ICU return risk prediction method based on patient state perception

Abstract

The invention relates to the technical field of ICU return risk prediction, in particular to an ICU return risk prediction method based on patient state perception. According to the method, a patient state time sequence matrix is constructed through multi-source heterogeneous data collection, and risk scores and a key factor set are generated based on feature extraction and integrated learning; constructing a dynamic risk association graph by adopting a two-stage association analysis method of Pearson correlation analysis and Granger causality test, and revealing a causality and a time sequence relationship among risk factors; performing topological optimization and path weight calculation on the association graph to identify a high-risk evolution path; and finally, integrating the risk score, the association map and the evolution path to generate a visual decision support report. According to the method, the limitation that only isolated risk factors can be provided in a traditional method is broken through, the dynamic association and conduction path of the risk factors can be clearly displayed, a complete basis is provided for a clinician to formulate a targeted intervention scheme, and the accuracy and safety of ICU patient transfer decision making are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ICU reentry risk prediction, in particular to an ICU reentry risk prediction method based on patient state perception. BACKGROUND

[0002] In the clinical management of intensive care units (ICUs), the decision of patient transfer timing is directly related to medical safety and resource utilization efficiency. Currently, this critical decision mainly relies on the subjective judgment of doctors' clinical experience, lacks objective and systematic quantitative tools, and is prone to increase the reentry rate or underutilize bed resources.

[0003] The prior art usually uses a machine learning model to predict the risk and provides a list of key risk factors (such as blood oxygen saturation, respiratory rate, etc.) through feature importance analysis technology, to explain the basis of the model's decision while outputting a risk score.

[0004] However, this method has certain limitations: the explanation information it provides is static and isolated, only showing the independent contribution of a single factor, and it is difficult to reveal the dynamic relationship and collaborative evolution mechanism between multiple factors, limiting its practical guiding value for clinical intervention.

[0005] Therefore, there is an urgent need in the art for an innovative solution that can not only maintain accurate risk warning capabilities but also deeply analyze risk causes and clearly present the dynamic relationship between key indicators, thereby providing more insightful support for clinical decision-making. SUMMARY

[0006] The present application provides an ICU reentry risk prediction method based on patient state perception, which constructs a patient state time series matrix and performs risk prediction, then creates a dynamic risk association graph to reveal the causal relationship between key risk factors, and finally identifies and outputs high-risk evolution paths and generates integrated visual decision support reports, thereby solving the problems raised in the background art, i.e.: In the decision-making of ICU patient transfer, the existing warning system can only list isolated risk factors and cannot show the dynamic relationship and evolution process between these factors, making it difficult for doctors to develop targeted intervention measures.

[0007] To achieve the above purpose, the risk prediction method includes the following steps: S1, collect multi-source heterogeneous patient data, clean, align and normalize the patient data, and generate a patient state time series matrix; S2, feature extraction is performed on the patient state time series matrix to construct a patient comprehensive feature vector, the patient comprehensive feature vector is input into a prediction model to obtain an initial reentry risk score, and a set of key risk factors is generated based on model interpretability analysis; S3, extracting corresponding data from the patient state time series matrix based on the key risk factor set to form a key factor time series sub-matrix, performing two-stage correlation analysis on the key factor time series sub-matrix, including calculating the Pearson correlation coefficient between factors and performing Granger causality test, and constructing a dynamic risk correlation graph based on the results of the correlation analysis, wherein the nodes represent risk factors and the directed edges represent causal relationships; S4, topologically optimizing the dynamic risk correlation graph to obtain a directed acyclic graph, calculating the weight of each path in the directed acyclic graph, identifying candidate high-risk paths based on the path weight, and determining a high-risk evolution path set by evaluating the clinical significance index of the candidate high-risk paths; S5, receiving an initial return risk score, a key risk factor set, a dynamic risk correlation graph, and a high-risk evolution path set, and integrating to generate an integrated visual decision support report.

