Cardiac intensive care monitoring method and system based on spatio-temporal graph neural network

Abstract

The application provides a kind of cardiac intensive care monitoring method and system based on space-time diagram neural network.The method comprises: signal quality evaluation, time alignment and individual baseline normalization are carried out on multi-source cardiac intensive care monitoring data, and uniform space-time monitoring sequence is generated;individualized cardiac intensive care space-time diagram is constructed based on uniform space-time monitoring sequence; individualized cardiac intensive care space-time diagram is input into space-time diagram neural network, message propagation and state update are carried out across time window for graph node and dynamic associated edge, and node dynamic state representation, edge dynamic coupling representation and global intensive care state representation are generated; cardiac intensive care monitoring result is generated according to global intensive care state representation, and risk evolution chain and key evidence subgraph corresponding to cardiac intensive care monitoring result are extracted; individual baseline, dynamic associated edge and monitoring result calibration parameter are updated according to posterior feedback data.The application can improve multi-source monitoring fusion precision, enhance space-time correlation identification ability and improve early warning traceability.

Description

Technical Field

[0001] This application relates to the fields of medical monitoring and artificial intelligence technology, and in particular to a method and system for monitoring critical cardiac conditions based on spatiotemporal graph neural networks. Background Technology

[0002] Cardiac critical care monitoring is a crucial technical aspect of intensive care, emergency transport, and postoperative management. It typically involves the continuous collection and comprehensive analysis of data on the target patient's electrocardiogram, blood pressure, blood oxygen saturation, respiration, body temperature, laboratory indicators, and clinical events to promptly identify risk states such as circulatory dysfunction, insufficient perfusion, and rhythm disturbances. With the development of bedside monitoring equipment and clinical information systems, the ability to acquire multi-source monitoring data has continuously improved, and artificial intelligence-based risk prediction methods are increasingly being applied to critical condition identification.

[0003] In existing technologies, cardiac critical care monitoring often employs fixed threshold alarms, single-indicator trend analysis, or conventional time series models for risk assessment. Some solutions also input multi-source monitoring data into deep learning models to generate risk scores. While these methods can improve the level of automation to some extent, they typically involve simple splicing or static correlation of data from different sources, making it difficult to reflect the time-varying coupling and propagation relationships between cardiac electrical activity, circulatory pressure, peripheral perfusion, oxygenation status, and clinical events.

[0004] Furthermore, critical care monitoring data suffers from issues such as inconsistent sampling frequencies, signal artifacts, missing data, and changes in equipment status. Existing methods do not adequately consider data quality and individual baseline variations, easily leading to false alarms or missed alarms. Simultaneously, current risk prediction results are mostly presented as single risk scores, lacking traceable representations of anomalies, correlation paths, and evidence sources, making it difficult for healthcare professionals to determine the risk formation process and the reliability of alarms. Therefore, there is an urgent need for a cardiac critical care monitoring method that can integrate multi-source monitoring data, characterize spatiotemporal correlation evolution, and output interpretable monitoring results. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and system for monitoring cardiac intensive care based on spatiotemporal graph neural networks, in order to solve the problems of insufficient multi-source monitoring fusion, weak spatiotemporal correlation characterization, and difficulty in tracing early warning results in the prior art.

[0006] A first aspect of this application provides a method for cardiac critical care monitoring based on a spatiotemporal graph neural network, comprising: acquiring multi-source cardiac critical care monitoring data of a target patient; performing signal quality evaluation, time alignment, and individual baseline normalization on the multi-source cardiac critical care monitoring data to generate a unified spatiotemporal monitoring sequence; constructing an individualized cardiac critical care spatiotemporal graph based on the unified spatiotemporal monitoring sequence, the individualized cardiac critical care spatiotemporal graph including graph nodes for representing monitoring channels, physiological states, event contexts, and individual baselines, and dynamically associated edges generated based on physiological prior relationships, temporal propagation relationships, and data quality credibility; inputting the individualized cardiac critical care spatiotemporal graph into a spatiotemporal graph neural network, performing cross-time window message propagation and state updates on the graph nodes and dynamically associated edges to generate node dynamic state representations, edge dynamic coupling representations, and global critical care state representations; generating cardiac critical care monitoring results based on the global critical care state representations, and extracting risk evolution chains and key evidence subgraphs corresponding to the cardiac critical care monitoring results based on the node dynamic state representations and edge dynamic coupling representations; and updating individual baselines, dynamically associated edges, and monitoring result calibration parameters based on posterior feedback data.

[0007] A second aspect of this application provides a cardiac critical care monitoring system based on a spatiotemporal graph neural network, comprising: an acquisition module for acquiring multi-source cardiac critical care monitoring data of a target patient, performing signal quality evaluation, time alignment, and individual baseline normalization on the multi-source cardiac critical care monitoring data, and generating a unified spatiotemporal monitoring sequence; and a construction module for constructing an individualized cardiac critical care spatiotemporal graph based on the unified spatiotemporal monitoring sequence, wherein the individualized cardiac critical care spatiotemporal graph includes graph nodes for representing monitoring channels, physiological states, event contexts, and individual baselines, and is based on physiological prior relationships, temporal propagation relationships, and data quality reliability. The system comprises the following modules: a dynamic linking edge generation module, which inputs the individualized cardiac critical care spatiotemporal graph into the spatiotemporal graph neural network, performs cross-time window message propagation and state updates on graph nodes and dynamic linking edges, and generates node dynamic state representations, edge dynamic coupling representations, and global critical care state representations; an extraction module, which generates cardiac critical care monitoring results based on the global critical care state representations, and extracts the risk evolution chain and key evidence subgraphs corresponding to the cardiac critical care monitoring results based on the node dynamic state representations and edge dynamic coupling representations; and an update module, which updates the individual baseline, dynamic linking edges, and monitoring result calibration parameters based on posterior feedback data.

[0008] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By acquiring multi-source cardiac critical care monitoring data from target patients, this application performs signal quality evaluation, time alignment, and individual baseline normalization on the multi-source cardiac critical care monitoring data to generate a unified spatiotemporal monitoring sequence. Based on the unified spatiotemporal monitoring sequence, an individualized cardiac critical care spatiotemporal graph is constructed. This individualized spatiotemporal graph includes graph nodes representing monitoring channels, physiological states, event contexts, and individual baselines, as well as dynamically related edges generated based on physiological prior relationships, temporal propagation relationships, and data quality reliability. The individualized cardiac critical care spatiotemporal graph is input into a spatiotemporal graph neural network, and cross-time window message propagation and state updates are performed on the graph nodes and dynamically related edges to generate node dynamic state representations, edge dynamic coupling representations, and a global critical care state representation. Cardiac critical care monitoring results are generated based on the global critical care state representation, and risk evolution chains and key evidence subgraphs corresponding to the cardiac critical care monitoring results are extracted based on the node dynamic state representations and edge dynamic coupling representations. Individual baselines, dynamically related edges, and monitoring result calibration parameters are updated based on posterior feedback data. This application can improve the accuracy of multi-source monitoring fusion, enhance spatiotemporal correlation identification capabilities, and improve the traceability of early warnings. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the cardiac critical care monitoring method based on spatiotemporal graph neural networks provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the cardiac critical care monitoring system based on spatiotemporal graph neural network provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0012] In existing technologies, cardiac critical care monitoring typically relies on bedside monitoring equipment and clinical information systems to acquire data such as the patient's electrocardiogram, blood pressure, blood oxygen saturation, respiration, body temperature, laboratory indicators, and clinical events. Patient status is then assessed using fixed threshold alarms, single-indicator trend analysis, or conventional time-series models. While some existing solutions incorporate deep learning models for risk prediction based on multi-source monitoring data, they usually focus on feature stitching or static fusion of data from different sources, outputting a single risk score. This approach fails to adequately represent the time-varying correlations and propagation relationships between various physiological signals.

[0013] However, the development of critical cardiac conditions is usually not caused by a single abnormality, but rather by the continuous coupling and evolution of cardiac electrical activity, circulatory pressure, peripheral perfusion, oxygenation status, and clinical intervention events. Existing methods do not adequately consider the temporal alignment of multi-source data, signal quality reliability, individual baseline variations, and cross-channel spatiotemporal correlations. They are easily affected by differences in sampling frequency, signal artifacts, changes in equipment status, and missing data, leading to insufficient accuracy in monitoring results. Furthermore, existing early warning results lack structured representations of abnormal nodes, correlation paths, and sources of evidence, making it difficult for healthcare professionals to trace and assess the risk formation process and the reliability of alarms.

[0014] To address the aforementioned issues, this application provides a method for cardiac critical care monitoring based on a spatiotemporal graph neural network. This method acquires multi-source cardiac critical care monitoring data from target patients, performs signal quality evaluation, time alignment, and individual baseline normalization on the multi-source cardiac critical care monitoring data, generating a unified spatiotemporal monitoring sequence. Based on the unified spatiotemporal monitoring sequence, an individualized cardiac critical care spatiotemporal graph is constructed. Graph nodes represent monitoring channels, physiological states, event contexts, and individual baselines, while dynamic edges represent physiological prior relationships, temporal propagation relationships, and data quality reliability. The individualized cardiac critical care spatiotemporal graph is then input into a spatiotemporal graph neural network, where cross-time window message propagation and state updates are performed on graph nodes and dynamic edges, generating dynamic state representations of nodes, dynamic coupling representations of edges, and a global critical care state representation.

[0015] Furthermore, this application generates cardiac critical care monitoring results based on the global critical care status representation, and extracts risk evolution chains and key evidence subgraphs based on node dynamic status representation and edge dynamic coupling representation. This ensures that the monitoring results not only include risk level, risk trend, and risk confidence, but also reflect the propagation path of risk from early state deviation to the formation of a critical state. Simultaneously, this application updates individual baselines, dynamic correlation edges, and monitoring result calibration parameters based on posterior feedback data, enabling the monitoring model to adapt to changes in the target patient's status, equipment status, and feedback confirmation results.

