A method and system for predicting postoperative complications in thoracic surgery

Abstract

The thoracic surgery postoperative complication prediction method and system of the present application belong to the technical field of medical care information science, and comprise the following steps: S1, collecting preoperative physiological reserve data and postoperative time series monitoring data of a patient; S2, constructing an initialized physiological network graph model based on the preoperative physiological reserve data, and generating an initial state vector for each node in the model; S3, using a cyclic updating unit, combining the last time state vector of the node and the postoperative time series monitoring data to generate an updated node state vector; S4, calculating the vulnerability score of each node and the attention coefficient between nodes based on the updated node state vector; S5, calculating the system resilience index of the physiological network based on the vulnerability scores of all nodes, and quantifying the cascade propagation risk by combining the vulnerability scores and the attention coefficients, which is significantly superior to the traditional static alarm system based on threshold.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of healthcare informatics, in particular to a thoracic surgery postoperative complication prediction method and system. BACKGROUND

[0002] In the postoperative monitoring of thoracic surgery, the traditional complication monitoring method mainly relies on the threshold alarm system of the isolated physiological indicators of the patient. These systems usually trigger an alarm when individual indicators such as heart rate and blood pressure exceed the preset safety range. This status quo makes it difficult for clinicians to obtain effective early warning in the early stages of complications and to understand the internal transmission relationship between organ function failure.

[0003] The above status and deficiencies are mainly due to the limitations of existing analysis techniques:

[0004] One-sidedness and lag of risk assessment. Traditional methods treat each physiological indicator as an independent event, lacking a holistic assessment of the patient as a complex physiological system. This piecemeal approach cannot capture the complex nonlinear dynamic relationships between multiple organs. As a result, alarms are often triggered when organ function has already deteriorated significantly, missing the best opportunity for early intervention.

[0005] Lack of explainability and predictability of risk transmission. Existing technologies are usually a black box. Even if an alarm is issued, it cannot clearly reveal the potential transmission path of complications between different organs. For example, it cannot quantitatively predict whether lung function deterioration is more likely to cause acute kidney injury or heart failure.

[0006] Passivity and blindness of clinical intervention. Due to the inability to predict the cascading transmission path of risks, the clinical team has difficulty in making prospective and targeted interventions. Medical resources cannot be preferentially allocated to the next weak link where problems are most likely to occur, resulting in passive measures that are difficult to achieve precise interruption of the chain reaction of complications.

[0007] The above information disclosed in the background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0008] The purpose of the present application is to provide a thoracic surgery postoperative complication prediction method and system to solve the problems raised in the background art.

[0009] The technical solution of the present application is as follows:

[0010] S1, collecting preoperative physiological reserve data and postoperative time series monitoring data of the patient;

[0011] S2, constructing an initialized physiological network graph model based on preoperative physiological reserve data, and generating an initial state vector for each node in the model;

[0012] S3, using a cycle update unit to generate an updated node state vector in combination with the previous time state vector of the node and postoperative time series monitoring data;

[0013] S4, calculating the vulnerability score of each node and the attention coefficient between nodes based on the updated node state vector;

[0014] S5, calculating the system resilience index of the physiological network based on the vulnerability scores of all nodes, and quantifying the cascade propagation risk in combination with the vulnerability scores and the attention coefficients;

[0015] S6, triggering an early warning when the vulnerability score of any node exceeds a preset first threshold or the system resilience index is lower than a preset second threshold, and identifying and outputting a cascade failure critical path according to the cascade propagation risk.

[0016] Preferably, the nodes in the physiological network graph model represent key physiological units; the initial state vector is composed of preoperative physiological reserve data, and the preoperative physiological reserve data includes lung function test results and cardiac ultrasound parameters.

[0017] Preferably, the cycle update unit is a long short-term memory unit or a gated recurrent unit.

[0018] Preferably, the step of calculating the vulnerability score of each node includes inputting the updated node state vector into a feedforward neural network for nonlinear transformation and dimension reduction to generate the vulnerability score.

[0019] Preferably, the step of quantifying the cascade propagation risk includes inputting the vulnerability score of the risk source node, the vulnerability score of the target node, and the attention coefficient between the two nodes into a risk fusion function to generate the cascade propagation risk.