[0008] In the above technical solution, the cooperative design of each step builds a complete technical chain from data perception to clinical decision, and the design concept is to systematically solve the core defect of the traditional risk prediction model that the explanation is insufficient through layer-by-layer progressive data processing and analysis. Specifically, if only steps S1 and S2 are included, the system can complete risk quantification evaluation, but the output is still a discrete feature importance ranking, which cannot explain the internal interaction mechanism and dynamic evolution law between each risk factor. The two-stage correlation analysis introduced in step S3 is designed to address this deficiency. Through Pearson correlation analysis, the statistical correlation between factors is established, and then Granger causality test is combined to further explore the driving relationship in time sequence, thereby converting the static risk snapshot into a dynamic topological network with causal direction. The graph optimization and path discovery mechanism of step S4 further overcomes the problems of poor readability and lack of emphasis of complex networks. Through directed acyclic graph conversion, the clarity of risk transmission logic is ensured, and through a weight formula that comprehensively considers causal strength and path integrity, the key transmission chain with the most clinical intervention value is accurately selected. Finally, the integrated report of step S5 realizes the conversion of analysis results to clinical insights, and through multi-dimensional visualization, the abstract risk mechanism is analyzed into a concrete guide that conforms to the logic of medical decision-making, so that doctors not only know the risk level, but also can clearly grasp the source, transmission path and key intervention nodes of the risk. The establishment of this progressive technical path enables the present application to evolve from a simple "risk warning" tool to a "risk analysis and decision support" system, realizing the qualitative improvement of technical value.

[0009] On this basis, the two-stage correlation analysis includes first calculating the Pearson correlation coefficient to determine the covariation relationship, and then performing Granger causality test to analyze the lead-lag relationship in time sequence.

[0010] In another technical solution, the size of the node in the dynamic risk correlation graph is proportional to the contribution weight of the corresponding factor, the direction of the edge is determined by the Granger causality, and the thickness of the edge is proportional to the causality strength.

[0011] In this technical solution, the two-stage correlation analysis and the collaborative design of the dynamic risk correlation graph constitute a complete analysis chain from statistical correlation to causal inference. If only Pearson correlation analysis is performed, although the covariation pattern between factors can be identified, the causal timing cannot be distinguished, and the symptoms occurring at the same time may be misjudged as a causal relationship. Although the Granger causality test can determine the timing relationship, the lack of covariation relationship may lead to the establishment of false causal links. The combination of the two forms a mutually verified analysis framework: Pearson correlation ensures the substantial correlation between factors, and Granger causality further reveals the driving direction. This double verification mechanism significantly improves the reliability of the correlation relationship. On this basis, the visual design of the dynamic risk correlation graph converts the abstract mathematical relationship into an intuitive expression consistent with clinical cognition: the size difference of the node highlights the core risk factors, the directed edge clearly shows the conduction direction of the pathological mechanism, and the thickness of the edge quantifies the influence strength. This multi-dimensional visual coding enables doctors to quickly grasp the key path and core link in the complex risk network, providing a clear target orientation for precise intervention.

[0012] Compared with the prior art, the beneficial effects of the present application are: 1. The present application realizes an important breakthrough from traditional single risk factor early warning to systematic risk mechanism analysis by constructing a dynamic risk correlation graph and identifying high-risk evolution paths. This technical solution adopts a two-stage correlation analysis method, first determines the statistical correlation between risk factors through Pearson correlation analysis, and then reveals the timing driving relationship through Granger causality test, establishing a risk network with clear causal direction. This design enables clinicians to break through the limitations of traditional methods that can only focus on single risk factors, and to understand the dynamic correlation and conduction mechanism between risk factors from a system level, providing complete evidence support for developing precise intervention programs targeting the source of risk.