[0016] Through the above technical solutions, this application can improve the fusion accuracy of multi-source monitoring data, enhance the ability to identify the spatiotemporal correlation of cardiac critical conditions, reduce the risk of false alarms caused by signal artifacts and individual differences, and improve the traceability and reliability of cardiac critical monitoring results.

[0017] The technical solution of this application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] Figure 1 This is a schematic flowchart of the cardiac critical care monitoring method based on spatiotemporal graph neural networks provided in an embodiment of this application. Figure 1 As shown, the method may specifically include: S101: Acquire multi-source cardiac critical care monitoring data of the target patient, evaluate the signal quality, time alignment and individual baseline normalization of the multi-source cardiac critical care monitoring data, and generate a unified spatiotemporal monitoring sequence. S102, Construct an individualized spatiotemporal graph of cardiac critical care based on a unified spatiotemporal monitoring sequence. The individualized spatiotemporal graph of cardiac critical care includes graph nodes used to characterize monitoring channels, physiological states, event contexts and individual baselines, as well as dynamic association edges generated based on physiological prior relationships, temporal propagation relationships and data quality credibility. S103: Input the individualized cardiac critical care spatiotemporal graph into the spatiotemporal graph neural network, perform cross-time window message propagation and state update on graph nodes and dynamically associated edges, and generate node dynamic state representation, edge dynamic coupling representation and global critical care state representation. S104, Generate cardiac critical care monitoring results based on global critical care status representation, and extract risk evolution chains and key evidence subgraphs corresponding to cardiac critical care monitoring results based on node dynamic state representation and edge dynamic coupling representation. S105 updates the individual baseline, dynamic correlation edge, and monitoring result calibration parameters based on posterior feedback data.

[0019] In some embodiments, multi-source cardiac critical care monitoring data of the target patient are acquired, and signal quality evaluation, time alignment, and individual baseline normalization are performed on the multi-source cardiac critical care monitoring data to generate a unified spatiotemporal monitoring sequence, including: Access bedside continuous monitoring data and clinical event data according to the target patient's monitoring object identification, and establish original monitoring data packages based on data source, collection time and equipment status; Quality features are extracted from the original monitoring data packets. Quality evaluation results are generated based on waveform stability, temporal integrity, and cross-source consistency. Low-confidence monitoring segments are marked based on the quality evaluation results. A unified time index is established based on the sampling interval and event occurrence time of different data sources. The original monitoring data packets are mapped to the unified time index to generate a time-series monitoring representation with missing state codes and quality weights. Individual baseline parameters are generated based on stable monitoring segments of the target patient, and the deviation of the temporal monitoring representation is normalized using the individual baseline parameters to generate a unified spatiotemporal monitoring sequence.

[0020] Specifically, after a target patient enters the cardiac intensive care unit, the monitoring platform assigns a monitoring target identifier to the patient. For example, it combines the patient's hospital number, bed number, monitoring device number, and admission time to generate a unique monitoring target identifier, which is then used as an index for multi-source data aggregation. Bedside continuous monitoring data may include electrocardiogram waveforms, arterial blood pressure waveforms, pulse oximetry waveforms, respiratory waveforms, and body temperature trends. Clinical event data may include medication adjustments, fluid infusions, changes in body position, blood tests, device disconnections, and sensor replacements.

[0021] The platform records the data source, acquisition time, device status, sampling frequency, and data format upon data access, and encapsulates data from the same target patient within the same monitoring cycle into a raw monitoring data package. For example, if the ECG monitor acquires ECG waveforms at 250 Hz, the arterial pressure module acquires blood pressure waveforms at 125 Hz, the nursing terminal records a vasopressor dosage adjustment event at 09:12, and the laboratory system transmits lactate results at 09:40, the platform writes all of the above data into the same raw monitoring data package.

[0022] After generating the raw monitoring data packets, the platform performs quality feature extraction on each data source. For ECG waveforms, it calculates baseline drift amplitude, signal loss ratio, spike noise density, and cardiac cycle stability; for arterial blood pressure waveforms, it calculates waveform continuity, pulse pressure abnormal jump amplitude, systolic peak identification stability, and waveform area abrupt change ratio; for pulse oximetry waveforms, it calculates perfusion amplitude stability, peak interval consistency, and low perfusion interference markers; for clinical event data, it calculates event timestamp integrity, event source consistency, and event recording delay.

[0023] The platform generates quality evaluation results based on the aforementioned quality characteristics and assigns a quality weight to each time segment. When a segment experiences ECG lead loss, excessive blood pressure waveform damping, continuous loss of pulse oximetry waveform, or abnormal event time recording, the platform marks the corresponding segment as a low-confidence monitoring segment. For example, if the ECG signal loss rate reaches 35% between 09:18 and 09:20, and the device status indicates lead loss during the same period, the platform marks this time segment as a low-confidence ECG segment and reduces the weight of this data source during subsequent feature fusion.

[0024] Subsequently, the platform establishes a unified time index based on the sampling intervals and event occurrence times of different data sources. The unified time index can employ a multi-scale window structure, such as using a 1-second waveform base window, a 1-minute vital sign trend window, and a 15-minute clinical event summary window, maintaining temporal correspondence between different scales through window mapping relationships. The platform maps continuous waveforms, periodic vital signs, test results, and clinical events from the original monitoring data packets to the unified time index.

[0025] In some examples, for high-frequency waveforms, the platform extracts morphological and stability features within a basic window; for low-frequency test results, the platform performs time-based mapping based on the result feedback time and the effective observation period; for clinical events, the platform generates event location codes and event duration codes based on the occurrence time. If a window does not receive valid data, the platform does not directly replace it with a zero value, but instead generates a missing state code to characterize the source of the missing data, the duration of the missing data, and the interval between the most recent valid data. Simultaneously, the platform converts the quality assessment results into quality weights and binds them to the corresponding windows, generating a time-series monitoring representation with missing state codes and quality weights.

[0026] Furthermore, after completing the time-series mapping, the platform generates individual baseline parameters based on stable monitoring segments for the target patient. Stable monitoring segments can be determined from the initial period when vital signs are relatively stable after admission, stable periods confirmed by medical staff, or effective monitoring segments from the same historical disease stage. The platform calculates the baseline center, fluctuation range, short-term change slope, and baseline reliability for key monitoring variables in the stable monitoring segments.

[0027] For example, if a target patient's average heart rate was 86 beats per minute, mean arterial pressure was 78 mmHg, and blood oxygen saturation remained stable at around 96% between 08:30 and 09:00, and the corresponding data quality weights were all higher than 0.85, then the platform would include this time period in the baseline calculation. For subsequent time-series monitoring, the platform uses individual baseline parameters to calculate the deviation of each window, converting the original monitoring values ​​into relative baseline changes, trend offsets, and reliable normalized values, generating a unified spatiotemporal monitoring sequence.

[0028] Through the above processing, multi-source cardiac critical care monitoring data can be reliably aggregated, quality screened, time-aligned, and individualized normalized under a unified monitoring object identifier. This preserves the sampling differences and missing states of different data sources while reducing the interference of signal artifacts, equipment malfunctions, and individual differences on subsequent mapping and risk identification, thereby improving the accuracy of the unified spatiotemporal monitoring sequence in representing the true cardiac critical care status of the target patients.

[0029] In some embodiments, a personalized spatiotemporal map of cardiac critical care is constructed based on a unified spatiotemporal monitoring sequence, including: Based on the monitoring source, time index, and baseline deviation status in the unified spatiotemporal monitoring sequence, generate graph node attributes corresponding to the target patient; Based on knowledge of cardiac critical care monitoring and the temporal status changes of target patients, candidate associations between graph nodes are determined. By comprehensively considering the physiological coupling state, temporal lag state, event response state, and quality reliability state of candidate associations, dynamic association edges and corresponding edge weight parameters are generated. By associating graph nodes, dynamically related edges, and corresponding time windows, an individualized spatiotemporal graph of cardiac critical illness is formed to characterize the evolution of the target patient's cardiac critical illness status.

[0030] Specifically, after generating a unified spatiotemporal monitoring sequence, the monitoring platform reads the monitoring source, time index, quality weight, and baseline deviation status within each time window according to the monitoring object identifier of the target patient, and establishes a graph node set for the target patient. Graph nodes can be abstracted according to the functional meaning of monitoring data in the state of cardiac critical illness. Instead of directly using a single raw indicator as an isolated node, it combines features with common physiological significance under the same monitoring source into node attributes.

[0031] For example, within the 09:00 to 09:01 time window, the node attributes corresponding to ECG sources include heart rate deviation, rhythm stability, decrease in heart rate variability, and quality weight; the node attributes corresponding to arterial blood pressure sources include mean arterial pressure deviation, pulse pressure change, waveform ascending limb slope, and damping reliability; the node attributes corresponding to pulse oximetry sources include oxygenation deviation, perfusion amplitude change, and waveform integrity; and the node attributes corresponding to clinical events include vasopressor adjustment flag, fluid input status, and event occurrence time code. The platform binds these node attributes to the corresponding time index to form the target patient graph node within the current time window.

[0032] After identifying the graph nodes, the platform invokes the cardiac critical care monitoring knowledge base and the target patient's historical time-series status changes to determine candidate associations between the graph nodes. The cardiac critical care monitoring knowledge base is pre-configured with the physiological influence directions and common propagation delays between cardiac electrical activity, circulatory pressure, peripheral perfusion, oxygenation status, and clinical events. For example, a sustained increase in heart rate may be associated with a decrease in circulatory pressure; a slowing of the ascending limb of the arterial blood pressure waveform may be associated with weakened peripheral perfusion; and vasopressor adjustment events may affect mean arterial pressure and pulse amplitude within several subsequent time windows.

