A postoperative care risk early warning system for crohn's disease

Abstract

The application provides a postoperative care risk early warning system for Crohn's disease, and the system comprises: a multi-modal time sequence feature extraction component for encoding Crohn's disease postoperative test index time sequence, imaging pixel distribution and nursing text, and outputting each modal time sequence embedding vector; a dynamic graph construction component for constructing a patient state dynamic space-time graph evolving over time by taking a timestamp as a node and taking inter-modal pathological correlation strength as an edge weight, and updating node features by a graph attention mechanism to aggregate cross-modal pathological information; and a graph attention risk classification component for calculating a complication risk probability distribution based on the updated node features of the dynamic space-time graph. The system solves the problem that the prior art cannot capture the dynamic evolution characteristics of cross-modal pathological correlation, overcomes the defect that static feature splicing is insufficient to capture early weak evolution characteristics, and realizes objective quantitative early warning of the early evolution trend of Crohn's disease postoperative complications.

Description

Technical Field

[0001] This invention relates to the field of medical informatics, specifically to a postoperative care risk warning system for Crohn's disease. Background Technology

[0002] Crohn's disease, a chronic inflammatory granulomatous disease, has a high incidence of postoperative complications, necessitating continuous monitoring and risk warning of patients' physiological status in clinical care. Existing postoperative risk warning systems typically rely on single physiological parameter monitoring devices or conventional electronic medical record management systems. These systems continuously collect time-series data of patients' laboratory indicators, such as inflammatory response protein levels and white blood cell counts. When the value of a specific indicator exceeds a pre-set static safety threshold, an alarm mechanism is immediately triggered. This approach, based on comparing a single value with a static threshold, only reflects the absolute deviation of a single physiological indicator at the current moment. It fails to comprehensively consider the patient's overall physiological state, resulting in a fragmented warning mechanism that struggles to assess the patient's postoperative risk trajectory from a holistic perspective.

[0003] To overcome the limitations of single-indicator monitoring, existing technologies have introduced general postoperative scoring systems or risk prediction models based on traditional machine learning. General scoring systems calculate risk scores by manually inputting multiple test indicators and using a linear combination with fixed weights. Traditional machine learning models concatenate heterogeneous data such as postoperative test indicators, imaging pixel distribution, and nursing texts into a one-dimensional feature vector, which is then input into a classifier for training and prediction. While these approaches incorporate multi-source data, they flatten heterogeneous data into static feature vectors during data processing, severing the dynamic evolutionary relationships between different modalities over time. Linear combination or static concatenation methods cannot capture the deep-seated nonlinear interactions between different pathological indicators and cannot reflect the time-dependent characteristics of pathological effects between modalities.

[0004] Based on the specific implementation methods of the existing technologies, current Crohn's disease postoperative nursing risk warning schemes suffer from a core technical problem: they fail to capture the dynamic evolution characteristics of cross-modal pathological associations. The occurrence and development of postoperative complications in Crohn's disease are not caused by isolated mutations of a single indicator, but rather by the interaction and nonlinear dynamic evolution of multimodal pathological data over time. Existing technologies isolate or statically stitch together test indicators, imaging data, and nursing texts, disrupting the topological structure of pathological associations between modalities and lacking a graph structure evolution mechanism over time. This results in the system's inability to extract implicit associations and dynamic temporal evolution logic between cross-modal pathological indicators, causing the loss of early, subtle evolutionary characteristics of complications and leading to delayed warnings. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide a postoperative care risk warning system for Crohn's disease, in order to solve or at least partially solve the above-mentioned problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides a postoperative care risk warning system for Crohn's disease, the system comprising: Multimodal temporal feature extraction component: used to encode the time series of laboratory indicators, image pixel distribution and nursing text after Crohn's disease surgery, and output temporal embedding vectors of each modality; Dynamic graph construction component: used to construct a dynamic spatiotemporal graph of patient status that evolves over time, with timestamps as nodes and the pathological association strength between modalities as edge weights. The pathological association strength refers to the pathological association strength between different modalities at the same timestamp. The node features are updated through a graph attention mechanism to aggregate cross-modal pathological information. Graph Attention Risk Classification Component: Used to calculate the probability distribution of complication risk based on the node features of the updated dynamic spatiotemporal graph.

[0007] Preferably, the multimodal temporal feature extraction component includes a temporal convolution sub-component, a visual coding sub-component, and a textual semantic sub-component; The temporal convolutional sub-component performs dilated causal convolution on the time series of the test index to extract temporal dependency features; The visual coding sub-component performs spatial feature mapping on the image pixel distribution to obtain image spatial features; The text semantic sub-component performs word embedding transformation on the nursing text to extract text semantic features; The multimodal temporal feature extraction component performs heterogeneous feature alignment and concatenation of the temporal dependency features, the image spatial features, and the text semantic features to generate the temporal embedding vectors of each modality.