[0013] 2、The present application realizes the whole chain optimization from data acquisition to clinical decision by establishing a complete end-to-end risk analysis and decision support process. The scheme first constructs a patient state time series matrix through multi-source heterogeneous data fusion, then combines feature engineering and ensemble learning for risk prediction, and then adopts graph structure optimization and path weight calculation to identify the key risk transmission path, and finally presents the analysis results through a multi-level visualization system. This systematic integration of technology not only provides accurate risk warning, but more importantly, by revealing the internal mechanism and dynamic evolution law of risk formation, doctors can make clinical decisions based on complete pathophysiological evidence, significantly improving the accuracy of ICU patient discharge risk assessment and the pertinence of clinical intervention. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The present application is a patient state perception-based ICU re-entry risk prediction method overall structure flow chart; Figure 2 The present application is a step S3 flow chart; Figure 3 The present application is a step S4 flow chart. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0016] Currently, in the ICU patient discharge decision, the existing early warning system can only list isolated risk factors and cannot show the dynamic relationship and evolution process between these factors, which makes it difficult for doctors to develop targeted intervention measures. The present application provides a patient state perception-based ICU re-entry risk prediction method, as shown in Figure 1 The method comprises the following steps: S1, collect multi-source heterogeneous patient data, clean, align and normalize the patient data, and generate a patient state time series matrix; S2, feature extraction is performed on the patient state time series matrix to construct a patient comprehensive feature vector, the patient comprehensive feature vector is input into a prediction model to obtain an initial re-entry risk score, and a key risk factor set is generated based on model interpretability analysis; S3, extracting corresponding data from the patient state time series matrix based on the key risk factor set to form a key factor time series sub-matrix, performing two-stage correlation analysis on the key factor time series sub-matrix, including calculating the Pearson correlation coefficient between factors and performing Granger causality test, and constructing a dynamic risk correlation graph based on the results of the correlation analysis, wherein the nodes represent risk factors and the directed edges represent causal relationships; S4, topological optimization of the dynamic risk correlation graph to obtain a directed acyclic graph, calculating the weight of each path in the directed acyclic graph, identifying candidate high-risk paths based on path weights, and determining a high-risk evolution path set by evaluating the clinical significance index of the candidate high-risk paths; S5, receiving an initial return risk score, a key risk factor set, a dynamic risk correlation graph and a high-risk evolution path set, and integrating to generate an integrated visual decision support report.

[0017] Through the technical leap from static early warning to dynamic mechanism interpretation, the ultimate goal of improving the accuracy of clinical decision-making and the safety of patient transfer is achieved.

[0018] Step S1 of the present application is the perception and fusion of multi-source heterogeneous data. Specifically, the primary task of this step is to build a unified data view that can comprehensively and accurately reflect the real-time physiological state of the patient, providing a high-quality, standardized data basis for subsequent risk prediction and mechanism analysis. To achieve this goal, the system collects key patient data from the heterogeneous environment of the intensive care unit through three special data interface modules.

[0019] First, the vital sign data acquisition module connects directly to the data output port of the bedside monitor through a device interface that conforms to the IEEE 11073 specification. This module continuously reads and parses the digital signal stream output by the monitor at a frequency of once per second, extracting continuous heart rate values, blood oxygen saturation values, arterial systolic and diastolic pressure values, respiratory rate values, and body temperature values. These values constitute a high-frequency time series data stream reflecting the patient's basic life status.

[0020] Second, the laboratory test data acquisition module interacts with the hospital's laboratory information system through the HL7 protocol. After receiving a query instruction from the system or a push notification from the LIS system, this module will obtain and parse specific observation result messages. The system fixedly collects test items including white blood cell count, C-reactive protein concentration, arterial blood lactic acid concentration, platelet count, creatinine concentration and procalcitonin concentration. Each test result is accurately bound to the sampling timestamp in the laboratory report.