[0033] For example, the platform simultaneously calculates the lag correlation, common deviation trend, and event response consistency between different graph nodes based on the time-series changes of the target patient over the past 30 minutes, thereby screening out candidate associations. If the target patient experiences an increase in vasopressor dosage at 09:05, and the mean arterial pressure gradually recovers from 09:07 to 09:12, the platform will establish a candidate association between the clinical event node and the circulatory pressure node, and record the event impact delay.

[0034] For each candidate association, the platform further calculates the physiological coupling state, temporal lag state, event response state, and quality reliability state. The physiological coupling state is determined based on the prior association strength in the cardiac critical care monitoring knowledge base; the temporal lag state is determined based on whether the changes in node attributes within the preceding and following time windows have a stable sequential relationship; the event response state is determined based on whether the relevant nodes show the expected changes after the occurrence of a clinical event; and the quality reliability state is determined based on the quality weights of the nodes at both ends of the candidate association graph and the marking of low-reliability segments.

[0035] The platform inputs the above states into the edge weight generation model to obtain dynamically associated edges and their corresponding edge weight parameters. Edge weight parameters may include the current coupling strength, propagation direction, effective time range, and confidence coefficient. If the arterial blood pressure node has a high quality weight, the pulse oximetry node is continuous and intact, and the baseline deviation trend of both increases synchronously within multiple time windows, the platform increases the edge weight between the circulatory pressure node and the peripheral perfusion node. If the ECG signal is marked as a low-confidence segment due to lead disconnection, the platform reduces the edge weight contribution of the ECG node to propagation to other nodes.

[0036] After generating the dynamic correlation edges, the platform organizes the graph nodes, dynamic correlation edges, and time indexes within each time window to form a personalized spatiotemporal graph of cardiac critical care for the target patient. This spatiotemporal graph not only records the spatial correlations between different physiological states within the same time window but also records the evolutionary relationships of node states and edge weight parameters between adjacent time windows. For example, from 09:00 to 09:15, the platform sequentially records changes in state such as increased heart rate deviation, decreased mean arterial pressure, weakened pulse wave amplitude, intervention of vasopressor events, and subsequent pressure recovery, and expresses the propagation and response process of risk states between different nodes through dynamic correlation edges. Thus, the personalized spatiotemporal graph of cardiac critical care can continuously characterize the evolution of the target patient's cardiac critical care status within the current monitoring period.

[0037] In this way, the unified spatiotemporal monitoring sequence is transformed into an individualized spatiotemporal graph of cardiac critical care with node attributes, dynamic associated edges and time window relationships. This allows multi-source monitoring data to no longer participate in subsequent analysis in a static splicing form, but to participate in spatiotemporal graph neural network calculations in a graph structure that conforms to the physiological coupling and temporal propagation laws, thereby improving the accuracy and interpretability of modeling the evolution of cardiac critical care states.

[0038] In some embodiments, cross-time-window message propagation and state updates are performed on graph nodes and dynamically associated edges to generate dynamic state representations of nodes, dynamic coupling representations of edges, and global critical state representations, including: According to the time window corresponding to the individualized cardiac critical care spatiotemporal graph, the node attributes of the graph nodes and the edge weight parameters of the dynamically associated edges are input into the spatiotemporal graph neural network. Graph messages are aggregated between adjacent graph nodes based on edge weight parameters, and the aggregated messages are gating and corrected in conjunction with the data quality reliability status to generate node update representations within the window. Based on the node update representation and edge weight change status between adjacent time windows, perform temporal state transfer and dynamic associated edge re-estimation to generate node dynamic state representation and edge dynamic coupling representation. Cross-scale fusion of node dynamic state representation and edge dynamic coupling representation is performed to generate a global critical illness state representation.

[0039] Specifically, after constructing a personalized spatiotemporal graph for cardiac critical care, the monitoring platform divides the spatiotemporal graph into continuous time windows according to a unified time index. The graph node attributes, edge weights of dynamically associated edges, and data quality reliability within each time window are input into the spatiotemporal graph neural network. For example, the platform uses a 1-minute monitoring window and a 15-minute trend observation window. Within the monitoring period from 09:00 to 09:15, the heart rate deviation, rhythm stability, and quality weights of ECG nodes; the mean arterial pressure deviation, pulse pressure change, and waveform reliability of circulatory pressure nodes; the pulse wave amplitude change and perfusion stability of peripheral perfusion nodes; and the dose change encoding of vasopressor event nodes are all input into the network. Dynamically associated edges carry the current coupling strength, propagation direction, effective delay, and reliability coefficient, used to constrain message propagation between different nodes.

[0040] During the message propagation phase within the window, the spatiotemporal graph neural network first retrieves the set of neighboring graph nodes for the target graph node and calculates the message contribution of each neighboring node based on the edge weight parameters of the dynamically associated edges. For example, within the 09:06 time window, the circulatory pressure node receives messages from the ECG node, the peripheral perfusion node, and the vasopressor event node. If the ECG node is marked as low confidence due to lead loosening, the network generates a gating coefficient based on the data quality confidence status, attenuating the incoming messages from the ECG node to a lower level. If the arterial blood pressure waveform has high continuity and is synchronized with the pulse wave amplitude change, the aggregation weight between the circulatory pressure node and the peripheral perfusion node is increased. The network fuses the gating-corrected aggregated messages with the previous window state representation of the target graph node to obtain the node update representation within the window, thereby avoiding the amplification of errors by low-quality data during graph propagation.

[0041] During the cross-time window state update phase, the network performs temporal state propagation based on the node update representation and edge weight changes between adjacent time windows. For nodes with continuous and stable sampling, the network uses higher state-maintaining parameters to preserve recent trends. For irregular data such as test results and medication events, the network combines time interval encoding and event persistence impact encoding to perform state decay or extension. For example, after a vasopressor dosage increase event occurs at 09:05, the impact of the event node on the circulatory pressure node gradually propagates from 09:06 to 09:12. If the mean arterial pressure subsequently recovers and the pulse wave amplitude improves, the network re-estimates the coupling strength from the event node to the circulatory pressure node and from the circulatory pressure node to the peripheral perfusion node, generating a dynamic edge coupling representation. If the pressure does not improve and perfusion continues to decline, the corresponding edge weights are strengthened, forming an abnormal propagation indication.

[0042] In the cross-scale fusion phase, the platform inputs second-level waveform status, minute-level vital sign trends, and longer-term clinical event status into the fusion layer. The fusion layer assigns fusion weights based on the timeliness, quality weight, and risk contribution of each scale representation. For short-term heart rate fluctuations without sustained deviations in blood pressure and perfusion, the fusion layer reduces the impact of transient waveform abnormalities on the overall critical condition. For cases where decreased blood pressure, weakened perfusion, and slowed oxygenation recovery occur simultaneously within multiple windows, the fusion layer increases the contribution of trend abnormalities, ultimately generating a global critical condition representation.

[0043] Through the above processing, the spatiotemporal graph neural network can identify credible associations between multiple physiological nodes within the same time window, characterize the continuous changes in state and the evolution of coupling strength between adjacent time windows, and form a representation reflecting the overall cardiac critical condition of the target patient through multi-scale fusion, thereby improving the stability of risk identification, reducing misjudgments caused by artifacts, and providing a reliable foundation for subsequent cardiac critical condition monitoring results, risk evolution chains, and key evidence subgraph generation.

[0044] In some embodiments, graph messages are aggregated based on edge weight parameters for adjacent graph nodes, and the aggregated messages are gating and corrected in conjunction with the data quality reliability status to generate an in-window node update representation, including: For the target graph node within the current time window, determine the set of adjacent graph nodes based on the dynamically associated edges, and calculate the message contribution of each adjacent graph node based on the edge weight parameters; The message contribution is weighted and aggregated with the corresponding node attributes to generate the initial aggregated message; Gating coefficients are generated based on the data quality trustworthiness status, and the gating coefficients are used to attenuate and correct the initial aggregated messages formed from low-trust sources. The corrected aggregated message is merged with the historical state representation of the target graph nodes to generate the node update representation within the window.

[0045] Specifically, within the current time window, the spatiotemporal graph neural network first determines the target graph node that needs to be updated. The target graph node can be a circulatory stress node, an electrocardiographic activity node, a peripheral perfusion node, an oxygenation status node, or a clinical event node. Based on the dynamic association edges connected to the target graph node in the individualized cardiac critical care spatiotemporal graph, the platform reads the set of adjacent graph nodes and the edge weight parameters of each dynamic association edge within the current time window. The edge weight parameters include coupling strength, propagation direction, effective delay, and confidence coefficient, which represent the degree of influence of adjacent graph nodes on the target graph node.

[0046] For example, within the 09:06 time window, when the circulatory pressure node is used as the target graph node, adjacent graph nodes may include ECG activity nodes, peripheral perfusion nodes, oxygenation status nodes, and vasopressor event nodes. If the edge weight between the ECG activity node and the circulatory pressure node is 0.42, the edge weight between the peripheral perfusion node and the circulatory pressure node is 0.76, and the edge weight between the vasopressor event node and the circulatory pressure node is 0.68, then the network calculates the message contribution of each adjacent graph node based on the corresponding edge weights.

[0047] During message contribution calculation, the network maps the node attributes of adjacent graph nodes to node message vectors of a unified dimension, and multiplies the node message vectors with the corresponding edge weight parameters to obtain weighted message contributions. For ECG activity nodes, node attributes may include heart rate deviation, rhythm stability, heart rate variability, and ECG quality weights; for peripheral perfusion nodes, node attributes may include pulse amplitude variation, perfusion stability, and waveform integrity; for vasopressor event nodes, node attributes may include dose change direction, event occurrence time, and event duration impact encoding. The network weighted aggregates the message contributions from multiple adjacent graph nodes to generate an initial aggregated message. If the mean arterial pressure decreases significantly between 09:06 and 09:07, the pulse amplitude decreases synchronously, and a vasopressor adjustment event has just occurred, then the initial aggregated message proportion of peripheral perfusion nodes and vasopressor event nodes for circulatory pressure nodes is higher than that of ECG activity nodes.