[0008] Preferably, before performing the dilated causal convolution operation, the temporal convolution sub-component also performs missing value imputation and outlier smoothing operations; The missing value imputation operation is based on the known data points before and after the missing position in the time series of the test index, which are constructed using a cubic spline interpolation function to fill the missing value. The outlier smoothing operation calculates the local mean and local standard deviation of the sliding window of the test index time series. Data points that deviate from the local mean of the sliding window by more than a preset multiple of the local standard deviation are identified as outliers, and the outliers are replaced with the local mean of the sliding window. The processed sequence is then input into the dilated causal convolution operation.

[0009] Preferably, the dynamic graph construction component includes an association computation subcomponent and a graph structure generation subcomponent; The correlation calculation subcomponent calculates the Pearson correlation coefficient of any two modes at the same timestamp based on the temporal embedding vector of each mode, maps the Pearson correlation coefficient to a Gaussian kernel function value, and obtains the initial correlation degree between modes. The graph structure generation sub-component compares the initial correlation degree between the modalities with a preset pathological correlation threshold, retains the correlation degree greater than the preset pathological correlation threshold as the edge weight, and constructs edges between the corresponding nodes to form a dynamic spatiotemporal graph of patient status that evolves over time.

[0010] Preferably, the edge weights of the patient state dynamic spatiotemporal graph undergo dynamic decay and state transition updates as the timestamp progresses, specifically including: Based on adjacent timestamps, the edge weight at the current time is equal to the edge weight at the previous time multiplied by the time decay factor, plus the initial correlation degree between the modes calculated at the current time multiplied by the state transition matrix. The time decay factor is set according to the half-life of inflammatory markers after Crohn's disease surgery, and the state transition matrix is ​​set according to the physiological conduction delay between different modes.

[0011] Preferably, updating node features through a graph attention mechanism specifically includes the following steps: Based on any target node in the dynamic spatiotemporal graph of the patient's state, obtain a set of neighboring nodes that are connected to the target node by an edge. The feature vector of the target node is concatenated with the feature vectors of each neighboring node in the set of neighboring nodes and then input into the activation function to calculate the attention coefficient between the target node and each of the neighboring nodes. The attention coefficients are normalized to obtain the attention weights; The feature vectors of each of the neighboring nodes are weighted and summed according to the attention weights to update the feature vector of the target node.

[0012] Preferably, the dynamic graph construction component employs a multi-head graph attention mechanism to update the node features, specifically including the following steps: The input features of the target node and the set of neighboring nodes are respectively input into multiple groups of graph attention mechanisms, and the graph attention mechanisms of each group calculate and output multi-head update features in parallel; All the multi-head update features are concatenated along the feature dimension and then linearly transformed to obtain the fused update features; The fused updated features are added element-wise to the original feature vector of the target node using a residual join operation, and the residual join result is subjected to layer normalization processing to output the final feature vector of the target node.

[0013] Preferably, the graph attention risk classification component includes a spatiotemporal feature pooling sub-component and a fully connected classification sub-component; The spatiotemporal feature pooling subcomponent performs max pooling operation on the node features of the updated dynamic spatiotemporal graph of the patient state along the time axis to extract the target global temporal features; The fully connected classification subcomponent maps the target global temporal features to the complication category space after Crohn's disease surgery, and outputs the complication risk probability distribution through a normalized exponential function. The complication risk probability distribution includes the anastomotic leakage risk probability, the abdominal infection risk probability, and the postoperative bleeding risk probability.

[0014] Preferably, it also includes a preoperative baseline alignment compensation component, which is used to extract the historical test feature baseline of Crohn's disease patients before surgery, calculate the covariance matrix between the historical test feature baseline and the temporal embedding vector of each modality, and perform eigenvalue decomposition on the covariance matrix to obtain an orthogonal transformation matrix. The historical test feature baseline is mapped to the postoperative dynamic space using the orthogonal transformation matrix to obtain the baseline compensation vector; The graph attention risk classification component calculates the probability distribution of complication risk after fusing the baseline compensation vector with the node features.

