output (CO): MSE = 15.36 ,; Cardiac Index (CI): MSE = 4.29 ; Systemic vascular resistance (SVR): MSE = 1.77 ; Systemic vascular resistance index (SVRI): MSE=6.14 ; Stroke volume (SV): MSE = 1,182.7 ; SVI (Stakes Index): MSE = 314.96 .

[0112] Relative accuracy: Cardiac output (CO): 80.36%; Cardiac index (CI): 84.10%; Systemic vascular resistance (SVR): 85.55%; Systemic vascular resistance index (SVRI): 86.03%; Stroke Count (SV): 82.16%; SVI (Stakes Index): 80.62%.

[0113] The method provided in this application will be illustrated below with specific examples.

[0114] To verify the model's effectiveness on patients with different physiological conditions, data from patients requiring simultaneous ECG monitoring, central venous pressure monitoring, and minimally invasive hemodynamic monitoring were selected. This data was obtained through ECG monitoring (blood pressure, heart rate, respiration, oxygen saturation, central venous pressure) and minimally invasive hemodynamic monitoring using the pulse contour method (cardiac output, cardiac index, stroke volume, stroke index, systemic vascular resistance, systemic vascular resistance index).

[0115] Specific detection methods include blood pressure monitoring in a resting state. The middle of the cuff is placed on the patient's upper arm at the level of the right atrium (midpoint of the sternum). An appropriate cuff size is used, ensuring the cuff covers 75%–100% of the arm, with a width of 37%–50% of the arm circumference. The cuff is smoothly wrapped around the middle of the patient's upper arm, with a tightness that allows one finger to fit comfortably. The lower edge is 2–3 cm from the elbow crease. Blood pressure monitoring is set to automatic measurement every 1 hour. For heart rate and respiration monitoring, the standard electrode placement is: RA is right upper-right, located at the first intercostal space along the right midclavicular line; RL is right lower-right, located at the level of the xiphoid process along the right midclavicular line; V or C are precordial leads, located at the fourth intercostal space along the left sternal border; LA is left upper-left, located at the first intercostal space along the left midclavicular line; LL is left lower-left, located at the level of the xiphoid process along the left midclavicular line. Oxygen saturation sensor: The oxygen saturation sensor is correctly placed on the patient's finger, toe, or ear. Finger oxygen saturation measurement should ensure good blood circulation at the fingertip, and measurement should be performed on fingertips without nail bed staining or disease. Central venous pressure (CVP): Measure CVP in the supine position at rest. Set the CVP waveform to the optimal scale and use the main lumen of the central venous catheter to monitor CVP. It is recommended to fix the sensor at the level of the 4th intercostal space along the midaxillary line, at the level of the upper 1 / 3 of the vertical distance between the anteroposterior diameter of the chest, or at the level 5 cm below the sternal angle. The same surface landmark should be used for the same patient, and the pressure monitoring catheter should be kept patent. Simultaneous continuous pulse contour method for minimally invasive hemodynamic monitoring: Select the radial artery (the radial artery of the arm opposite to the cuff pressure measurement). Zero the pressure sensor at the level of the right atrium (the cardiac zero-point surface reference point, at the level of the 4th intercostal space along the midaxillary line), and keep the pressure monitoring catheter patent.

[0116] The overall framework of the Transformer-based intracardiac intelligent prediction method based on static physiological parameters is implemented in steps. First, static physiological parameters (HR, SpO2, NIBP, etc.) are collected using conventional monitoring equipment, and data cleaning and standardization preprocessing are completed. Then, a lightweight model architecture based on Transformer is constructed, and the self-attention mechanism is used to efficiently model the global correlation features between multiple parameters. The architecture is described in [link to architecture]. Figure 2 During model training, a multi-layer Transformer encoder is used to progressively uncover the deep nonlinear relationships between parameters, establishing an end-to-end mapping between static physiological parameters and intracardiac functional parameter labels (CO, CI, SVR), without relying on dynamic signal input. See also... Figure 3 , Figure 4 The prediction model achieved a relative accuracy of 87.47% in CO prediction and 88.67% in SV prediction. Its non-invasive prediction results showed high consistency with traditional invasive detection methods, providing a convenient and accurate tool for clinical cardiac function assessment. The final prediction results were output through a visualization module and validated experimentally using an intensive care unit dataset. (See [link to relevant documentation]). Figure 5 .