[0048] Subsequently, the platform generates a gating coefficient based on the data quality reliability status. The data quality reliability status originates from the preceding quality evaluation results and reflects whether there are artifacts, missing data, device disconnections, or cross-source inconsistencies within the current time window. The gating coefficient can be determined jointly based on quality weights, low-reliability segment identifiers, and cross-source consistency. If the data quality weight of a certain adjacent graph node is higher than 0.85 and there are no low-reliability monitoring segments, the gating coefficient is close to 1. If the ECG lead experiences a brief disconnection at 09:06, and the quality weight of the ECG activity node drops to 0.35, the network attenuates the initial aggregation message formed by the ECG activity node to prevent ECG artifacts from being amplified into circulatory risks through graph message propagation. If the arterial blood pressure waveform quality is high and the pulse oximetry waveform is in the same direction as the blood pressure change, the aggregation message between the circulatory pressure node and the peripheral perfusion node is preserved or enhanced.

[0049] Furthermore, after gating correction, the network fuses the corrected aggregated message with the historical state representation of the target graph nodes. The historical state representation records the state change trend of the target graph nodes within the previous time window or multiple historical time windows. For example, if the circulatory pressure node has continuously deviated from its baseline mean arterial pressure by 8 to 12 mmHg between 09:03 and 09:05, and the current corrected aggregated message indicates decreased peripheral perfusion and insufficient response to vasopressor events, the network will increase the proportion of this abnormal trend in the node update representation within the window by gating the circulatory unit or the time attention unit. If the current aggregated message is mainly generated by low-confidence sources, the historical stable state can suppress the impact of anomalous mutations on the node update results.

[0050] Through the above method, adjacent graph node messages within the current time window can be reliably aggregated according to the edge weight parameters of dynamically associated edges. Data quality gating weakens interference from low-reliability sources, and then the messages are fused with the historical state representation of the target graph node to form a node update representation within the window. This process improves the responsiveness of graph message propagation to real physiological coupling changes, reduces the risk of mispropagation caused by noise, missing data, and equipment malfunctions, and provides a more stable node state foundation for subsequent cross-window state transmission and global critical condition identification.

[0051] In some embodiments, based on the node update representation and edge weight change state between adjacent time windows, temporal state propagation and dynamic associated edge reestimation are performed to generate node dynamic state representation and edge dynamic coupling representation, including: Obtain the node update representation and edge weight change status corresponding to the current time window and adjacent historical time windows; Based on time interval encoding and state preservation parameters, the node update representation is propagated in a temporal state to generate cross-window node state representation. Based on the changing relationship between the state representations of cross-window nodes and the corresponding edge weight changes, the coupling strength of dynamically associated edges is re-estimated. By integrating cross-window node state representations and re-estimated coupling strength, dynamic state representations of nodes and dynamic coupling representations of edges are generated.

[0052] Specifically, after completing the node update within the current time window, the monitoring platform reads the node update representations, edge weight change states, and time interval codes corresponding to the current time window and adjacent historical time windows from the spatiotemporal graph cache. For example, when using 09:08 to 09:09 as the current time window, the platform simultaneously reads the update representations of ECG activity nodes, circulatory pressure nodes, peripheral perfusion nodes, oxygenation status nodes, and vasopressor event nodes from 09:05 to 09:08, and reads the edge weight changes of each dynamically related edge within the continuous window. If a vasopressor dosage increase event occurs at 09:05, the mean arterial pressure is still 12 mmHg below the individual baseline at 09:06, the pulse wave amplitude decreases at 09:07, and the blood oxygen recovery time is prolonged at 09:08, then the platform uses these continuous changes as input for the time-series state transmission.

[0053] During the execution of time-series state transmission, the platform describes the data interval, event duration, and monitoring segment continuity between adjacent windows based on time interval encoding, and determines the degree of influence of historical states on the current state by combining state preservation parameters. For waveform nodes with continuous sampling and reliable quality, the state preservation parameter can be set to a higher value, allowing trends such as decreased heart rate variability, decreased mean arterial pressure, and weakened pulse wave amplitude to be expressed continuously across multiple windows. For low-frequency or discrete event nodes such as test results and medication adjustments, the platform generates decaying state preservation parameters based on the event occurrence time and preset influence duration. For example, the vasopressor event node maintains a high influence weight for 3 to 10 minutes after the event occurs, gradually decaying after exceeding the effective response window. The platform fuses the current node update representation with the historical node update representation according to the time interval encoding and state preservation parameters to generate a cross-window node state representation.

[0054] During the dynamic edge reassessment phase, the platform determines whether the coupling strength between different graph nodes needs to be enhanced, maintained, or weakened based on the changing relationships between the state representations of cross-window nodes and the edge weight changes. For example, if the circulatory pressure node continuously deviates from the individual baseline, while the pulse wave amplitude of the peripheral perfusion node decreases synchronously, and the order of their changes remains stable, the platform increases the coupling strength from the circulatory pressure node to the peripheral perfusion node. If the electrocardiogram (ECG) activity node exhibits a short-term abnormality but has a low ECG quality weight, and there are no corresponding changes in blood pressure, perfusion, or oxygenation status, the platform decreases the coupling strength from the ECG activity node to the risk-related node. If the mean arterial pressure recovers and the perfusion amplitude improves after a vasopressor event, the platform adjusts the edge weights from the clinical event node to the circulatory pressure node to a positive response enhancement state. The edge weight reassessment process can be implemented using a gated attention unit, inputting the node state difference, historical edge weights, quality confidence, and event response delay into the edge weight update function, and outputting the reassessed coupling strength and edge state labels.

[0055] Subsequently, the platform integrates cross-window node state representations and re-estimated coupling strength to generate node dynamic state representations and edge dynamic coupling representations. Node dynamic state representations are used to indicate the state trend, anomaly persistence, and dynamic deviation of target graph nodes relative to the individual baseline within a continuous time window; edge dynamic coupling representations are used to indicate the propagation direction, coupling strength changes, and reliable evolution state of dynamically associated edges within a continuous time window.

[0056] For example, if the circulatory pressure node decreases before the peripheral perfusion node decreases during the period from 09:05 to 09:12, and the oxygenation status node subsequently shows a recovery delay, the platform can generate a dynamic coupling characterization of the propagation from circulatory pressure anomalies to perfusion and oxygenation anomalies; if subsequent vasopressor events lead to a pressure rebound, the relevant event response edges are marked as enhanced restorative coupling.

[0057] Through the above processing, the monitoring status between adjacent time windows is no longer analyzed in a fragmented manner, but can form a continuous dynamic representation under the constraints of time interval, state preservation, and edge weight estimation. This embodiment can improve the accuracy of identifying the evolution trend of cardiac critical state, enhance the ability of dynamic association edges to express the real physiological propagation relationship, reduce the interference of short-term artifacts on cross-window judgment, and provide a stable data foundation for subsequent global critical state representation, risk evolution chain, and key evidence subgraph generation.

[0058] In some embodiments, generating cardiac critical care monitoring results based on a global critical care status characterization includes: The global critical illness status representation is input into the risk decoding network to analyze the distribution of the critical illness status of the target patient within the current time window and the predicted time window. Based on the distribution of severe illness status, the characteristics of continuous changes in status, and the reliability of monitoring, risk level, risk trend, and risk confidence level are generated. The risk level, risk trend, and risk confidence level are matched with the preset graded output rules to determine the monitoring prompt level and output trigger status. When the monitoring confidence level is lower than the confidence threshold, the output trigger state is suppressed and corrected based on the corresponding low-confidence monitoring segment to generate cardiac critical care monitoring results.

[0059] Specifically, after obtaining the global critical illness status representation, the monitoring platform inputs this representation into a pre-trained risk decoding network. The risk decoding network includes a current state parsing branch and a predictive state parsing branch. The current state parsing branch is used to identify the distribution of the critical illness status of the target patient within the current time window, and the predictive state parsing branch is used to infer the risk changes in subsequent predictive time windows based on the status evolution trend within consecutive time windows.

[0060] For example, within the 09:10 time window, the global critical condition status representation simultaneously includes information such as a continuous decline in circulatory pressure nodes, increased fluctuations in peripheral perfusion nodes, slower recovery of oxygenation status nodes, and insufficient response to vasopressor events. The risk decoding network maps the above information into probability distributions of multiple critical condition status categories and calculates the status probabilities within the current window, the next 5-minute window, and the next 15-minute window, respectively.

[0061] When analyzing the distribution of severe illness status, the platform does not directly use a single highest probability category as the final result, but rather makes a comprehensive judgment by combining the characteristics of continuous status change and monitoring reliability. The characteristics of continuous status change can be determined by the changes in the probability of risk status within multiple consecutive windows, the duration of deviation from key nodes, and the degree of enhanced coupling of dynamic related edges.

[0062] The reliability of monitoring can be determined by the quality weights of graph nodes participating in the current decoding, the proportion of low-reliability monitoring segments, and cross-source consistency results. For example, if the mean arterial pressure is continuously lower than the individual baseline by more than 10 mmHg between 09:06 and 09:10, the pulse wave amplitude decreases synchronously, the blood oxygen recovery time is prolonged, and the quality weights of both arterial blood pressure and pulse oxygenation data are higher than 0.85, then the platform judges that the risk has high persistence and high reliability. If only ECG nodes show abnormalities during the same period, and the ECG quality weight is 0.32 and is accompanied by lead dislodgement indicators, then the platform reduces the contribution of the abnormal state to the final risk.

[0063] The platform generates risk levels, risk trends, and risk confidence scores based on the distribution of critical illness status, the characteristics of continuous status changes, and monitoring reliability. The risk level indicates the severity of the current cardiac critical illness monitoring result; the risk trend indicates whether the risk increases, stabilizes, or decreases within the prediction time window; and the risk confidence score indicates the degree to which the result is supported by data quality and multi-source consistency. For example, if the current probability of a critical illness status is 0.72, and the probability of risk rises to 0.81 in the next 5 minutes, and there is a stable propagation relationship between key nodes, the platform can generate a higher risk level, an increasing risk trend, and a higher risk confidence score. If the current probability is high but mainly originates from low-reliability electrocardiogram segments, the platform generates a lower risk confidence score and proceeds to the output suppression judgment process.