[0020] Preferably, the step of identifying and outputting the cascade failure critical path specifically includes: constructing a risk propagation directed graph based on the cascade propagation risks of all node pairs, the weight of the edge connecting the nodes in the graph being determined by the negative logarithm of the cascade propagation risk; and starting from any node with a vulnerability score exceeding the first threshold, identifying the path with the minimum cumulative risk in the risk propagation directed graph as the cascade failure critical path using a graph shortest path algorithm.

[0021] Preferably, it further includes: using a backpropagation algorithm on historical case data containing complication occurrence time sequence information to jointly optimize the learnable parameters in the cycle update unit, the feedforward neural network, the graph attention mechanism, and the risk fusion function with the goal of minimizing the model prediction error.

[0022] A thoracic surgery postoperative complication prediction system comprises:

[0023] A data acquisition module is configured to acquire preoperative physiological reserve data and postoperative time series monitoring data of a patient.

[0024] A model construction module is configured to construct an initial physiological network graph model based on the preoperative physiological reserve data, generate an initial state vector, and dynamically update the state vector of each node by using a recurrent update unit.

[0025] A risk parameter calculation module is configured to calculate the vulnerability score of each node and the attention coefficient between nodes based on the updated node state vector.

[0026] A risk index generation module is configured to calculate a system resilience index based on the vulnerability score, and to quantify the cascading propagation risk by combining the vulnerability score and the attention coefficient.

[0027] An early warning and path identification module is configured to trigger an early warning based on the vulnerability score and the system resilience index, and to identify and output a cascading failure critical path according to the cascading propagation risk.

[0028] The present application provides a thoracic surgery postoperative complication prediction method and system, which has the following improvements and advantages compared with the prior art:

[0029] 1. The scheme adopts a recurrent update unit, combines the last time state vector of each node and the postoperative time series monitoring data to generate an updated node state vector, and is obtained by training and learning on a large number of historical case data containing complication occurrence time sequence information. This data-driven modeling method enables the model to capture the nonlinear dynamic law of physiological state evolution over time, so that weak deterioration signals can be identified in the early stage of complication, which is significantly ahead of traditional threshold-based static alarm systems.

[0030] 2. The scheme not only evaluates isolated risks, but also comprehensively considers the system level. By inputting the updated node state vector into a feedforward neural network, the vulnerability score of each node is calculated. This score compresses high-dimensional and complex physiological information into a standardized dimensionless risk scalar, which intuitively quantifies the instability degree of a single physiological unit. The weighted sum of the vulnerability of all nodes provides a macroscopic index for measuring the overall stability of the system. This multi-level evaluation from nodes to systems can help clinicians distinguish local problems from impending systemic collapse and provide more comprehensive evaluation.

[0031] 3. The core progress of this scheme is to introduce a graph attention mechanism to calculate the attention coefficient between nodes, and input it into the risk fusion function together with the vulnerability score of the risk source node and the vulnerability score of the target node, establishing an interpretable probability model: the occurrence probability of a cascade event depends on the dangerous degree of the risk source, the susceptibility of the target node, and the coupling strength between them; this makes the model no longer a black box, but can reveal which organ deterioration is most likely to affect another specific organ;

[0032] 4. This scheme provides forward-looking and operational clinical decision support; when the early warning is triggered, the scheme does not stop at the alarm, but further constructs a risk propagation directed graph according to the cascade propagation risk of all node pairs, and identifies and outputs the critical path of cascade failure by using the shortest path algorithm of the graph; so that the clinical team can prioritize the allocation of monitoring and intervention resources to the potential risk chain that needs the most attention, which is beyond the reach of the prior art; this transition from risk assessment to path prediction is the most significant progress of this scheme. BRIEF DESCRIPTION OF DRAWINGS

[0033] The present application will be further explained in conjunction with the accompanying drawings and embodiments:

[0034] Figure 1 is a step flow chart of a thoracic surgery postoperative complication prediction method of the present application. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical scheme and advantages of the present application more clear and explicit, the present application will be further described in detail in conjunction with specific embodiments.