[0015] Preferably, it also includes an intervention strategy generation and feedback correction component, which generates a corresponding nursing intervention strategy vector based on the complication risk probability distribution; After executing the nursing intervention strategy vector, the actual test index time series of the next time stamp is obtained, and the predicted test index series of the next time stamp is predicted based on the node features of the current time and the nursing intervention strategy vector. The root mean square error between the actual test index time series and the predicted test index series is calculated. If the root mean square error exceeds a preset correction threshold, the gradient is calculated based on the root mean square error and backpropagated to the fully connected classification sub-component in the graph attention risk classification component. The weight parameters of the fully connected classification sub-component are then fine-tuned.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a postoperative nursing risk early warning system for Crohn's disease. By constructing a dynamic graph construction component and a graph attention risk classification component, it maps multimodal heterogeneous data after Crohn's disease surgery into a dynamic spatiotemporal graph structure that evolves over time. This solves the core technical problem that existing technologies cannot capture the dynamic evolution features of cross-modal pathological associations. The multimodal temporal feature extraction component outputs temporal embedding vectors for each modality, preserving the inherent temporal dependencies of each modality. The dynamic graph construction component constructs a dynamic spatiotemporal graph of patient states that evolves over time, reconstructing fragmented heterogeneous data into a graph structure with topological relationships, making implicit pathological associations between modalities explicit. By updating node features through a graph attention mechanism, cross-modal pathological information is aggregated, and nonlinear interaction features between cross-modal pathological indicators are extracted. The graph attention risk classification component calculates the probability distribution of complication risks, overcoming the deficiency of traditional static feature splicing in capturing early weak evolution features. This achieves objective quantitative early warning of the early evolution trend of postoperative complications in Crohn's disease. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall structure of a postoperative care risk warning system for Crohn's disease provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall business process of a postoperative care risk warning system for Crohn's disease provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the entire process of multimodal temporal feature extraction provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the dynamic spatiotemporal graph construction and edge weight dynamic update process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the multi-head graph attention node feature optimization and update process provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the preoperative baseline compensation fusion and complication risk classification calculation process provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the nursing intervention generation and model closed-loop feedback correction process provided in an embodiment of the present invention. Detailed Implementation

[0019] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0020] Reference Figure 1 This embodiment provides a postoperative care risk warning system for Crohn's disease, the system comprising: Multimodal temporal feature extraction component: used to encode the time series of laboratory indicators, image pixel distribution and nursing text after Crohn's disease surgery, and output temporal embedding vectors of each modality; Dynamic graph construction component: used to construct a dynamic spatiotemporal graph of patient status that evolves over time, with timestamps as nodes and the pathological association strength between modalities as edge weights. The pathological association strength refers to the pathological association strength between different modalities at the same timestamp. The node features are updated through a graph attention mechanism to aggregate cross-modal pathological information. Graph Attention Risk Classification Component: Used to calculate the probability distribution of complication risk based on the node features of the updated dynamic spatiotemporal graph.

[0021] Reference Figure 2 In one embodiment, the postoperative care risk warning system for Crohn's disease runs on the server side of the hospital information system, interacting with the electronic medical record system, medical image archiving and communication system, and nursing information system. The system acquires real-time time series of laboratory indicators, image pixel distribution, and nursing text data of postoperative Crohn's disease patients. It encodes this heterogeneous data using a multimodal temporal feature extraction component, outputting temporal embedding vectors for each modality. A dynamic graph construction component uses timestamps as nodes and the pathological correlation strength between modalities as edge weights to construct a dynamic spatiotemporal graph of patient status that evolves over time. It updates node features through a graph attention mechanism to aggregate cross-modal pathological information. A graph attention risk classification component calculates the probability distribution of complication risks based on the updated node features of the dynamic spatiotemporal graph. The system pushes the calculated probability distribution of complication risks to the nurse station terminal, providing decision support for clinical nursing staff.

[0022] The multimodal temporal feature extraction component receives time series of laboratory indicators from the electronic medical record system, pixel distribution of images from the medical image archiving and communication system, and nursing text from the nursing information system. The time series of laboratory indicators includes numerical sequences of 12 physiological indicators collected every 6 hours post-surgery, including white blood cell count, neutrophil percentage, C-reactive protein, procalcitonin, hemoglobin, platelet count, albumin, creatinine, blood urea nitrogen, electrolytes, prothrombin time, and activated partial thromboplastin time. The pixel distribution of images is a pixel matrix of daily post-operative enhanced abdominal CT images, with an image resolution of 512×512, including plain scan, arterial phase, venous phase, and delayed phase sequences. The nursing text consists of text information recorded by nurses every 2 hours, detailing patient symptoms, signs, nursing procedures, and changes in condition, with each record ranging from 50 to 200 characters in length. The multimodal temporal feature extraction component encodes the data from these three modalities separately, preserving the inherent structural features and temporal dependencies of each modality, and outputting a unified-dimensional temporal embedding vector for each modality.

[0023] The dynamic graph construction component receives the temporal embedding vectors of each modality from the multimodal temporal feature extraction component. It creates an independent node for each timestamp, with each node containing the temporal embedding vectors of all modalities at that timestamp. The component calculates the pathological association strength between any two modalities at the same timestamp and uses this association strength as the edge weight between corresponding nodes. As time progresses, new timestamp nodes are continuously added to the graph structure, and the edge weights are dynamically updated based on the newly acquired modal data, forming a dynamic spatiotemporal graph of the patient's state evolving over time. The component updates the features of each node through a graph attention mechanism, aggregating the cross-modal pathological information of the node and its neighboring nodes, enabling the node features to reflect the patient's overall physiological state and pathological evolution trend at that time point.