[0021] Finally, the clinical event and assessment data collection module extracts relevant information from the electronic medical record system through the FHIR standard interface. This module queries and obtains the patient's Glasgow coma score, pain numerical rating scale score, vasoactive drug use flag (yes / no), mechanical ventilation support mode, and daily fluid balance value at regular intervals. These data record the clinical assessments and interventions led by medical staff.

[0022] After the collection of the above three types of data is completed, the system starts a standardized data preprocessing process. First, the system applies quality control algorithms based on rules and signal quality indices to real-time check the vital sign data stream, identify and eliminate abnormal data points caused by motion artifacts, pipeline flushing, and other factors. For laboratory data and clinical assessment data, range verification logic is used to mark abnormal values that are obviously beyond the physiological possibility. Subsequently, the system aligns all vital sign data, laboratory data, and clinical event data to a unified time axis based on Unix timestamps. For non-continuous data, the system uses a time window-based forward filling method for interpolation to ensure that each time point has a complete set of parameters. Finally, the system uses the min-max normalization algorithm to map the value of each clinical parameter to the interval [0, 1] to eliminate the influence of different dimensions on subsequent model calculations.

[0023] After the above complete perception, cleaning, alignment, and normalization, all multi-source heterogeneous data are integrated into a structured patient state time series matrix. Each row of the matrix represents a specific minute-level time point, and each column corresponds to a specific processed clinical parameter. This complete data package containing heart rate, oxygen saturation, systolic blood pressure, diastolic blood pressure, respiratory rate, body temperature, white blood cell count, C-reactive protein, blood lactate, platelet count, creatinine, procalcitonin, Glasgow coma score, pain score, vasoactive drug use flag, mechanical ventilation mode, and fluid balance value will serve as the core carrier representing the patient's comprehensive state and be passed on to the subsequent step S2 for deep analysis and risk prediction.

[0024] Step S2 of the present application is the initial prediction of return risk based on ensemble learning. Specifically, the core task of this step is to perform deep feature engineering and model calculation on the patient state time series matrix generated in step S1, thereby generating quantitative risk prediction results and their explanatory basis. The system first starts a multi-level feature extraction process that processes different types of data in the time series matrix. For vital sign parameters, the system uses a sliding window statistical method to calculate the mean, standard deviation, coefficient of variation, and trend slope within each predetermined time window, thereby converting the high-frequency original sequence into statistical features reflecting its stability and change direction. The trend slope is calculated using the least squares method, and its calculation formula is:

[0025]

[0026] For laboratory test parameters, the system calculates the slope of change in the last 24 hours in addition to recording the latest value, to capture the dynamic evolution trend. For clinical assessment parameters, the system mainly extracts the latest state and the proportion of abnormal state duration within the observation period.

[0027] For laboratory test parameters, the system calculates the slope of change in the last 24 hours in addition to recording the latest value, to capture the dynamic evolution trend. For clinical assessment parameters, the system mainly extracts the latest state and the proportion of abnormal state duration within the observation period.

[0028] After completing feature extraction, the system combines all generated features with the latest values of the original parameters to construct a high-dimensional patient comprehensive feature vector. This vector contains hundreds of carefully designed feature dimensions such as mean heart rate, standard deviation of blood oxygen saturation, respiratory rate coefficient of variation, white blood cell count change slope, latest Glasgow coma score, etc. This comprehensive feature vector completely represents the patient's physiological state stability, trend, and severity in the near future. Subsequently, the system inputs the patient comprehensive feature vector into a pre-trained gradient boosting tree model. The model is trained using historical ICU patient data, which contains a large number of patient cases with known outcomes after discharge and their corresponding comprehensive feature vectors. The model outputs an initial return risk score between 0 and 1 by integrating the prediction results of multiple decision trees, which directly represents the predicted probability of the current patient needing to return to the ICU within 48 hours after discharge from the ICU.