[0036] Example 1:

[0037] Please refer to Figure 1 The present application provides a thoracic surgery postoperative complication prediction method, comprising the following steps: S1, collecting preoperative physiological reserve data and postoperative time series monitoring data of a patient;

[0038] S2, constructing an initialized physiological network graph model based on the preoperative physiological reserve data, and generating an initial state vector for each node in the model;

[0039] S3, using a cycle update unit to generate an updated node state vector by combining the last time state vector of the node and the postoperative time series monitoring data;

[0040] S4, calculating the vulnerability score of each node and the attention coefficient between nodes based on the updated node state vector;

[0041] S5, calculating the system resilience index of the physiological network based on the vulnerability scores of all nodes, and quantifying the cascade propagation risk by combining the vulnerability scores and the attention coefficients;

[0042] S6, triggering a warning when the vulnerability score of any node exceeds a preset first threshold or the system resilience index is lower than a preset second threshold, and identifying and outputting a cascading failure critical path according to a cascading propagation risk;

[0043] The initialization construction of the physiological network graph model is to abstract the patient as a system composed of multiple key physiological units. Each physiological unit, such as the lung, heart or kidney, is represented as a node in the model To quantify the inherent physiological reserve of each node before surgery, an initial state vector is constructed, which is composed of objective indicators such as forced vital capacity obtained from lung function tests or left ventricular ejection fraction from echocardiography, aiming to establish a standardized digital baseline for the deep heterogeneity of physiological reserves of different patients as the starting point for all subsequent dynamic calculations;

[0044] To ensure the explicitness and reproducibility of the model, in this embodiment, the key physiological unit nodes for common postoperative complications in thoracic surgery can be specifically selected as: lung, heart, kidney, liver, and coagulation system; each unit corresponds to a node in the model; in the initial construction stage of the model, we assume that it is a fully connected graph, i.e. there is potential mutual influence between any two nodes. Therefore, the 'neighbor node set ' of any node is defined as the set of all other nodes in the network except the node itself. This setting allows the model to autonomously learn the dynamic coupling strength between any two physiological units during the training process without presetting fixed connections based on prior medical knowledge;

[0045] To dynamically track the physiological state evolution of the patient after surgery, the state of the physiological network is updated in real time; this step introduces a recurrent update unit, which can be implemented as a long short-term memory unit or a gated recurrent unit, effectively fusing historical information and current input using its gating mechanism;

[0046] The preoperative physiological reserve data can include but is not limited to forced vital capacity, forced expiratory volume in one second, left ventricular ejection fraction, etc.; the postoperative time series monitoring data is a multi-dimensional vector that can include but is not limited to regular monitoring indicators of the corresponding physiological units, such as heart rate, mean arterial pressure, oxygen saturation, respiratory rate, hourly urine output, platelet count, creatinine level and bilirubin level, etc., which are usually collected in hours;

[0047] Before entering the cyclic update unit, the postoperative time series monitoring data is Z-score standardized to eliminate the influence of different indicator dimensions; the calculation formula is:

[0048]

[0049] in, is a node At the moment The original multi-dimensional monitoring data vector, is the normalized data vector, and are the mean and standard deviation of the indicator in the training dataset; : subscripts, representing specific nodes in the physiological network;

[0050] The mathematical expression of this dynamic update process is as follows:

[0051]

[0052] This formula is derived from the principle of recurrent neural networks, and the technical motivation is to capture and simulate the continuity and dependence of physiological states in the time dimension;

[0053] in, is the state vector, Subscript, representing the physiological unit node , Represents the current moment , Represents physiological unit node At the current moment The state vector of At time t, the external monitoring data of the i-th physiological unit after normalization is: is the state vector of the node at the previous moment, As a nonlinear transformation function of the cyclic update unit, Represents the learnable weight parameter inside the update function, and the value is obtained by training on a large amount of historical case data; this formula is applied to each time point to calculate the latest state vector for each node in the network, and the output high-dimensional vector It contains rich feature expressions of node history information and current status, which constitutes the direct input for all subsequent risk quantification calculations, ensuring the real-time and accuracy of predictions;

[0054] The dynamics and accuracy of the prediction of the present invention are improved. This scheme adopts a cyclic update unit to combine the node's previous state vector with the postoperative time series monitoring data to generate an updated node state vector. The mathematical form of the update process is defined as ; this formula is not mathematically derived from first principles, but is a model constructed based on the recognized effectiveness of recurrent neural networks in processing time series data; the specific parameters of this model are not preset, but are learned through training on a large amount of historical case data containing timing information of the occurrence of complications; this data-driven modeling approach enables the model to capture the nonlinear dynamic rules of physiological state evolution over time, thereby identifying weak deterioration signals at an early stage of complication onset, significantly ahead of traditional static threshold-based alarm systems.