[0024] The graph attention risk classification component receives updated node features from the dynamic graph construction component, performs pooling operations on these features along the time axis, and extracts global temporal features that reflect the overall evolution of the patient's condition after surgery. The component maps these global temporal features to the Crohn's disease postoperative complication category space and outputs a complication risk probability distribution using a normalized exponential function. This distribution includes three dimensions: anastomotic leakage risk probability, abdominal infection risk probability, and postoperative bleeding risk probability. Each dimension ranges from 0 to 1, with higher values ​​indicating a greater risk of the corresponding complication.

[0025] This embodiment maps multimodal heterogeneous data into a dynamic spatiotemporal graph structure, using graph nodes to represent timestamps and edge weights to represent the strength of pathological associations between modalities, thus achieving a topological representation of cross-modal pathological information. By aggregating cross-modal pathological information through a graph attention mechanism, it captures the nonlinear interactions between different pathological indicators, solving the problem that existing technologies cannot capture the dynamic evolution characteristics of cross-modal pathological associations.

[0026] Reference Figure 3 In a preferred embodiment, the multimodal temporal feature extraction component includes a temporal convolution sub-component, a visual encoding sub-component, and a textual semantic sub-component. The temporal convolution sub-component performs dilated causal convolution on the time series of the test indicators to extract temporal dependency features. The visual encoding sub-component performs spatial feature mapping on the pixel distribution of the images to obtain image spatial features. The textual semantic sub-component performs word embedding transformation on the nursing text to extract textual semantic features. The multimodal temporal feature extraction component performs heterogeneous feature alignment and concatenation of the temporal dependency features, image spatial features, and textual semantic features to generate temporal embedding vectors for each modality.

[0027] The temporal convolutional sub-component receives the preprocessed time series of test indicators. The input sequence has a dimension of T×12, where T is the time step. The temporal convolutional sub-component uses a three-layer dilated causal convolutional network to extract temporal dependency features. The first layer has a dilation coefficient of 1, a kernel size of 3, and 64 output channels; the second layer has a dilation coefficient of 2, a kernel size of 3, and 128 output channels; the third layer has a dilation coefficient of 4, a kernel size of 3, and 256 output channels. The formula for calculating the dilated causal convolution is:

[0028] in, The output feature at time 1 The value, It is the first The weights of each convolutional kernel, The input sequence at time... The value, It is the coefficient of thermal expansion. This refers to the size of the convolution kernel. Dilated causal convolution expands the receptive field without increasing the number of kernel parameters by setting different dilation coefficients, enabling it to capture long-range temporal dependencies in the time series of test indicators. Simultaneously, the structure of causal convolution ensures that the output depends only on the input at the current time step and previous steps, without revealing information from future time steps, thus meeting the causality requirements of time series data processing.

[0029] The visual coding subcomponent receives the pixel matrix of an enhanced abdominal CT image. The input image has dimensions of 512×512×4, where 4 corresponds to the four sequences: plain scan, arterial phase, venous phase, and delayed phase. The visual coding subcomponent uses a pre-trained ResNet-50 model as the basic feature extractor, removing the last two fully connected layers and retaining the first four convolutional and pooling layers. The input CT image pixel matrix is ​​fed into the ResNet-50 model, outputting a feature map with dimensions of 16×16×2048. The visual coding subcomponent performs global average pooling on this feature map, averaging the feature maps of each channel to obtain a 2048-dimensional image spatial feature. Global average pooling preserves the global spatial information of the image while reducing the feature dimensionality, thus reducing the computational cost of subsequent processing.

[0030] The text semantic subcomponent receives nursing text data and first preprocesses the text, including word segmentation, stop word removal, and case neutralization. The text semantic subcomponent then uses a pre-trained BERT-base-chinese model to perform word embedding transformations on the preprocessed text. Each sentence is input into the BERT model, and the vector corresponding to the 0th token in the last layer's output is obtained; this is the sentence's semantic vector, with a dimension of 768. For nursing records containing multiple sentences, the text semantic subcomponent performs average pooling on the semantic vectors of all sentences to obtain the 768-dimensional text semantic features at that timestamp.