[0029] While generating the risk score, the system performs model interpretability analysis. By calculating the average contribution of each feature in the model prediction process, i.e., using the SHAP value analysis method, the system accurately quantifies the positive or negative impact of each feature on the final risk score. This analysis generates a feature importance ranking list, which explicitly lists the top N clinical parameters that contribute most to this high-risk prediction and their corresponding contribution weight. Thus, this step receives the patient state time series matrix from step S1, and through feature engineering and model calculation, it finally generates the initial return risk score and the key risk factor set as two core outputs. These two results together form the basis for subsequent deep mechanism analysis and are completely passed on to the subsequent step S3 for constructing a dynamic risk association graph.

[0030] Referring to Figure 2As shown, step S3 of the present application is to construct a dynamic risk correlation graph. Specifically, the core task of this step is to break through the limitation of isolated static interpretation of risk factors in the prior art, to construct a visual network graph that can intuitively show the risk evolution path by quantitatively analyzing the time sequence causal relationship between key risk factors. This step receives the key risk factor set from step S2 and the patient state time sequence matrix from step S1, and based on these two inputs, a deep analysis is carried out.

[0031] The system first extracts the data columns corresponding to the key risk factor set from the patient state time sequence matrix to form a simplified key factor time sequence sub-matrix. This sub-matrix only contains those clinical parameters identified by the model as having a significant contribution to this high-risk prediction, but it retains the complete and high-resolution time series information of these parameters, providing a data basis for subsequent dynamic correlation analysis.

[0032] After obtaining the key factor time sequence data, the system starts a two-stage correlation relationship calculation process. In the first stage, a linear correlation analysis method is used to calculate the Pearson correlation coefficient between each pair of key factors to determine the strength of their static covariation relationship. In the second stage, a time series-based Granger causality test method is used to analyze the lead-lag relationship of these key factors in time sequence. The core of Granger causality test is to determine whether the historical information of one factor helps to predict the future value of another factor. Its calculation formula is based on the vector autoregressive model, which specifically compares the residual sum of squares of two regression models to achieve:

[0033] This statistic follows an F distribution, and the system uses the p value obtained to determine the statistical significance of the causal relationship.

[0034] After completing the correlation calculation of all factor pairs, the system begins to construct the dynamic risk correlation graph. This graph takes the key risk factors as nodes, and the size of the nodes is proportional to the SHAP contribution weight obtained in step S2. The construction rule of the edges in the graph is as follows: only when both factors satisfy significant correlation (Pearson correlation coefficient p value less than 0.05) and significant Granger causality (Granger test p value less than 0.05) at the same time, a directed edge is established between the corresponding nodes. The direction of the edge is determined by the direction of the Granger causality, i.e. from the Granger cause to the Granger result, and the thickness of the edge is proportional to the F statistic of the Granger causality, which intuitively reflects the strength of the causal relationship.

[0035] Thus, this step, through temporal causal relationship mining of key risk factors, transforms the static set of key risk factors output in step S2 into a dynamic risk correlation graph containing node attributes, directed edges, and weight attributes. This correlation graph not only shows which factors are important, but more importantly, clearly reveals the dynamic transmission mechanism of "who changes first and who affects whom" among these important factors. This structured graph object, as the final output of this step, is fully passed to the subsequent step S4 for the identification and extraction of high-risk evolution paths.

[0036] See Figure 3 As shown, step S4 of this invention involves identifying and outputting high-risk evolution paths. Specifically, the core task of this step is to perform deep topological analysis on the dynamic risk association graph generated in step S3, extracting the most clinically significant risk transmission paths and transforming complex network relationships into clear and actionable decision insights. This step receives the complete dynamic risk association graph from step S3, which includes node attributes, directed edges, and weight attributes, forming the basic data structure for path analysis.

[0037] The system first performs graph preprocessing and optimization. Considering the simplicity and effectiveness of clinical interpretation, the system detects and marks any loop structures in the graph. When a loop is found, the system retains the edge with the highest weight and removes the edges with lower weights, thus transforming the original graph into a directed acyclic graph (DAG). This process ensures the unidirectional transmission characteristics of risk paths, better aligning with the logic of clinical disease progression, and also removes structural obstacles for subsequent critical path algorithms.