[0055] Embodiment 2

[0056] The nodes in the physiological network graph model represent key physiological units; the initial state vector is composed of preoperative physiological reserve data, including pulmonary function test results and echocardiographic parameters;

[0057] The recurrent update unit is a long short-term memory unit or a gated recurrent unit;

[0058] The step of calculating the vulnerability score of each node includes inputting the updated node state vector into a feedforward neural network for nonlinear transformation and dimensionality reduction to generate the vulnerability score;

[0059] The step of quantifying the cascading propagation risk includes inputting the vulnerability score of the risk source node, the vulnerability score of the target node, and the attention coefficient between the two nodes into a risk fusion function to generate the cascading propagation risk;

[0060] Based on the real-time updated node state vector , the system risk is further quantified in multiple dimensions; the process of calculating the vulnerability score of each physiological unit node is realized through a feedforward neural network, which performs nonlinear transformation and dimensionality reduction on the high-dimensional state vector, aiming to map the high-dimensional, complex physiological state vector to a standardized, easily interpretable single risk scalar, thereby realizing the quantitative monitoring of the deterioration degree of individual organ function state; the calculation of the node vulnerability score is defined by the following formula:

[0061]

[0062] wherein, : the vulnerability score of the node, which is a dimensionless scalar, : subscript, representing the node , : superscript, representing the transpose of the vector, is the vulnerability score of the node at time , which is a dimensionless scalar, is the real-time state vector of the node, are learnable weight and bias parameters obtained through model training to constitute the feedforward network, and is a standard nonlinear activation function, and the vulnerability score is the basis for triggering individual node early warning and is a key input parameter for calculating the overall resilience of the system and the subsequent cascade propagation risk, with a dimension of one;

[0063] the macroscopic index of the overall stability of the system, i.e., the system resilience index , is obtained by weighted integration of the vulnerability of each node; the calculation logic aims to provide a macroscopic index to measure the overall stability of the system, because the instability of a single organ does not necessarily represent a systemic collapse, while the decline in the overall resilience of the system is a more critical early warning signal; the calculation formula of the index is:

[0064]

[0065] wherein, is the resilience index of the system at time , : subscript, representing the node in the summation index , is the dimensionless vulnerability score of node , is a dimensionless physiological importance weight preset based on medical expert consensus or clinical scoring system, is the total number of nodes in the network;

[0066] To make the assignment of weight more objective and operable, a preferred embodiment is to refer to a recognized clinical scoring system, such as the sequential organ failure assessment (SOFA) score; can be set according to the scoring weight of each organ system in the SOFA scoring system ; when there is a lack of explicit prior knowledge or for the purpose of simplifying the model, the physiological importance weight of all nodes can also be set to 1, i.e. , at this time the system resilience index is the complement of the arithmetic mean of all node vulnerability scores;

[0067] The formula generates a resilience index between 0 and 1 by weighted summation and normalization of the vulnerability of all nodes, ensuring dimensional consistency;

[0068] To predict the conduction of complications, a graph attention mechanism is introduced to quantify the dynamic interaction strength between nodes, which is derived from the core formula of the graph attention network:

[0069]

[0070] wherein, is the node to node At time , the dimensionless attention coefficient, : subscript, representing different physiological unit nodes, : represents time , : exponential function, is the shared weight parameter learned by model training, is the node state vector, denotes vector concatenation, is the neighbor node set of node , is the LeakyReLU activation function; this formula is normalized by Softmax, ensuring is the dimensionless weight describing relative importance, : neighbor node set of node , : indicates that node belongs to the neighbor node set of node ;

[0071] The occurrence probability of a cascading failure event depends on the severity of the risk source, the susceptibility of the target node, and the functional coupling strength between the two; to mathematically integrate these three core elements, the cascading propagation risk is quantified; its risk fusion function takes the form of a logistic regression:

[0072]

[0073] where, is the conditional probability of risk propagating from node to node , : subscript, representing the source node and the target node, respectively, and are the dimensionless vulnerability scores of the source node and the target node, respectively, is the dimensionless attention coefficient between the two, is the learnable regression coefficient, is the Sigmoid function; the inputs of this formula are all dimensionless quantities, so the output probability value is also dimensionless, consistent in dimension; this calculation integrates all previous parameters and outputs a cascading propagation probability value with clear physical meaning, providing a quantitative basis for identifying the most dangerous concurrent complication chain reaction path, : dimensionless attention coefficient between node and node ;