[0064] Subsequently, the platform matches the risk level, risk trend, and risk confidence level with preset tiered output rules. These tiered output rules can be set based on the monitoring alert level, output duration, repetitive trigger interval, and confirmation conditions. When the risk level is at the "Attention" level and the risk trend is rising, the platform can generate a trend alert on the bedside terminal; when the risk level is at the "Warning" level and the risk confidence level is higher than 0.75, the platform generates a warning trigger status; when the risk level reaches the "High Risk" level, the risk trend continues to rise, and multiple key nodes are abnormal simultaneously, the platform generates a high-priority alarm trigger status. If the risk trend declines or the risk confidence level is insufficient, the platform can adjust the output trigger status to delayed observation or review alerts to avoid instantaneous fluctuations directly triggering strong alarms.

[0065] When the monitoring confidence level is below the confidence threshold, the platform suppresses and corrects the output trigger status based on the corresponding low-confidence monitoring segment. For example, if the risk decoding result at 09:10 shows a high contribution from abnormal heart rate, but the ECG signal from 09:09 to 09:10 shows lead disconnection, spike noise, and cross-source inconsistency, the platform corrects the high-priority alarm trigger status to a device verification prompt, while retaining the continuous observation status of blood pressure, perfusion, and oxygenation nodes. If the low-confidence segment accounts for only a small proportion, and other high-confidence sources all support an increased risk, the platform only reduces the risk confidence level without canceling the warning. The final generated cardiac critical care monitoring results include risk level, risk trend, risk confidence level, monitoring prompt level, and output trigger status.

[0066] Through the above processing, the global critical condition representation can be transformed into cardiac critical care monitoring results with hierarchical meaning and credibility constraints. This can identify risk changes within the current window and prediction window, and can also modify alarm outputs by combining status persistence and data quality. This improves the stability of early warning results, reduces false alarms caused by low-credibility segments, and enhances the reliability of cardiac critical care monitoring results in bedside applications.

[0067] In some embodiments, risk evolution chains and key evidence subgraphs corresponding to cardiac critical care monitoring results are extracted based on node dynamic state representation and edge dynamic coupling representation, including: Determine the target risk status and its corresponding monitoring time range based on the results of cardiac critical care monitoring; Within the monitoring time range, the target graph nodes whose state deviations are continuously enhanced are identified based on the dynamic state representation of the nodes, and the target dynamic associated edges whose coupling contributions meet the preset conditions are identified based on the dynamic coupling representation of the edges. Based on the chronological order and the direction of propagation, path tracing is performed on the target graph nodes and the dynamic edges associated with the target to generate a risk evolution chain; Based on the contribution of each graph node and dynamic associated edge in the risk evolution chain to the target risk state, a key evidence subgraph is selected and formed.

[0068] Specifically, after generating cardiac critical care monitoring results, the monitoring platform first analyzes the risk level, risk trend, risk confidence level, and output trigger status in the results to determine the target risk status that needs to be interpreted. Then, based on the trigger time, prediction time window, and persistent abnormal interval corresponding to the target risk status, the monitoring time range is determined. For example, if the platform generates a high-risk circulatory perfusion risk alert at 09:12, and the risk trend continues to rise within the next 5 minutes, then 09:12 is taken as the risk trigger time. The platform traces back the node status changes from 09:02 to 09:12, and uses 09:12 to 09:17 as the prediction extension interval for subsequent risk evolution chain and key evidence subgraph extraction.

[0069] Within the monitoring timeframe, the platform reads the dynamic state representation of nodes, calculates the deviation magnitude, duration, and slope of deviation enhancement for each graph node relative to the individual baseline, and filters target graph nodes with continuously increasing state deviation. For example, if the circulatory pressure node shows a mean arterial pressure continuously lower than the individual baseline by 8 to 15 mmHg between 09:04 and 09:12, the peripheral perfusion node shows a decrease in pulse wave amplitude and reduced perfusion stability after 09:06, and the oxygenation status node shows a prolonged blood oxygen recovery time after 09:09, the platform identifies these nodes as target graph nodes. If the electrocardiographic activity node shows only a short-term abnormality at 09:08 and has a low electrocardiographic quality weight, the platform reduces the priority of this node in the target graph node selection to avoid artifacts forming the interpretation path.

[0070] Subsequently, the platform reads the dynamic coupling characterization of the edges and calculates the coupling contribution of each dynamically associated edge within the monitoring time range. The coupling contribution can be determined by the edge weight enhancement magnitude, the consistency of the propagation direction, the time lag matching degree, and the data quality reliability. For example, the dynamic associated edge from the circulatory pressure node to the peripheral perfusion node increases in edge weight from 0.48 to 0.83 between 09:05 and 09:10, and the pressure drop occurs 2 minutes before the perfusion drop, which meets the preset propagation direction and time lag conditions. The platform identifies this dynamic associated edge as the target dynamic associated edge. If the clinical event node is a vasopressor dosage increase event, and the pressure node does not recover in time after the event, the platform can mark the dynamic associated edge from the clinical event node to the circulatory pressure node as an insufficient intervention response associated edge and include it in the candidate propagation path.

[0071] During the path tracing phase, the platform uses the target risk state as the endpoint and backtracks from the risk trigger window to the historical window according to the temporal sequence and the direction of propagation. Starting from the circulatory perfusion risk state at 09:12, the platform sequentially searches for nodes that have continuously increased circulatory pressure, peripheral perfusion, and oxygenation status, and generates candidate paths along the target dynamic correlation edges. For each candidate path, the platform calculates the path contribution characterization, including node deviation contribution, edge coupling contribution, temporal continuity contribution, and quality reliability contribution. If a path exhibits a decrease in circulatory pressure at 09:04, a weakening of peripheral perfusion at 09:06, a delayed oxygenation recovery at 09:09, and a risk trigger at 09:12, and the data quality weight of each node is higher than 0.8, the platform identifies this path as the primary risk evolution chain. If another path depends on low-reliability electrical segments, the platform downgrades this path to an auxiliary path or excludes it.

[0072] After forming the risk evolution chain, the platform further filters key evidence subgraphs based on the contribution of each graph node and dynamic connection edge to the target risk state. The contribution can be determined by the attention weight of the risk decoding network, the node deviation intensity, the magnitude of edge weight changes, and the stability of path occurrence. The platform retains target graph nodes and target dynamic connection edges with contribution values ​​higher than a preset threshold, and removes nodes and edges with low contribution values ​​or insufficient credibility, forming key evidence subgraphs. For example, key evidence subgraphs may include circulatory pressure nodes, peripheral perfusion nodes, oxygenation state nodes, vasopressor event nodes, and dynamic connection edges from circulatory pressure to peripheral perfusion, from peripheral perfusion to oxygenation state, and from vasopressor events to circulatory pressure, recording the time window, contribution value, and quality weight corresponding to each node and dynamic connection edge.

[0073] Through the above processing, the results of cardiac critical care monitoring can be expanded from a single risk warning to a structured interpretation with time sequence, direction of propagation, and contribution of evidence. This enables the platform to clearly express the evolution of risk from early deviation to final triggering, improves the traceability and verifiability of risk warnings, and reduces the interference of low-reliability monitoring segments on the risk interpretation results.

[0074] In some embodiments, path tracing is performed on target graph nodes and target dynamic associated edges according to chronological order and direction of propagation to generate a risk evolution chain, including: Using the time window corresponding to the target risk status as the backtracking starting point, candidate backtracking paths are determined based on the propagation direction of the target's dynamic associated edges; Extract the state deviation changes of each target graph node and the coupling enhancement changes of each target dynamic association edge along the candidate backtracking path to generate a path contribution representation; Candidate backtracking paths are screened and ranked based on path contribution characteristics and time continuity constraints. Candidate backtracking paths that meet the preset contribution criteria are identified as risk evolution chains.

[0075] Specifically, after generating target graph nodes and target dynamic association edges, the monitoring platform uses the target risk status in the cardiac critical care monitoring results as the path tracing endpoint and the time window corresponding to the target risk status as the backtracking starting point. For example, if the platform generates an elevated circulatory perfusion risk alert at 09:18, with the risk triggering window from 09:17 to 09:18, and the target risk status corresponds to a joint anomaly in the circulatory pressure node, peripheral perfusion node, and oxygenation status node, the platform will backtrack from 09:18 to 09:03, reading the node dynamic status representation of each target graph node and the edge dynamic coupling representation of each target dynamic association edge within the monitoring range. The platform determines candidate backtracking paths based on the propagation direction, effective delay, and edge weight enhancement period recorded in the dynamic association edges. If the propagation direction from the circulatory pressure node to the peripheral perfusion node is a decrease in pressure affecting weakened perfusion, and the propagation direction from the peripheral perfusion node to the oxygenation status node is insufficient perfusion affecting oxygenation recovery, then the platform will prioritize generating candidate backtracking paths formed by connecting the circulatory pressure node, peripheral perfusion node, and oxygenation status node.

[0076] After candidate backtracking paths are determined, the platform extracts the state deviation changes of each target graph node along each candidate backtracking path. State deviation changes include the magnitude of deviation relative to the individual baseline, the duration of deviation, the slope of deviation growth, and the confidence level of deviation. For example, between 09:05 and 09:18, the mean arterial pressure deviation of the circulatory pressure node increased from 6 mmHg below the individual baseline to 16 mmHg below the individual baseline; the pulse wave amplitude of the peripheral perfusion node continued to decrease after 09:08; and the blood oxygen recovery time of the oxygenation status node significantly prolonged after 09:12. The platform encodes these changes as node deviation contributions.