[0117] Based on this, personalized decision-making suggestions are generated through steps 5-8.

[0118] For example, the final personalized decision recommendations are as follows: Cardiac index (CI) threshold: <2.2 L / (min·m²) indicates low cardiac output, ≥4.0 L / (min·m²) indicates high cardiac output. Systemic vascular resistance index (SVRI) threshold: <1200 dyn·s / cm 5 / m 2 For low resistance, 1200-1800 dyn·s / cm 5 / m 2 Normal resistance, >1800 dyn·s / cm 5 / m 2 It has high resistance.

[0119] Treatment principles: For all cases of shock, the cause should be treated first (such as infection or ischemia), and volume status should be maintained (volume expansion). The use of vasoactive drugs requires monitoring of blood pressure and heart rate.

[0120] Rule 1: High-exhaustion, low-resistance shock (CI≥4.0L / (min·m²), SVRI<1200dyn·s / cm) 5 / m 2 ) Pathophysiology: Commonly seen in distributive shock (such as septic shock), characterized by vasodilation, low resistance, and relatively high cardiac output.

[0121] Treatment recommendations: Vasoactive drugs: Norepinephrine is the first choice to increase vascular resistance and improve blood pressure. If ineffective, metaraminol can be added.

[0122] Anti-infection: If infection is suspected, administer empirical antibiotics immediately (such as ceftazidime for injection or levofloxacin hydrochloride), adjusting the dosage based on culture and drug sensitivity results.

[0123] Anti-inflammatory treatment: In cases of septic shock, consider glucocorticoids or ulinastatin to suppress the inflammatory response.

[0124] Other: Ensure adequate volume (crystal resuscitation); avoid diuretics (unless there is heart failure or clear volume overload).

[0125] Rule 2: Low-emission, high-resistance shock (CI < 2.2 L / (min·m²), SVRI > 1800 dyn·s / cm) 5 / m 2 ) Pathophysiology: Commonly seen in cardiogenic shock (such as after myocardial infarction) or hypovolemic shock, characterized by low cardiac output and high resistance.

[0126] Treatment recommendations: Positive inotropic agents: The use of positive inotropic agents is determined based on the etiology. In low-output shock caused by acute myocardial infarction, positive inotropic agents should be used with caution during the acute phase. In patients with low-output, high-resistance shock not caused by acute myocardial infarction, dobutamine may be considered as the first-line treatment to enhance myocardial contractility. If the effect is unsatisfactory, digitalis glycosides (such as digoxin), phosphodiesterase inhibitors (such as milrinone), or calcium sensitizers (such as levosimendan) can be added.

[0127] Vasoactive drugs: Use norepinephrine with caution (only in small doses in cases of severe hypotension to maintain perfusion pressure, but avoid increasing afterload). In cases of hypovolemia, prioritize volume expansion (crystalloid solutions); small doses of dopamine can be used for renal protection, but large doses should be avoided.

[0128] Mechanical circulatory support devices: After ruling out contraindications, mechanical circulatory support devices such as intra-aortic balloon pump (IABP) and extracorporeal membrane oxygenation (ECMO) can be considered.

[0129] Anti-inflammatory treatment: only used for infection-related shock (such as ulinastatin or glucocorticoids), and not routinely used for non-infectious shock (such as cardiac shock).

[0130] Other measures include: treating the underlying cause (e.g., if ischemia is suspected, performing emergency coronary revascularization); monitoring blood volume to avoid over-expansion.

[0131] Rule 3: Low-emission, low-resistance shock (CI < 2.2 L / (min·m²), SVRI < 1200 dyn·s / cm) 5 / m 2 ) Pathophysiology: Commonly seen in late-stage shock (such as severe infection or heart failure), characterized by low cardiac output, low resistance, and insufficient perfusion of multiple organs.

[0132] Treatment recommendations: Vasoactive drugs: Combined with norepinephrine, dopamine, metaraminol and other vasoactive drugs.

[0133] Positive inotropic drugs: Dobutamine is the first choice, and phosphodiesterase inhibitors or calcium sensitizers can be added if necessary.

[0134] Mechanical circulatory support devices: After ruling out contraindications, mechanical circulatory support devices such as intra-aortic balloon pump (IABP) and extracorporeal membrane oxygenation (ECMO) can be considered.