[0074] The depth and breadth of risk assessment in this invention are expanded; the solution not only assesses isolated risks, but also comprehensively considers them from a system level; by inputting the updated node state vector into the feedforward neural network, the vulnerability score of each node is calculated. This score compresses high-dimensional and complex physiological information into a standardized dimensionless risk scalar, intuitively quantifying the degree of instability of a single physiological unit; on this basis, the formula Calculate the system resilience index of the physiological network. The practical significance of this formula is that it takes the weighted sum of the vulnerabilities of all nodes, providing a macro-metric for measuring the overall stability of the system. This multi-level assessment, from nodes to systems, can help clinicians distinguish between local problems and impending systemic collapse, providing a more comprehensive assessment.

[0075] The interpretability and predictability of the complication conduction pathway of the present invention are achieved; the core advancement of this solution is the introduction of a graph attention mechanism to calculate the attention coefficient between nodes and compare it with the vulnerability score of the risk source node, , vulnerability score of target node Input together into the risk fusion function, quantifying the slave node To Node Cascade risk ;formula The practical significance of this is that it establishes an interpretable probabilistic model: the probability of a cascade event depends on the dangerousness of the risk source, the susceptibility of the target node, and the coupling strength between the two; this makes the model no longer a black box, but can reveal which organ's deterioration is most likely to affect another specific organ.

[0076] Example 3

[0077] The steps of identifying and outputting the critical path of cascading failures specifically include: constructing a risk propagation directed graph based on the cascading propagation risks of all node pairs, where the weights of the edges connecting the nodes in the graph are determined by the negative logarithm of the cascading propagation risk; and using a graph shortest path algorithm in the risk propagation directed graph, starting from any node whose vulnerability score exceeds a first threshold, identifying the path with the minimum cumulative risk as the critical path of cascading failures;

[0078] In the early warning and path identification stage, the system sets two early warning trigger conditions: the vulnerability score of any node Exceeds the preset first threshold , or the system resilience index Below the preset second threshold The setting of these thresholds is not arbitrary, but is based on statistical analysis of historical clinical data, and is iteratively optimized to optimize specific performance indicators, such as F1 score, of the prediction model on an independent validation set, aiming to ensure that the sensitivity and specificity of the early warning reach a balance point required by clinical application;

[0079] The optimization process of the threshold is as follows: on an independent validation data set, by adjusting the threshold, a series of different confusion matrices can be obtained. Based on these confusion matrices, the true positive rate and false positive rate are calculated, and the receiver operating characteristic curve is drawn; in this embodiment, the point on the ROC curve that maximizes the F1 score is selected, and the corresponding vulnerability score and system resilience index are determined as the optimal early warning threshold and The formula for calculating the F1 score is: wherein, : F1 score, a comprehensive indicator for measuring the prediction accuracy of the model, which is the harmonic mean of precision and recall, : precision, the proportion of samples predicted as positive examples by the model that are actually positive examples, : recall, the proportion of all samples that are actually positive examples that are successfully predicted as positive examples by the model;

[0080] Once the early warning is triggered, the system will perform identification of the cascading failure critical path; this process is first based on the cascading propagation risk of all node pairs A risk propagation directed graph is constructed; in this graph, the weight of the directed edge from node to node is defined as the negative logarithm of the propagation risk, wherein, : the weight of the directed edge from node to node in the risk propagation directed graph, : natural logarithm function, : cascading propagation risk of risk from node to node , this mathematical transformation converts the problem of finding the maximum probability propagation path into the problem of finding the path with the minimum cumulative weight in the graph;

[0081] The cumulative risk of the path is defined by the following formula:

[0082]

[0083] This formula provides a clear mathematical objective for finding the most likely path of the complication chain reaction;

[0084] wherein, : represents path All the from nodes to the node , is the dimensionless cumulative risk of the path , is a path composed of a sequence of nodes, is the cascading propagation risk from node to node on the path; the system starts from any high-vulnerability node and uses the graph shortest path algorithm, such as Dijkstra algorithm, in the risk propagation directed graph to efficiently identify the path with the minimum cumulative risk ; this identified path is the cascading failure critical path, which reveals the sequence of organ function failure most likely to occur under the current state, providing prospective guidance for targeted clinical intervention.