[0077] Simultaneously, the platform extracts the enhanced coupling changes of the dynamic associated edges of each target along the candidate backtracking path, including edge weight increment, consistency of propagation direction, temporal lag matching degree, and quality reliability. For example, the edge weight from the circulatory pressure node to the peripheral perfusion node increases from 0.46 to 0.82, and the pressure decrease precedes the perfusion decrease by about 2 minutes, which conforms to the preset propagation delay; the edge weight from the peripheral perfusion node to the oxygenation status node increases from 0.38 to 0.71, and the perfusion decrease precedes the blood oxygenation recovery delay by about 3 minutes. The platform encodes these changes as edge coupling contributions.

[0078] The platform further integrates node deviation contributions and edge coupling contributions to generate a path contribution representation. This representation includes not only individual contribution scores but also path start time, path end time, node anomaly sequence, edge weight enhancement sequence, and path confidence level. For example, a candidate path exhibits a decrease in circulatory pressure at 09:05, a decrease in peripheral perfusion at 09:08, a delay in oxygenation recovery at 09:12, and a risk trigger at 09:18, with each node's corresponding data quality weight exceeding 0.8; the platform sets this candidate path's confidence level to high. Another candidate path exhibits a short-term abnormality in ECG activity at 09:10 and an increased risk probability at 09:11, but the ECG source shows lead dislodgement markers and does not form a stable temporal relationship with changes in blood pressure and perfusion; the platform sets this candidate path's confidence level to low.

[0079] Subsequently, the platform filters and ranks candidate backtracking paths based on path contribution characteristics and temporal continuity constraints. The temporal continuity constraint is used to determine whether anomalies in adjacent nodes of a candidate path occur sequentially within the effective time delay range, and whether there are long-term breaks or spurious connections composed of low-confidence segments in the path.

[0080] In some examples, if a candidate path has a strong node deviation, but the occurrence times of anomalies in adjacent nodes are not clearly sequential, or there is data loss exceeding the preset window length, the platform will lower the ranking of that candidate path. If a candidate path exhibits increased node deviation and synchronously increased edge weights across multiple consecutive time windows, and the propagation direction conforms to the physiological propagation direction set in the cardiac critical care monitoring knowledge base, the platform will raise the ranking of that candidate path.

[0081] Finally, the platform identifies candidate backtracking paths that meet preset contribution criteria as risk evolution chains. These criteria may include a path contribution score reaching a threshold, a required number of key nodes, a required level of credibility for key dynamic connections, and a requirement for temporal continuity. For the aforementioned example, the platform ultimately identifies the continuous decrease in circulatory pressure, the continuous weakening of peripheral perfusion, the delay in oxygenation recovery, and the triggering of circulatory perfusion risk as the primary risk evolution chain, and uses low-confidence, short-term abnormal electrical paths as exclusion paths or auxiliary reference paths.

[0082] Through this implementation method, the platform can filter out risk evolution chains that conform to time sequence, propagation direction and credible contribution from multiple candidate explanatory paths, improve the accuracy of expressing the formation process of cardiac critical disease risk, and enhance the interpretability and verification value of monitoring results.

[0083] In some embodiments, updating the individual baseline, dynamic correlation edge, and monitoring result calibration parameters based on posterior feedback data includes: Obtain the posterior feedback data corresponding to the cardiac critical care monitoring results, and map the posterior feedback data to the corresponding time window, graph node and dynamic association edge; The monitoring result confirmation status, signal interference status, and stability status are determined based on the post-test feedback data. When the stable state indicator meets the baseline update conditions, the individual baseline is updated based on the unified spatiotemporal monitoring sequence of the corresponding time window; Based on the monitoring results, confirm the status and signal interference status, adjust the edge weight calibration parameters of the corresponding dynamic associated edges, and update the monitoring result calibration parameters corresponding to the risk level, risk trend, and risk confidence level.

[0084] Specifically, after the cardiac critical care monitoring results are output, the monitoring platform continuously receives posterior feedback data related to those results. Posterior feedback data can come from bedside terminal verification records, central monitoring station confirmation operations, equipment maintenance records, nursing event records, and subsequent monitoring results. For example, if the platform generates an increased risk warning for circulatory perfusion at 09:18, and medical staff confirm at 09:22 that the patient does indeed have a decrease in mean arterial pressure and weakened peripheral perfusion, while the nursing system records that the vasopressor dosage was adjusted at 09:05, and the equipment system shows that the arterial pressure sensor connection is normal, the platform uses this feedback as confirmation of actual risk. If another patient triggers an ECG abnormality warning at 10:12, but at 10:13 the equipment status record shows lead detachment, and blood pressure, blood oxygen, and pulse wave remain stable, the platform uses this feedback as signal interference data.

[0085] After receiving the posterior feedback data, the platform maps the data to corresponding time windows, graph nodes, and dynamic correlation edges based on the monitored object identifier, risk trigger time, and feedback timestamp. For confirmed risk data, the platform maps the feedback to circulatory pressure nodes, peripheral perfusion nodes, oxygenation status nodes, and related dynamic correlation edges before and after the risk trigger. For signal interference data, the platform maps the feedback to ECG activity nodes, ECG source quality status, and dynamic correlation edges connected to the ECG activity nodes. During the mapping process, the platform records the feedback source, confirmation time, feedback type, reliability level, and corresponding monitoring segment to form a feedback mapping representation, avoiding the misapplication of the same feedback to irrelevant time periods or nodes.

[0086] The platform further determines the monitoring result confirmation status, signal interference status, and stability status based on the post-test feedback data. The monitoring result confirmation status is used to distinguish between situations such as confirmed actual risks, confirmed false alarms, pending verification, and inconsistent results; the signal interference status is used to mark interference sources such as lead detachment, pressure wave damping, probe displacement, and communication interruption; and the stability status indicator is used to mark whether a certain time period can be used as the basis for individual baseline updates.

[0087] For example, if the target patient's heart rate, mean arterial pressure, pulse wave amplitude, and blood oxygen status are all within the individual's historical stable range between 09:30 and 09:50, and the data quality weight is higher than 0.85, the medical staff will mark the status as stable, and the platform will set this time period as a stable state. If there are events such as blood drawing, turning over, and probe re-attaching between 10:00 and 10:05, even if some indicators are not exceeded, the platform will still restrict this time period as a baseline update sample.

[0088] When the stable state indicator meets the baseline update conditions, the platform reads the unified spatiotemporal monitoring sequence within the corresponding time window and removes low-confidence monitoring segments, signal interference segments, and abnormal confirmation segments. Subsequently, the platform performs consistency verification between the remaining stable segments and the current individual baseline, calculating the baseline center change, fluctuation range change, and trend change of monitoring nodes such as heart rate, circulatory pressure, perfusion intensity, and oxygenation status. For example, if the target patient's mean arterial pressure remains stable at around 82 mmHg for 30 consecutive minutes, which is 4 mmHg higher than the original individual baseline, and there are no drastic medication adjustment events, the platform generates baseline update increments according to quality weights and temporal continuity, smoothly updating the individual baseline parameters instead of directly writing a short-term increase into the baseline.

[0089] For dynamic correlation edge updates, the platform adjusts the edge weight calibration parameters based on the monitoring results to confirm the status and signal interference status. If a dynamic correlation edge from a circulatory pressure node to a peripheral perfusion node appears multiple times in the actual risk confirmation path, and the edge weight enhancement has a stable temporal relationship with subsequent risk triggering, the platform increases the initial calibration weight of this dynamic correlation edge under similar monitoring conditions. If a dynamic correlation edge related to an ECG activity node causes multiple false alarms due to lead detachment, the platform reduces the edge weight calibration coefficient when low-quality ECG segments participate in the propagation of graph messages. For dynamic correlation edges from clinical event nodes to circulatory pressure nodes, if the pressure status improves after adjusting vasopressors, the platform enhances the recovery response calibration of this edge; if the pressure continues to decrease after adjustment, the platform records the characteristics of insufficient intervention response and improves the sensitivity of subsequent risk identification.

[0090] For the calibration parameters of the monitoring results, the platform updates the output calibration relationship of risk level, risk trend, and risk confidence based on the feedback confirmation results. If a certain type of high-risk alert is confirmed as a real risk after multiple reviews, the platform can reduce the trigger delay and increase the risk confidence under similar conditions; if a certain type of alert is mainly caused by low-confidence segments, the platform increases the output trigger conditions or converts the alert into a device verification alert. For example, when an ECG abnormality rises alone but blood pressure, perfusion, and oxygenation are stable, the platform reduces the probability of outputting a severe risk level; if a decrease in circulatory pressure, weakened perfusion, and delayed oxygenation recovery occur simultaneously, the platform maintains a high risk confidence even if some ECG data is missing.

[0091] Through the above processing, the posterior feedback data can influence the individual baseline, dynamic correlation edges, and monitoring result calibration parameters, enabling the monitoring platform to gradually adapt to changes in the target patient's condition, differences in equipment status, and verification results over long-term operation. This embodiment can reduce the risk of repeated false alarms, improve the stability of true risk identification, and enhance the individualized adaptation capability of cardiac critical care monitoring results in continuous bedside applications.

[0092] In some embodiments, when the stable state identifier meets the baseline update conditions, the individual baseline is updated based on the unified spatiotemporal monitoring sequence of the corresponding time window, including: The candidate baseline update window is determined based on the stable state identifier, and time windows containing low-confidence monitoring segments, signal interference states, or abnormal confirmation states are removed. The consistency of the unified spatiotemporal monitoring sequence within the candidate baseline update window with the current individual baseline is verified to determine the monitoring segments that can be updated. Based on the baseline deviation status, quality weight, and temporal continuity of the updatable monitoring segments, baseline update increments are generated. The baseline center, fluctuation range, trend of change, and baseline reliability of individual baselines are updated incrementally according to the baseline update.