[0031] The multimodal temporal feature extraction component performs heterogeneous feature alignment and concatenation on the 256-dimensional temporal dependency features output by the temporal convolution sub-component, the 2048-dimensional image spatial features output by the visual encoding sub-component, and the 768-dimensional text semantic features output by the text semantic sub-component. First, the three types of features are mapped to the same dimension through linear transformations, with the weight matrices of the linear transformations obtained through training. The temporal dependency features are mapped to 256 dimensions using a 256×256 linear transformation matrix; the image spatial features are mapped to 256 dimensions using a 2048×256 linear transformation matrix; and the text semantic features are mapped to 256 dimensions using a 768×256 linear transformation matrix. Then, the aligned three types of features are concatenated along their feature dimensions to obtain 768-dimensional temporal embedding vectors for each modality. The multimodal data types and corresponding feature dimensions are shown in Table 1 below.

[0032] Table 1. Multimodal data types and corresponding feature dimensions.

[0033] This embodiment extracts features from three modalities using three independent sub-components, preserving the inherent structural features of each modality. Through heterogeneous feature alignment and concatenation, features of different dimensions are mapped to a unified feature space, laying the foundation for subsequent dynamic graph construction and cross-modal information aggregation.

[0034] Reference Figure 4 In a preferred embodiment, the dynamic graph construction component includes an association calculation subcomponent and a graph structure generation subcomponent. The association calculation subcomponent calculates the Pearson correlation coefficient between any two modalities at the same timestamp based on the temporal embedding vectors of each modality, maps the Pearson correlation coefficient to a Gaussian kernel function value, and obtains the initial correlation degree between modalities. The graph structure generation subcomponent compares the initial correlation degree between modalities with a preset pathological correlation threshold, retains correlation degrees greater than the preset pathological correlation threshold as edge weights, and constructs edges between corresponding nodes to form a dynamic spatiotemporal graph of patient states evolving over time.

[0035] The correlation computation subcomponent receives the temporal embedding vectors of each modality output by the multimodal temporal feature extraction component, for each timestamp. Obtain the test modality embedding vector under this timestamp. Image modal embedding vector and text modality embedding vectors Each vector has a dimension of 256. The association calculation subcomponent calculates the Pearson correlation coefficient between any two modal embedding vectors. The formula for calculating the Pearson correlation coefficient is:

[0036] in, It is a moment Lower mode With mode Pearson correlation coefficient, It is modal At any moment The first of the time-series embedding vectors 3D eigenvalues It is modal At any moment The first of the time-series embedding vectors 3D eigenvalues It is modal At any moment The mean of the temporal embedding vector, This is the dimension of the temporal embedding vector, with a value of 256. The Pearson correlation coefficient ranges from -1 to 1; a larger absolute value indicates a stronger linear correlation between the two modes.

[0037] The correlation calculation subcomponent maps the calculated Pearson correlation coefficient to a Gaussian kernel function value to obtain the initial correlation degree between modes. The formula for calculating the Gaussian kernel function is:

[0038] in, It is a moment Lower mode With mode The initial degree of correlation, This is the bandwidth parameter of the Gaussian kernel, with a value of 0.3. The Gaussian kernel function maps the Pearson correlation coefficient from the interval [-1,1] to the interval [0,1]. When the Pearson correlation coefficient is 1, the initial correlation is 1; when the Pearson correlation coefficient is -1, the initial correlation is close to 0. Through the Gaussian kernel function mapping, the linear correlation between modalities is converted into a measure of the pathological correlation strength, which is consistent with the characteristic that positive correlations are predominant among indicators in the pathological process.

[0039] The graph structure generation subcomponent receives the initial intermodal correlation degree output by the correlation calculation subcomponent and compares it with a preset pathological correlation threshold. The preset pathological correlation threshold is set to 0.3. When the initial correlation degree is greater than 0.3, a significant pathological correlation is considered to exist between the two modalities. This initial correlation degree is retained as an edge weight, and an edge is constructed between the corresponding nodes. When the initial correlation degree is less than or equal to 0.3, the pathological correlation between the two modalities is considered insignificant, and no edge is constructed. For each timestamp node, the graph structure generation subcomponent constructs edges between three pairs of modalities: test-image, test-text, and image-text, forming a static subgraph at that time. As time progresses, the graph structure generation subcomponent connects the static subgraphs at each time point in chronological order, constructing edges between nodes of the same modality at adjacent timestamps. The edge weight is set to 0.8, forming a dynamic spatiotemporal graph of the patient's state evolving over time. The preset pathological correlation threshold and edge construction rules are shown in Table 2 below.

[0040] Table 2 Preset pathological association thresholds and edge construction rules

[0041] This embodiment calculates the initial correlation between modalities using the Pearson correlation coefficient and Gaussian kernel function, converting the linear correlation between modalities into a quantitative measure of the strength of pathological associations. Significant associations are screened by setting a preset pathological association threshold, and a dynamic spatiotemporal graph structure is constructed, enabling the graph structure to accurately reflect the pathological association topology between modalities at different time points for the patient.