[0135] Anti-inflammatory and anti-infective treatment: If infection is considered to be the main cause, use antibiotics (as in Rule 1) and anti-inflammatory drugs (glucocorticoids or ulinastatin).

[0136] Other: Optimize capacity status and monitor cardiac function.

[0137] Rule 4: Intermediate or unclassified shock (CI and SVRI are in other ranges, such as CI < 2.2 L / (min·m²), SVRI 1200-1800 dyn·s / cm²). 5 / m 2 Or CI ≥ 4 L / (min·m²), SVRI > 1800 dyn·s / cm 5 / m 2 ) Pathophysiology: This could be early shock, a mixed or non-shock state (such as compensated hypertension). Further etiology needs to be evaluated.

[0138] Treatment recommendations: Prioritize etiological diagnosis (e.g., echocardiography, infection markers). Basic support: ensure adequate fluid intake and oxygenation.

[0139] Individualized medication: If CI is low and SVRI is normal, refer to rule 2 or 3; if CI is high and SVRI is high, lower blood pressure or reduce afterload.

[0140] The non-invasive hemodynamic detection and decision support device provided by the present invention is described below. The non-invasive hemodynamic detection and decision support device described below can be referred to in correspondence with the non-invasive hemodynamic detection and decision support method described above.

[0141] Figure 6 This is a schematic diagram of the non-invasive hemodynamic detection and decision support device provided in an embodiment of the present invention, as shown below. Figure 6 As shown, it includes: an acquisition module 61, a preprocessing module 62, a first construction module 63, an output module 64, a second construction module 65, an analysis module 66, an inference module 67, and a generation module 68, wherein: The acquisition module 61 is used to acquire a first dataset, which is obtained by collecting real-time electrocardiogram physiological parameters of the cardiac critical care monitoring subjects. Preprocessing module 62 preprocesses the first dataset to obtain the second dataset; The first building block 63 is used to build and train a cardiac prediction model based on the Transformer architecture; Output module 64 is used to input the second dataset into the trained intracardiac prediction model and output the predicted hemodynamic parameters as the third dataset. The second construction module 65 is used to collect expert experience knowledge and organize it into a semantic network structure; using the knowledge representation method of semantic network, the acquired knowledge is represented in the form of nodes and edges to form the association between vital signs parameters and diseases, and then a relational expert knowledge base is constructed. Analysis module 66 is used to perform trend analysis and correlation analysis on the third dataset, the real-time electrocardiogram monitor physiological parameters of the cardiac critical care monitoring subjects, and the electronic medical records to obtain analysis results; The reasoning module 67 is used to match and reason with the treatment plans of corresponding cases in the expert knowledge base to obtain the reasoning results; The generation module 68 is used to generate personalized decision-making suggestions based on the reasoning results and the patient's individual circumstances.

[0142] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 710, a communications interface 720, a memory 730, and a communications bus 740. The processor 710, communications interface 720, and memory 730 communicate with each other via the communications bus 740. The processor 710 can call logical instructions from the memory 730 to execute non-invasive hemodynamic detection and decision support methods.

[0143] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0144] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the non-invasive hemodynamic detection and decision support methods provided by the above methods.

[0145] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the non-invasive hemodynamic detection and decision support methods provided by the above methods.

[0146] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0147] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-invasive hemodynamic detection and decision support method, characterized in that, The method comprises the following steps: acquiring a first data set by collecting real-time electrocardiogram monitor physiological parameters of a critical heart monitoring object; preprocessing the first data set to obtain a second data set; constructing and training an intracardiac prediction model based on a Transformer architecture; inputting the second data set into the trained intracardiac prediction model to output a predicted blood flow parameter prediction value as a third data set; collecting expert experience knowledge and arranging the knowledge into a semantic network structure; using a semantic network knowledge representation method to represent the acquired knowledge in the form of nodes and edges to form the association relationship between the sign parameters and diseases, and then constructing a relational expert knowledge base; performing trend analysis and correlation analysis on the third data set, real-time electrocardiogram monitor physiological parameters of the critical heart monitoring object and electronic medical records to obtain an analysis result; matching and reasoning the analysis result with a corresponding case treatment scheme in the expert knowledge base to obtain a reasoning result; generating a personalized decision suggestion based on the reasoning result and in combination with the individual condition of the patient.