[0093] Specifically, after the posterior feedback data is written, the monitoring platform first reads the set of time windows corresponding to the stable state identifier, and determines the candidate baseline update windows by combining the monitoring object identifier, risk trigger records, and equipment status records. For example, if the target patient did not trigger a high-risk monitoring result between 09:30 and 10:10, and the medical staff review record shows that the circulatory status is stable, the arterial blood pressure sensor connection is normal, and there are no ECG lead detachment records, the platform divides this time period into multiple candidate baseline update windows. If there is a turning care event between 09:42 and 09:44, a pulse oximeter shift between 09:50 and 09:52, or an increase in arterial blood pressure waveform damping between 10:02 and 10:05, the platform eliminates candidate windows based on low-confidence monitoring segments, signal interference status, and abnormal confirmation status, retaining only windows with good continuity and meeting the quality weight requirements for subsequent updates.

[0094] After obtaining the candidate baseline update window, the platform performs a consistency check between the unified spatiotemporal monitoring sequence within the window and the current individual baseline. The consistency check can calculate the baseline deviation magnitude, direction of change, short-term fluctuation frequency, and cross-source consistency for each monitoring node. For example, in the current individual baseline, the target patient's mean arterial pressure baseline center is 78 mmHg, the heart rate baseline center is 86 beats per minute, and the blood oxygen saturation baseline center is 96%.

[0095] For example, if the mean arterial pressure is stable between 80 and 84 mmHg, the heart rate is stable between 82 and 90 beats per minute, and the blood oxygen saturation is stable between 95% and 97% within a candidate window, and the pulse wave amplitude and blood pressure waveform change in the same direction, the platform determines that the window is consistent with the current individual baseline and can be used as an updatable monitoring segment. If a window has stable values ​​but also exhibits a short transitional state following a recent adjustment of the vasopressor dosage, the platform can mark the window as a temporarily delayed update segment to avoid mistakenly writing transient responses after clinical intervention into the long-term baseline.

[0096] For identified updatable monitoring segments, the platform further generates baseline update increments based on baseline deviation status, quality weights, and temporal continuity. Baseline deviation status indicates the degree of difference between the current stable segment and the original individual baseline; quality weights indicate the reliability of each data source within the stable segment; and temporal continuity indicates whether the stable segment has sufficient duration and consistency between adjacent windows. For example, if, for 20 consecutive minutes, the mean arterial pressure is consistently higher than the original baseline by approximately 4 mmHg, the heart rate differs from the original baseline by less than 3 beats per minute, the data quality weights are all higher than 0.88, and there are no interfering events, the platform will generate a small positive update increment centered on the mean arterial pressure baseline. If a certain indicator deviates significantly but for insufficient duration, the platform reduces the update increment for that indicator to prevent baseline drift caused by a single short-term segment.

[0097] During the update process, the platform employs a smooth update method for different monitoring nodes, rather than directly replacing the original individual baseline with the mean of candidate segments. For the baseline center, the platform calculates a weighted update value based on the current baseline reliability, candidate segment quality weight, and stable duration. For the fluctuation range, the platform updates the normal fluctuation boundary based on the upper and lower quantile values ​​within the candidate segments and the results of anomaly removal. For the trend, the platform updates the trend parameters based on the slope and directional consistency of the indicators within the continuous window. For the baseline reliability, the platform increases or decreases it based on the number of updatable monitoring segments, cross-source consistency, and posterior confirmation level. For example, when a target patient shows a stable upward shift in mean arterial pressure in multiple stable segments and there are no abnormal confirmation records, the platform smoothly updates the mean arterial pressure baseline center from 78 mmHg to 79.5 mmHg and adjusts the fluctuation range from 72 mmHg to 84 mmHg to 74 mmHg to 86 mmHg.

[0098] Through the above implementation method, the individual baseline update process can effectively filter out low-confidence data, signal interference, and abnormal confirmation segments, updating the baseline center, fluctuation range, trend, and baseline confidence only using stable monitoring segments that have undergone consistency verification. This embodiment can reduce the contamination of individual baselines by short-term artifacts and clinical operations, reduce baseline drift errors in long-term monitoring, and improve the individualized accuracy of subsequent deviation normalization, risk identification, and alarm calibration.

[0099] The following are system embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the system embodiments of this application, please refer to the method embodiments of this application.

[0100] Figure 2 This is a schematic diagram of the structure of a cardiac critical care monitoring system based on a spatiotemporal graph neural network provided in an embodiment of this application. Figure 2 As shown, the system includes: The acquisition module 201 is used to acquire multi-source cardiac critical care monitoring data of the target patient, perform signal quality evaluation, time alignment and individual baseline normalization on the multi-source cardiac critical care monitoring data, and generate a unified spatiotemporal monitoring sequence. Module 202 is used to construct an individualized spatiotemporal graph of cardiac critical care based on a unified spatiotemporal monitoring sequence. The individualized spatiotemporal graph of cardiac critical care includes graph nodes for characterizing monitoring channels, physiological states, event contexts and individual baselines, as well as dynamic association edges generated based on physiological prior relationships, temporal propagation relationships and data quality credibility. The generation module 203 is used to input the individualized cardiac critical care spatiotemporal graph into the spatiotemporal graph neural network, perform cross-time window message propagation and state update on graph nodes and dynamically associated edges, and generate node dynamic state representation, edge dynamic coupling representation and global critical care state representation. The extraction module 204 is used to generate cardiac critical care monitoring results based on the global critical care status representation, and extract the risk evolution chain and key evidence subgraph corresponding to the cardiac critical care monitoring results based on the node dynamic state representation and the edge dynamic coupling representation. The update module 205 is used to update the individual baseline, dynamic correlation edge, and monitoring result calibration parameters based on the posterior feedback data.

[0101] In some embodiments, Figure 2 The acquisition module 201 accesses bedside continuous monitoring data and clinical event data according to the monitoring object identifier of the target patient, and establishes an original monitoring data package based on the data source, acquisition time, and equipment status; it extracts quality features from the original monitoring data package, generates quality evaluation results based on waveform stability, temporal integrity, and cross-source consistency, and marks low-confidence monitoring segments based on the quality evaluation results; it establishes a unified time index based on the sampling interval and event occurrence time of different data sources, maps the original monitoring data package to the unified time index, and generates a temporal monitoring representation with missing state coding and quality weights; it generates individual baseline parameters based on the stable monitoring segments of the target patient, and uses the individual baseline parameters to normalize the deviation of the temporal monitoring representation to generate a unified spatiotemporal monitoring sequence.

[0102] In some embodiments, Figure 2 The construction module 202 generates graph node attributes corresponding to the target patient based on the monitoring source, time index, and baseline deviation status in the unified spatiotemporal monitoring sequence; it determines candidate associations between graph nodes based on cardiac critical care monitoring knowledge and the temporal state changes of the target patient; it generates dynamic association edges and corresponding edge weight parameters by comprehensively considering the physiological coupling state, temporal lag state, event response state, and quality reliability state of the candidate associations; and it associates the graph nodes, dynamic association edges, and corresponding time windows to form an individualized cardiac critical care spatiotemporal graph used to characterize the evolution of the target patient's cardiac critical care status.

[0103] In some embodiments, Figure 2 The generation module 203 inputs the node attributes of the graph nodes and the edge weight parameters of the dynamically associated edges into the spatiotemporal graph neural network according to the time window corresponding to the individualized cardiac critical care spatiotemporal graph. Based on the edge weight parameters, it performs graph message aggregation on adjacent graph nodes and performs gating correction on the aggregated messages in combination with the data quality credibility status to generate node update representations within the window. According to the node update representations and edge weight change status between adjacent time windows, it performs temporal state transmission and dynamic associated edge re-estimation to generate node dynamic state representations and edge dynamic coupling representations. It performs cross-scale fusion of node dynamic state representations and edge dynamic coupling representations to generate a global critical care status representation.

[0104] In some embodiments, Figure 2 The generation module 203, for the target graph node within the current time window, determines the set of adjacent graph nodes based on the dynamically associated edges, and calculates the message contribution of each adjacent graph node based on the edge weight parameters; it weights and aggregates the message contribution with the corresponding node attributes to generate an initial aggregated message; it generates a gating coefficient based on the data quality credibility status, and uses the gating coefficient to attenuate and correct the initial aggregated message formed by low credibility sources; it merges the corrected aggregated message with the historical state representation of the target graph node to generate the node update representation within the window.

[0105] In some embodiments, Figure 2 The generation module 203 obtains the node update representation and edge weight change state corresponding to the current time window and the adjacent historical time window; based on the time interval encoding and state preservation parameters, it performs temporal state transfer on the node update representation to generate cross-window node state representation; based on the change relationship between cross-window node state representations and the corresponding edge weight change state, it re-estimates the coupling strength of dynamically associated edges; and it fuses the cross-window node state representation and the re-estimated coupling strength to generate node dynamic state representation and edge dynamic coupling representation.

[0106] In some embodiments, Figure 2 The extraction module 204 inputs the global critical illness status representation into the risk decoding network to analyze the distribution of the critical illness status of the target patient within the current time window and the predicted time window; based on the distribution of critical illness status, the characteristics of continuous status changes, and the monitoring credibility, it generates risk level, risk trend, and risk confidence; it matches the risk level, risk trend, and risk confidence with the preset graded output rules to determine the monitoring prompt level and output trigger status; when the monitoring credibility is lower than the credibility threshold, it suppresses and corrects the output trigger status based on the corresponding low-credibility monitoring segment to generate cardiac critical illness monitoring results.

[0107] In some embodiments, Figure 2The extraction module 204 determines the target risk state and its corresponding monitoring time range based on the cardiac critical care monitoring results; within the monitoring time range, it identifies target graph nodes whose state deviation is continuously enhanced based on the dynamic state representation of nodes, and identifies target dynamic associated edges whose coupling contribution meets preset conditions based on the dynamic coupling representation of edges; according to the temporal relationship and the direction of association propagation, it performs path tracing on the target graph nodes and target dynamic associated edges to generate a risk evolution chain; based on the contribution of each graph node and dynamic associated edge in the risk evolution chain to the target risk state, it selects and forms a key evidence subgraph.