[0042] In a preferred embodiment, the step of updating node features through a graph attention mechanism in the dynamic graph construction component includes: for any target node in the dynamic spatiotemporal graph, obtaining a set of neighboring nodes that are connected by edges to the target node; concatenating the feature vector of the target node with the feature vectors of each neighboring node in the set and inputting the concatenation into an activation function to calculate the attention coefficients between the target node and each neighboring node; normalizing the attention coefficients to obtain attention weights; and weighting and summing the feature vectors of each neighboring node according to the attention weights to update the feature vector of the target node.

[0043] After constructing a dynamic spatiotemporal graph of the patient's state evolving over time, the dynamic graph construction component performs feature update operations on each node in the graph. For any target node... First, obtain the set of neighboring nodes that are connected to the target node by an edge. The neighborhood node set includes nodes that have intermodal edge connections with the target node at the same time, and nodes that have temporal edge connections with the target node at adjacent timestamps. For example, for time... The test modal node, whose neighborhood nodes include time... The image modal nodes and text modal nodes, as well as the time... and time The test modal nodes.

[0044] The dynamic graph construction component will use the feature vector of the target node. The feature vectors of each neighboring node in the neighboring node set Using the same linear transformation matrix Perform the transformation to obtain the transformed target node features. and neighborhood node features The transformed target node features are concatenated with the features of each neighboring node to obtain a concatenated feature vector. ,in This represents the feature concatenation operation. The concatenated feature vector is input into the attention weight vector. The attention coefficients between the target node and its neighboring nodes are calculated using the LeakyReLU activation function. The formula for calculating the attention coefficients is:

[0045] in, The target node With neighboring nodes Attention coefficient between them It is an attention weight vector with dimensions 512×1. It is a linear transformation matrix with dimensions of 256×256, and the negative slope of LeakyReLU is 0.2.

[0046] The dynamic graph construction component normalizes the attention coefficients between the target node and all its neighboring nodes to obtain the attention weights. The normalization process uses the softmax function, and the calculation formula is as follows:

[0047] in, The target node For neighboring nodes Attention weights The target node With the set of neighboring nodes The Middle The attention weights of each neighboring node are calculated. Through softmax normalization, the sum of the attention weights of all neighboring nodes is 1, ensuring the stability of the feature update process.

[0048] The dynamic graph construction component weights and sums the transformed feature vectors of each neighboring node according to attention weights, then applies the sigmoid activation function to obtain the updated target node feature vector. The update formula is:

[0049] in, It is the updated target node feature vector. It is the Sigmoid activation function. Through a weighted summation operation, the features of the target node are fused with cross-modal pathological information from its neighboring nodes, which can more comprehensively reflect the patient's overall physiological state at that point in time.

[0050] This embodiment automatically learns the importance weights of neighboring nodes to the target node through a graph attention mechanism, achieving adaptive aggregation of cross-modal pathological information. The calculation of attention weights is based on the similarity of node features, which can automatically focus on neighboring nodes that are more closely related to the pathology of the target node, thus improving the effectiveness of feature aggregation.

[0051] In a preferred embodiment, the graph attention risk classification component includes a spatiotemporal feature pooling subcomponent and a fully connected classification subcomponent. The spatiotemporal feature pooling subcomponent performs max pooling on the node features of the updated dynamic spatiotemporal graph along the time axis to extract the target global temporal features. The fully connected classification subcomponent maps the target global temporal features to the Crohn's disease postoperative complication category space and outputs a complication risk probability distribution through a normalized exponential function. The complication risk probability distribution includes the probability of anastomotic leakage, the probability of abdominal infection, and the probability of postoperative bleeding.

[0052] The spatiotemporal feature pooling subcomponent receives updated node features from the dynamic graph construction component. The node features have a dimension of T×3×256, where T is the time step, 3 is the number of modalities, and 256 is the feature dimension of each node. The spatiotemporal feature pooling subcomponent performs max pooling on the feature sequence of each modality along the time axis. For the feature sequence of the test modality, the maximum value of each dimension on the time axis is taken, resulting in a 256-dimensional global temporal feature for the test modality. Similarly, 256-dimensional global temporal features for the image modality and 256-dimensional global temporal features for the text modality are obtained respectively. The spatiotemporal feature pooling subcomponent concatenates the global temporal features of the three modalities along their feature dimensions to obtain a 768-dimensional target global temporal feature. Max pooling preserves the maximum value of each feature dimension on the time axis, capturing key abnormal features in the development of postoperative complications.