2. The non-invasive hemodynamic monitoring and decision assistance method of claim 1, wherein, The electrocardiogram monitor physiological parameters include: heart rate HR, oxygen saturation SpO2, non-invasive systolic blood pressure NIBP-S, non-invasive diastolic blood pressure NIBP-D, respiratory rate RR, pulse rate PR and central venous pressure CVP.

3. The non-invasive hemodynamic monitoring and decision assistance method of claim 1, wherein, The intracardiac prediction model based on the Transformer architecture comprises: an input embedding layer composed of a full connection layer, a GELU activation function and layer normalization, which is used to map static parameters to a multi-dimensional feature space to obtain a multi-dimensional feature vector; a Transformer encoder containing multiple encoder layers, each layer being provided with a multi-head self-attention mechanism and a feedforward network dimension for global relationship modeling of the multi-dimensional feature vector to obtain standardized high-order features; and a regression output layer composed of a full connection layer, a GELU activation, a Dropout, a layer normalization and a final full connection layer, which is used to decode the high-order feature representation into a blood flow parameter prediction value through the full connection layer, the GELU activation and the regularization operation.

4. The non-invasive hemodynamic monitoring and decision assistance method of claim 1, wherein, The third data set includes: cardiac output CO, cardiac index CI, stroke volume SV, stroke volume index SVI, peripheral vascular resistance SVR and peripheral vascular resistance index SVRI.

5. The non-invasive hemodynamic monitoring and decision assistance method of claim 1, wherein, The semantic network knowledge representation method is used to represent the acquired knowledge in the form of nodes and edges to form the association relationship between the sign parameters and diseases, which comprises: The semantic network knowledge representation method is used to take the sign parameters, disease names, pathological physiological conditions and treatment schemes as nodes, establish the relationship between the nodes through edges, and then form the association relationship between the sign parameters and diseases.

6. The non-invasive hemodynamic monitoring and decision assistance method of claim 1, wherein, The matching and reasoning of the analysis result with the corresponding case treatment scheme in the expert knowledge base to obtain the reasoning result comprises: performing rule reasoning of the analysis result with the corresponding case treatment scheme in the expert knowledge base to obtain a first reasoning result; the rule reasoning is to classify the diseases based on the sign parameters according to a preset rule. The analysis result is case-inference with the treatment scheme of the corresponding case in the expert knowledge base, and a second inference result is obtained; the second inference result is a similar search according to the input parameters to search for the closest corresponding case; The first inference result and the second inference result are fused to obtain a final inference result.

7. The non-invasive hemodynamic monitoring and decision assistance method of claim 1, wherein, The personalized decision suggestion is generated based on the inference result and combined with the individual situation of the patient, including: Obtaining a clinical decision based on the inference result; Risk classification based on the clinical decision; Obtaining a basic scheme based on the risk classification; Based on the basic scheme, combined with Monte Carlo deduction and the individual situation of the patient, a personalized decision suggestion is generated.

8. Non-invasive hemodynamic detection and decision support device, characterized in that It includes: The acquisition module is used for acquiring a first data set, and the first data set is obtained by collecting real-time electrocardiogram monitor physiological parameters of a heart critical monitoring object; The preprocessing module is used for preprocessing the first data set to obtain a second data set; The first construction module is used for constructing and training an intracardiac prediction model based on a Transformer architecture; The output module is used for inputting the second data set to the trained intracardiac prediction model, and outputting a predicted hemodynamic parameter prediction value as a third data set; The second construction module is used for collecting expert experience knowledge and arranging the knowledge into a semantic network structure; the knowledge is represented in the form of nodes and edges by using a semantic network knowledge representation method, forming the association relationship between the signs and diseases, and then constructing a relational expert knowledge base; The analysis module is used for trend analysis and correlation analysis of the third data set, real-time electrocardiogram monitor physiological parameters of the heart critical monitoring object and electronic medical records to obtain an analysis result; The inference module is used for matching inference of the analysis result and the treatment scheme of the corresponding case in the expert knowledge base to obtain an inference result; The generation module is used for generating a personalized decision suggestion based on the inference result and combined with the individual situation of the patient.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the noninvasive hemodynamic detection and decision assistance method of any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the noninvasive hemodynamic detection and decision assistance method of any one of claims 1 to 7.

Images