[0038] After optimizing the graph structure, the system initiates the critical path discovery process. This process, based on dynamic programming, calculates the sum of the weights of all possible paths from each source node (a node with an in-degree of 0) to the sink node (a node with an out-degree of 0). The formula for calculating the path weights is:

[0039] This formula takes into account the multiplicative effect of the strength of all causal relationships in the path, and provides appropriate compensation for longer paths through the square root term of the path length, thus avoiding the algorithm's tendency to overly favor shorter paths.

[0040] The system then sorts all candidate paths in descending order according to their calculated path weights and selects the top K paths with the highest weights as candidate high-risk paths. For each candidate path, the system further calculates its clinical significance index, which comprehensively considers the path's weight ranking, the average SHAP value of the nodes in the path, and the clinical importance of the physiological systems involved in the path. Through weighted evaluation of these dimensions, the system finally determines the set of high-risk evolutionary paths to be output.

[0041] After identifying the final high-risk pathways, the system generates a corresponding semantic description for each pathway. This description clearly indicates the starting node, intermediate transmission links, and ending node of the pathway, and uses standardized clinical terminology to describe its underlying pathophysiological mechanism.

[0042] This step transforms the complex network structure output from step S3 into a set of high-risk evolutionary paths with clear clinical significance through deep topological analysis of the dynamic risk association graph. This path set not only quantifies the importance ranking of different risk transmission mechanisms, but more importantly, provides clinicians with a clear sequence of intervention targets. This structured path description set, as the final output of this step, is fully passed to the subsequent step S5 to generate an integrated and visualized decision support report.

[0043] Step S5 of this invention is to generate an integrated, visualized decision support report. Specifically, this step is the final stage of the workflow of this invention, and its core task is to systematically integrate and clinically contextualize the key results generated by all the preceding steps to generate a comprehensive report that can provide ICU physicians with all-round, multi-level decision support. This step receives the initial relapse risk score and key risk factor set from step S2, the dynamic risk correlation diagram from step S3, and the set of high-risk evolution paths from step S4, and uses these as the four core data pillars for constructing the report.

[0044] The system first activates the report generation engine, which organizes content based on a pre-defined template structure that aligns with clinical reasoning. The engine first creates a risk overview panel that prominently displays the initial regression risk score and maps it to an intuitive risk level (e.g., low, medium, high), along with a concise clinical interpretation. Next, the system embeds a set of key risk factors into the report as an ordered list. Each factor in the list includes its quantified contribution from step S2, providing physicians with a direct snapshot of the risk factors.

[0045] After presenting the basic risk information, the report generation engine invokes the graphics rendering module to visualize the dynamic risk association graph passed in step S3. This module transforms the nodes (risk factors) and directed edges (causal relationships) in the graph into graphical elements conforming to medical atlas specifications. Node size is proportional to contribution, and the arrows and thickness of the edges intuitively reflect the direction and strength of causality, generating a visual association graph that doctors can interactively explore. Next, the system links and annotates the set of high-risk evolution paths output in step S4 with the graph, clearly marking these key paths in the graph with highlighted colors. The semantic description of each path is presented side-by-side in clear text boxes, thus transforming abstract topological relationships into concrete pathophysiological narratives.

[0046] Furthermore, the system creates a clinical context and raw data validation area at the bottom of the report. This area extracts raw data related to core risk factors and critical pathways from the patient status time-series matrix in step S1, generating trend curves for key vital signs and laboratory indicators over the past 24 hours. This allows physicians to make decisions not only based on the model's abstract analysis but also to perform intuitive validation by going back to the raw data level, further enhancing the credibility of the system's conclusions.