[0053] The fully connected classification subcomponent receives the 768-dimensional global temporal features of the target from the spatiotemporal feature pooling subcomponent and performs classification through two fully connected layers. The first fully connected layer has an input dimension of 768 and an output dimension of 128, with ReLU activation. The second fully connected layer has an input dimension of 128 and an output dimension of 3, corresponding to three postoperative complications of Crohn's disease. The fully connected classification subcomponent converts the output of the second fully connected layer into a complication risk probability distribution using a normalized exponential function. The formula for calculating the normalized exponential function is:

[0054] in, It is the first The probability of such complications It is the output of the second fully connected layer. The logit value of the class. Corresponding to anastomotic leakage, Corresponding to abdominal infection, Corresponding to postoperative bleeding; The output of the second fully connected layer The logit value of the class. Table 3 below is a table of complication types and their corresponding output dimensions. Table 3. Complication Types and Corresponding Output Dimensions

[0055] This embodiment extracts global temporal features through spatiotemporal feature pooling, converting the node features of the dynamic spatiotemporal graph into fixed-dimensional feature vectors to provide input for the classifier. By using a fully connected classification sub-component and a normalized exponential function to output the probability distribution of complication risks, it achieves quantitative risk assessment of the three main postoperative complications of Crohn's disease.

[0056] In a preferred embodiment, the temporal convolution sub-component performs missing value imputation and outlier smoothing operations before performing the dilated causal convolution operation. The missing value imputation operation fills in the missing values ​​by constructing a cubic spline interpolation function based on known data points before and after the missing positions in the test index time series. The outlier smoothing operation calculates the sliding window local mean and local standard deviation of the test index time series, identifies data points deviating from the sliding window local mean by more than three times the local standard deviation as outliers, replaces the outliers with the sliding window local mean, and inputs the processed sequence into the dilated causal convolution operation.

[0057] After receiving the original time series of test indicators, the temporal convolutional component first performs missing value imputation. Some data points may be missing in the original time series of test indicators due to reasons such as patients being temporarily away for examination, testing equipment malfunction, or failed blood collection. For each missing location... The temporal convolutional sub-component selects three known data points before and after it. , If the missing position is located at the beginning or end of the sequence, and sufficient preceding and following data points cannot be obtained, the six most recent known data points are selected. The temporal convolutional sub-component constructs a cubic spline interpolation function based on the selected known data points. ,satisfy And the second derivative is continuous within each interval. (The missing positions are...) Substitute into the cubic spline interpolation function The missing values ​​are then filled in. The cubic spline interpolation function has second-order smoothness, which allows it to better fit the continuous changing trend of physiological indicators and avoids the abrupt changes caused by linear interpolation.

[0058] After imputation of missing values, the temporal convolutional subcomponent performs outlier smoothing. Outliers are data points that significantly deviate from the normal range due to factors such as testing errors, specimen contamination, or temporary physiological fluctuations in patients. The temporal convolutional subcomponent uses a sliding window method to calculate the local mean and local standard deviation, with a sliding window size set to 5. For each data point... Calculate the local mean within its window. and local standard deviation .like If the data point is not found to be an outlier, the temporal convolution sub-component will replace it with the local mean of its window. The three-standard-deviation criterion, based on the statistical properties of the normal distribution, can effectively identify extreme outliers with a probability of less than 0.3%.

[0059] This embodiment preprocesses the original time series of test indicators by imputing missing values ​​and smoothing outliers, removing the interference of dirty data on the subsequent feature extraction process. Cubic spline interpolation and sliding window statistical methods can preserve the true trend of physiological indicators, improving the accuracy and robustness of time series feature extraction.

[0060] In a preferred embodiment, the edge weights of the dynamic spatiotemporal graph of the patient's state, which evolves over time, undergo dynamic decay and state transition updates as the timestamps progress. For two adjacent timestamps, the edge weight at the current time is equal to the edge weight at the previous time multiplied by a time decay factor, plus the initial intermodal correlation calculated at the current time multiplied by the state transition matrix. The time decay factor is set based on the half-life of inflammatory markers after Crohn's disease surgery, and the state transition matrix is ​​set based on the physiological transmission delay between different modalities.

[0061] The dynamic graph construction component acquires the initial correlation between modalities at each new timestamp and then dynamically updates the weights of all edges in the dynamic spatiotemporal graph. This dynamic update of edge weights considers the time decay characteristics of pathological processes and the physiological transmission delay characteristics between modalities. The formula for calculating the dynamic update of edge weights is:

[0062] in, It is a moment Lower mode With mode edge weights, It is the time decay factor. It is the state transition matrix from mode pair to modal pair The transition probability, It is a moment Lower mode With mode The initial degree of correlation.

[0063] Time decay factor The timing was determined based on the half-life of key inflammatory markers after Crohn's disease surgery. C-reactive protein (CRP) is an important indicator of the degree of inflammation after Crohn's disease surgery, with a half-life of approximately 19 hours. The system's time sampling interval is 6 hours, therefore the time decay factor... The value is set to 0.8. The role of the time decay factor is to gradually weaken the pathological correlation strength at historical moments over time, which is consistent with the metabolic pattern of inflammatory markers in the body.