[0085] Example 4

[0086] On the historical case data containing the timing information of the occurrence of complications, the back propagation algorithm is used to minimize the prediction error of the model as the target, and the learnable parameters in the recurrent update unit, the feedforward neural network, the graph attention mechanism and the risk fusion function are jointly optimized;

[0087] To ensure the accuracy of the model parameters and the reliability of the prediction results, a key training step is performed; this step is performed on a dataset containing a large amount of historical case data, which not only contains the physiological monitoring data of patients, but also accurately labels the time sequence information of the occurrence of various complications; the goal of training is to minimize the prediction error of the model; by using the back propagation algorithm, the system performs end-to-end joint optimization of all learnable parameters in the entire model; these parameters specifically include the weights in the recurrent update unit, the weights and biases in the feedforward neural network for calculating vulnerability , the shared weights in the graph attention mechanism , and the regression coefficients in the risk fusion function ; This joint optimization method ensures that all components of the model work together to serve the final prediction goal, enabling the model to autonomously learn complex patterns of physiological changes and deep laws of organ interactions from data;

[0088] To achieve the above optimization goal, the present application uses binary cross entropy as the loss function to measure the prediction error of the cascading propagation risk probability , and the definition of the loss function is as follows:

[0089]

[0090] where, : value of loss function, used to measure the gap between model prediction and true result, : summation symbol, representing the accumulation of the calculation results of all samples from the 1st to the : subscript, representing the index of a particular sample in the batch, : total number of samples in the training batch, : true label of the sample, indicating whether the cascade event occurs, 0 or 1, : probability of the model predicting the occurrence of the cascade event, i.e. ; : natural logarithm function;

[0091] The optimization of the model uses the Adam optimizer, and the learning rate and other hyperparameters are determined by grid search on the validation set, for example, the initial learning rate can be set to ; The entire model is trained on a large-scale historical case data set containing at least thousands of patients' complete time series data and complication annotation information to ensure its generalization ability and prediction accuracy.

[0092] Embodiment 5

[0093] A postoperative complication prediction system for thoracic surgery, comprising: a data acquisition module for acquiring preoperative physiological reserve data and postoperative time series monitoring data of a patient;

[0094] a model construction module for constructing an initial physiological network graph model based on the preoperative physiological reserve data, generating an initial state vector, and dynamically updating the state vector of each node using a recurrent update unit;

[0095] a risk parameter calculation module for calculating the vulnerability score of each node and the attention coefficient between nodes based on the updated node state vector;

[0096] a risk index generation module for calculating the system resilience index based on the vulnerability score, and quantifying the cascade propagation risk by combining the vulnerability score and the attention coefficient;

[0097] an early warning and path identification module for triggering early warning based on the vulnerability score and the system resilience index, and identifying and outputting the key path of cascade failure according to the cascade propagation risk;

[0098] ​Corresponding to the above method, a postoperative complication prediction system for thoracic surgery is provided; the system is embodied by the organic combination of a series of functional units; the implementation of the system includes a functional unit for interfacing with a hospital information system or a monitoring device to obtain preoperative physiological reserve data and postoperative time series monitoring data of a patient; another functional unit receives these data, performs the initialization construction of the physiological network graph model, and continuously dynamically updates the node state vector using a cycle update unit; the risk parameter calculation functional unit calculates the vulnerability score of each node and the attention coefficient between nodes in real time based on the updated state vector; the risk index generation functional unit further generates the system resilience index and the quantified cascade propagation risk according to these parameters; finally, the early warning and path identification functional unit monitors the vulnerability score of the node and the system resilience index, and triggers an early warning once the index reaches the preset threshold, and simultaneously calls the graph algorithm to identify and output the most clinically alarming critical path of cascade failure; these functional units work together to form a complete technical closed loop from raw data input to provide precise clinical decision support, which is automated and intelligent;