[0047] Thus, through the systematic integration and presentation described above, this step transforms the scattered, technical intermediate results into a clearly structured, insightful, and highly actionable integrated visual decision support report. This report represents a fundamental leap from traditional "single risk warnings" to modern "multi-dimensional mechanism analysis and decision support," ultimately achieving the fundamental goals of improving the accuracy of ICU patient transfer decisions, ensuring patient safety, and optimizing the use of medical resources.

[0048] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An ICU reentry risk prediction method based on patient state awareness, characterized in that, The method comprises the following steps: S1, collecting multi-source heterogeneous patient data, cleaning, aligning and normalizing the patient data, and generating patient state time series matrix; S2, extracting features from the patient state time series matrix to construct a patient comprehensive feature vector, inputting the patient comprehensive feature vector into a prediction model to obtain an initial return risk score, and generating a key risk factor set based on model interpretability analysis; S3, extracting corresponding data from the patient state time series matrix based on the key risk factor set to form a key factor time series sub-matrix, performing two-stage correlation analysis on the key factor time series sub-matrix, including calculating the Pearson correlation coefficient between factors and performing Granger causality test, constructing a dynamic risk correlation graph based on the results of the correlation analysis, wherein the nodes represent risk factors, and the directed edges represent causal relationships; S4, topologically optimizing the dynamic risk correlation graph to obtain a directed acyclic graph, calculating the weights of each path in the directed acyclic graph, identifying candidate high-risk paths based on path weights, and determining a high-risk evolution path set by evaluating the clinical significance index of the candidate high-risk paths; S5, receiving the initial return risk score, the key risk factor set, the dynamic risk correlation graph and the high-risk evolution path set, and integrating to generate an integrated visual decision support report.

2. The ICU return risk prediction method based on patient state perception according to claim 1, wherein: the multi-source heterogeneous patient data includes vital sign data collected from bedside monitors, test item data collected from laboratory information systems, and clinical assessment data collected from electronic medical record systems.

3. The ICU return risk prediction method based on patient state perception according to claim 1, wherein: the cleaning, alignment and normalization of the patient data includes quality control of vital sign data based on signal quality index, time series alignment based on timestamp, and elimination of the influence of different dimensions by using normalization algorithm.

4. The ICU return risk prediction method based on patient state perception according to claim 1, wherein: the feature extraction includes calculating statistical features of vital sign parameters using sliding window statistical method, calculating the change slope of laboratory test parameters, and extracting state features of clinical assessment parameters.

5. The ICU return risk prediction method based on patient state perception according to claim 1, wherein: the prediction model uses gradient boosting tree model, and the model interpretability analysis uses SHAP value analysis method to quantify the contribution of each feature.

6. The ICU return risk prediction method based on patient state perception according to claim 1, wherein: the two-stage correlation analysis includes first calculating the Pearson correlation coefficient to determine the covariation relationship, and then performing Granger causality test to analyze the lead-lag relationship in time series.

7. The ICU return risk prediction method based on patient state perception according to claim 1, wherein: ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ The node size in the dynamic risk correlation graph is proportional to the contribution weight of the corresponding factor, the direction of the edge is determined by the Granger causality relationship, and the thickness of the edge is proportional to the causality strength.

8. The ICU reentry risk prediction method based on patient state perception according to claim 1, wherein: the topology optimization comprises detecting and eliminating a loop structure, and converting the original correlation graph into a directed acyclic graph.

9. The ICU reentry risk prediction method based on patient state perception according to claim 1, wherein: the path weight calculation adopts a formula of multiplying all edge weights in a connected path by the square root of the path length, and the clinical significance index comprehensively considers the path weight ranking, the average contribution of the node, and the importance of the involved physiological system.

10. The ICU reentry risk prediction method based on patient state perception according to claim 1, wherein: the integrated visual decision support report comprises a risk overview panel, a key risk factor list, a visual correlation graph, a high-risk evolution path description, and an original data verification area. ​ ​ ​ ​ ​ ​

Images