[0108] In some embodiments, Figure 2 The update module 205 acquires the posterior feedback data corresponding to the cardiac critical care monitoring results and maps the posterior feedback data to the corresponding time window, graph node, and dynamic association edge; it determines the monitoring result confirmation status, signal interference status, and stable status identifier based on the posterior feedback data; when the stable status identifier meets the baseline update conditions, it updates the individual baseline based on the unified spatiotemporal monitoring sequence of the corresponding time window; it adjusts the edge weight calibration parameters of the corresponding dynamic association edge according to the monitoring result confirmation status and signal interference status, and updates the monitoring result calibration parameters corresponding to the risk level, risk trend, and risk confidence.

[0109] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0110] Figure 3 This is a schematic diagram of the electronic device 3 provided in an embodiment of this application. Figure 3 As shown, the electronic device 3 of this embodiment includes: a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various system embodiments described above.

[0111] Electronic device 3 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 3 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0112] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0113] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0114] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0115] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which may be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0116] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application, and should all be included within the protection scope of this application.

Claims

1. A method for monitoring cardiac intensive care based on spatiotemporal graph neural networks, characterized in that, include: Acquire multi-source cardiac critical care monitoring data of the target patient, and perform signal quality evaluation, time alignment and individual baseline normalization on the multi-source cardiac critical care monitoring data to generate a unified spatiotemporal monitoring sequence; Based on the unified spatiotemporal monitoring sequence, a personalized spatiotemporal graph of cardiac critical care is constructed. The personalized spatiotemporal graph of cardiac critical care includes graph nodes for characterizing monitoring channels, physiological states, event contexts and individual baselines, as well as dynamic association edges generated based on physiological prior relationships, temporal propagation relationships and data quality credibility. The individualized cardiac critical care spatiotemporal graph is input into a spatiotemporal graph neural network, and cross-time window message propagation and state update are performed on the graph nodes and the dynamically associated edges to generate node dynamic state representations, edge dynamic coupling representations and global critical care state representations. Based on the global critical illness status representation, cardiac critical illness monitoring results are generated, and risk evolution chains and key evidence subgraphs corresponding to the cardiac critical illness monitoring results are extracted based on the node dynamic state representation and the edge dynamic coupling representation. The individual baseline, the dynamic correlation edge, and the monitoring result calibration parameters are updated based on the posterior feedback data.

2. The method according to claim 1, characterized in that, The process of acquiring multi-source cardiac critical care monitoring data from the target patient, performing signal quality evaluation, time alignment, and individual baseline normalization on the multi-source cardiac critical care monitoring data, and generating a unified spatiotemporal monitoring sequence includes: Access bedside continuous monitoring data and clinical event data according to the target patient's monitoring object identification, and establish original monitoring data packages based on data source, collection time and equipment status; Quality features are extracted from the original monitoring data packets, and quality evaluation results are generated based on waveform stability, temporal integrity, and cross-source consistency. Low-confidence monitoring segments are then marked based on the quality evaluation results. A unified time index is established based on the sampling interval and event occurrence time of different data sources. The original monitoring data packets are mapped to the unified time index to generate a time-series monitoring representation with missing state codes and quality weights. Individual baseline parameters are generated based on stable monitoring segments of the target patient, and the deviation of the temporal monitoring representation is normalized using the individual baseline parameters to generate the unified spatiotemporal monitoring sequence.

3. The method according to claim 1, characterized in that, The construction of an individualized spatiotemporal map of cardiac critical care based on the unified spatiotemporal monitoring sequence includes: Based on the monitoring source, time index, and baseline deviation status in the unified spatiotemporal monitoring sequence, graph node attributes corresponding to the target patient are generated. Based on knowledge of cardiac critical care monitoring and the temporal status changes of target patients, candidate associations between the graph nodes are determined; By combining the physiological coupling state, temporal lag state, event response state, and quality reliability state of the candidate associations, dynamic association edges and corresponding edge weight parameters are generated. The graph nodes, the dynamically associated edges, and the corresponding time windows are associated to form the individualized spatiotemporal graph of cardiac critical illness used to characterize the evolution of the target patient's cardiac critical illness state.

4. The method according to claim 1, characterized in that, The step of performing cross-time window message propagation and state updates on the graph nodes and the dynamically associated edges to generate dynamic state representations of nodes, dynamic coupling representations of edges, and global critical state representations includes: According to the time window corresponding to the individualized cardiac critical care spatiotemporal graph, the node attributes of the graph nodes and the edge weight parameters of the dynamically associated edges are input into the spatiotemporal graph neural network. Based on the edge weight parameters, graph messages are aggregated for adjacent graph nodes, and the aggregated messages are gating and corrected in conjunction with the data quality reliability status to generate a node update representation within the window. Based on the node update representation and edge weight change status between adjacent time windows, perform temporal state transfer and dynamic associated edge re-estimation to generate the node dynamic state representation and the edge dynamic coupling representation. The dynamic state representation of the nodes and the dynamic coupling representation of the edges are fused across scales to generate the global critical illness state representation.

5. The method according to claim 4, characterized in that, The process of aggregating graph messages between adjacent graph nodes based on the edge weight parameters, and performing gating correction on the aggregated messages in conjunction with the data quality reliability status to generate an in-window node update representation includes: For the target graph node within the current time window, determine the set of adjacent graph nodes based on the dynamically associated edges, and calculate the message contribution of each adjacent graph node based on the edge weight parameters. The message contribution is weighted and aggregated with the corresponding node attributes to generate an initial aggregated message; A gating coefficient is generated based on the data quality trust status, and the gating coefficient is used to attenuate and correct the initial aggregated message formed by low trust sources; The corrected aggregated message is fused with the historical state representation of the target graph node to generate the node update representation within the window.

6. The method according to claim 4, characterized in that, The step of performing temporal state propagation and dynamic associated edge reestimation based on the node update representation and edge weight change state between adjacent time windows to generate the node dynamic state representation and the edge dynamic coupling representation includes: Obtain the node update representation and edge weight change status corresponding to the current time window and adjacent historical time windows; Based on time interval encoding and state preservation parameters, the node update representation is subjected to temporal state transfer to generate cross-window node state representation. Based on the changing relationship between the cross-window node state representations and the corresponding edge weight change states, the coupling strength of the dynamically associated edges is re-estimated. By integrating the cross-window node state representation and the re-estimated coupling strength, the node dynamic state representation and the edge dynamic coupling representation are generated.

7. The method according to claim 1, characterized in that, The generation of cardiac critical care monitoring results based on the global critical care status characterization includes: The global critical illness status representation is input into the risk decoding network to analyze the distribution of the critical illness status of the target patient within the current time window and the predicted time window. Based on the distribution of severe illness status, the characteristics of continuous changes in status, and the reliability of monitoring, risk level, risk trend, and risk confidence level are generated. The risk level, risk trend, and risk confidence level are matched with preset graded output rules to determine the monitoring prompt level and output trigger status; When the monitoring confidence level is lower than the confidence threshold, the output trigger state is suppressed and corrected based on the corresponding low-confidence monitoring segment to generate the cardiac critical care monitoring result.

8. The method according to claim 7, characterized in that, The extraction of the risk evolution chain and key evidence subgraph corresponding to the cardiac critical care monitoring results based on the node dynamic state representation and the edge dynamic coupling representation includes: The target risk status and its corresponding monitoring time range are determined based on the cardiac critical care monitoring results. Within the monitoring time range, target graph nodes with continuously increasing state deviations are identified based on the node dynamic state characterization, and target dynamic associated edges whose coupling contributions meet preset conditions are identified based on the edge dynamic coupling characterization. Based on the chronological order and the direction of propagation, path tracing is performed on the target graph nodes and the target dynamic associated edges to generate the risk evolution chain; The key evidence subgraph is formed by selecting nodes and dynamically related edges in the risk evolution chain to contribute to the target risk state.

9. The method according to claim 1, characterized in that, The step of updating the individual baseline, the dynamic correlation edge, and the monitoring result calibration parameters based on posterior feedback data includes: Obtain the posterior feedback data corresponding to the cardiac critical care monitoring results, and map the posterior feedback data to the corresponding time window, graph node, and dynamic associated edge; The monitoring result confirmation status, signal interference status, and stability status identifier are determined based on the a posteriori feedback data. When the stable state indicator meets the baseline update conditions, the individual baseline is updated based on the unified spatiotemporal monitoring sequence of the corresponding time window; Based on the confirmed status of the monitoring results and the status of the signal interference, the edge weight calibration parameters of the corresponding dynamic associated edges are adjusted, and the monitoring result calibration parameters corresponding to the risk level, risk trend, and risk confidence level are updated.

10. A cardiac critical care monitoring system based on spatiotemporal graph neural networks, characterized in that, include: The acquisition module is used to acquire multi-source cardiac critical care monitoring data of the target patient, perform signal quality evaluation, time alignment and individual baseline normalization on the multi-source cardiac critical care monitoring data, and generate a unified spatiotemporal monitoring sequence. The construction module is used to construct an individualized spatiotemporal graph of cardiac critical care based on the unified spatiotemporal monitoring sequence. The individualized spatiotemporal graph of cardiac critical care includes graph nodes for characterizing monitoring channels, physiological states, event contexts and individual baselines, as well as dynamic association edges generated based on physiological prior relationships, temporal propagation relationships and data quality credibility. The generation module is used to input the individualized cardiac critical care spatiotemporal graph into the spatiotemporal graph neural network, perform cross-time window message propagation and state update on the graph nodes and the dynamically associated edges, and generate node dynamic state representations, edge dynamic coupling representations and global critical care state representations. The extraction module is used to generate cardiac critical care monitoring results based on the global critical care status representation, and to extract the risk evolution chain and key evidence subgraph corresponding to the cardiac critical care monitoring results based on the node dynamic status representation and the edge dynamic coupling representation. The update module is used to update the individual baseline, the dynamic correlation edge, and the monitoring result calibration parameters based on the posterior feedback data.

Images