[0099] The present scheme provides prospective and operable clinical decision support; when the early warning is triggered, the scheme does not stop at the alarm, but further constructs a risk propagation directed graph according to the cascade propagation risk of all node pairs, and uses the shortest path algorithm to identify and output the critical path of cascade failure; for example, a specific beneficial effect is that for a postoperative patient, the system can not only warn of lung function deterioration, but also clearly indicate that in the current state, the most likely next complication is acute kidney injury, not heart failure; the identification of this specific path enables the clinical team to prioritize monitoring and intervention resources to the potential risk chain that needs the most attention, such as strengthening the monitoring of kidney function indicators, adjusting fluid management strategies, etc., so as to achieve precise blocking of the chain reaction of complications, which is beyond the reach of the prior art; this shift from risk assessment to path prediction is the most significant progress of the present scheme.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for predicting complications after thoracic surgery, characterized in that: The steps include: S1. Collect the patient's preoperative physiological reserve data and postoperative time series monitoring data; S2. Construct an initialized physiological network graph model based on preoperative physiological reserve data and generate an initial state vector for each node in the model; S3, using a cyclic update unit to combine the node's previous state vector with the postoperative time series monitoring data to generate an updated node state vector; S4. Calculating a vulnerability score of each node and an attention coefficient between nodes based on the updated node state vector. The step of calculating the vulnerability score of each node includes inputting the updated node state vector into a feedforward neural network for nonlinear transformation and dimensionality reduction to generate a vulnerability score. S5. Calculating the system resilience index of the physiological network based on the vulnerability scores of all nodes, and combining the vulnerability score with the attention coefficient to quantify the cascade propagation risk. The step of quantifying the cascade propagation risk includes inputting the vulnerability score of the risk source node, the vulnerability score of the target node, and the attention coefficient between the two nodes into a risk fusion function to generate the cascade propagation risk. S6. When the vulnerability score of any node exceeds a preset first threshold or the system resilience index falls below a preset second threshold, an early warning is triggered, and a critical path of cascading failure is identified and output based on the cascading propagation risk. The steps of identifying and outputting the critical path of cascading failure specifically include: constructing a risk propagation directed graph based on the cascading propagation risk of all node pairs, where the weights of the edges connecting the nodes in the graph are determined by the negative logarithm of the cascading propagation risk; Starting from any node whose vulnerability score exceeds the first threshold, the graph shortest path algorithm is used in the risk propagation directed graph to identify the path with the minimum cumulative risk as the critical path of cascading failure.

2. A method for predicting complications after thoracic surgery according to claim 1, characterized in that: The nodes in the physiological network graph model represent key physiological units; the initial state vector is composed of preoperative physiological reserve data, which includes pulmonary function test results and cardiac ultrasound parameters.

3. The method for predicting postoperative complications of thoracic surgery according to claim 1, characterized in that: The recurrent update unit is a long short-term memory unit or a gated recurrent unit.

4. A method for predicting complications after thoracic surgery according to claim 1, characterized in that: Also includes: Based on historical case data containing temporal information of complication occurrence, the backpropagation algorithm is used to jointly optimize the learnable parameters in the cyclic update unit, feedforward neural network, graph attention mechanism, and risk fusion function with the goal of minimizing the model prediction error.

5. A system for predicting thoracic surgery postoperative complications, applied to a method for predicting thoracic surgery postoperative complications according to any one of claims 1 to 4, characterized in that: include: Data acquisition module, used to collect patients' preoperative physiological reserve data and postoperative time series monitoring data; A model building module is used to construct an initialized physiological network graph model based on preoperative physiological reserve data, generate an initial state vector, and dynamically update the state vector of the node using a cyclic update unit; The risk parameter calculation module is used to calculate the vulnerability score of each node and the attention coefficient between nodes based on the updated node state vector; The step of calculating the vulnerability score of each node includes inputting the updated node state vector into a feedforward neural network for nonlinear transformation and dimensionality reduction to generate a vulnerability score; A risk indicator generation module is used to calculate the system resilience index based on the vulnerability score and combine the vulnerability score with the attention coefficient to quantify the cascading risk. The steps for quantifying the cascading risk include inputting the vulnerability score of the risk source node, the vulnerability score of the target node, and the attention coefficient between the two nodes into the risk fusion function to generate the cascading risk. An early warning and path identification module is used to trigger early warnings based on vulnerability scores and system resilience indexes, and identify and output critical cascading failure paths based on cascading propagation risks. The steps for identifying and outputting critical cascading failure paths specifically include: constructing a risk propagation directed graph based on the cascading propagation risks of all node pairs, where the weights of the edges connecting the nodes in the graph are determined by the negative logarithm of the cascading propagation risk; Starting from any node whose vulnerability score exceeds the first threshold, the graph shortest path algorithm is used in the risk propagation directed graph to identify the path with the minimum cumulative risk as the critical path of cascading failure.

Images