[0064] State transition matrix The state transition matrix is ​​a 3×3×3×3 four-dimensional matrix, where the first two dimensions correspond to target modality pairs and the last two dimensions correspond to source modality pairs. The element values ​​of the state transition matrix are set according to the physiological transmission delay between different modalities. The physiological transmission delay between test indicators and imaging is approximately 6 hours, meaning changes in test indicators take about 6 hours to appear on imaging. Therefore, the transition probability from the test-imaging modality pair to itself is set to 0.7, and the transition probability from other modality pairs to the test-imaging modality pair is set to 0.1. The physiological transmission delay between test indicators and nursing text is approximately 2 hours, meaning changes in test indicators can be quickly reflected in patient symptoms and signs. Therefore, the transition probability from the test-text modality pair to itself is set to 0.9, and the transition probability from other modality pairs to the test-text modality pair is set to 0.05. The physiological transmission delay between imaging and nursing text is approximately 4 hours; therefore, the transition probability from the image-text modality pair to itself is set to 0.8, and the transition probability from other modality pairs to the image-text modality pair is set to 0.1. The time decay factor corresponds to the half-life of inflammatory markers, as shown in Table 4 below.

[0065] Table 4. Correspondence between time decay factors and the half-life of inflammatory markers

[0066] This embodiment employs a dynamic edge weight update mechanism based on a time decay factor and a state transition matrix to ensure that the evolution of the dynamic spatiotemporal graph closely reflects the physical laws of real pathological evolution. The time decay factor simulates the metabolic process of inflammatory markers, and the state transition matrix simulates the physiological transmission delay between modes, thereby improving the graph structure's ability to represent the evolution of the patient's pathological state.

[0067] Reference Figure 5 In a preferred embodiment, the dynamic graph construction component employs a multi-head graph attention mechanism to update node features. The input features of the target node and its neighboring node set are input into multiple independent graph attention mechanisms, which compute in parallel to output multi-head updated features. All multi-head updated features are concatenated along their feature dimensions and subjected to a linear transformation to obtain fused updated features. The fused updated features are then joined with the original feature vector of the target node using an element-wise residual join operation, and the residual join result is subjected to layer normalization to output the final feature vector of the target node.

[0068] The dynamic graph construction component employs an 8-head graph attention mechanism to update node features. The 8 independent graph attention mechanisms share the same structure but each has its own independent linear transformation matrix. and attention weight vector , The feature vector of the target node. eigenvectors of the neighboring node set The inputs are fed into eight independent graph attention mechanisms. Each graph attention mechanism, following the aforementioned graph attention calculation process, outputs a 256-dimensional multi-head update feature. .

[0069] The dynamic graph construction component concatenates eight multi-head update features along their feature dimensions, resulting in a 2048-dimensional concatenated feature. Then, a 2048×256 linear transformation matrix is ​​used to transform the concatenated feature, yielding a 256-dimensional fused update feature. The multi-head graph attention mechanism learns attention weights for neighboring nodes from different representation subspaces, capturing different types of cross-modal pathological associations and improving the richness and comprehensiveness of feature aggregation.

[0070] The dynamic graph construction component performs a residual connection operation, which involves element-wise addition of the merged updated features and the original feature vector of the target node, to obtain residual features. The residual connection operation alleviates the vanishing gradient problem during deep neural network training, enabling the network to learn identity mappings more easily and ensuring the effective transfer of original feature information.

[0071] The dynamic graph construction component performs layer normalization on the residual connection results, outputting the feature vector of the final target node. The formula for layer normalization is:

[0072] in, These are the residual features of the input. It is the mean of the input features. It is the variance of the input features. It is a very small constant, taking the value of This is used to avoid the denominator being zero. and These are learnable scaling and offset parameters. Layer normalization normalizes the features of each sample, which can accelerate the convergence speed of the network and improve the training stability of the model.

[0073] This embodiment further improves the effect of node feature updating by using a multi-head graph attention mechanism, residual connections, and layer normalization. The multi-head attention mechanism can capture multifaceted cross-modal pathological associations, while residual connections and layer normalization can improve the training efficiency and stability of the model.

[0074] Reference Figure 6In a preferred embodiment, the system further includes a preoperative baseline alignment compensation component. This component extracts the preoperative historical test feature baseline of Crohn's disease patients, calculates the covariance matrix between the preoperative historical test feature baseline and the multimodal temporal embedding vector, and performs eigenvalue decomposition on the covariance matrix to obtain an orthogonal transformation matrix. The orthogonal transformation matrix is ​​then used to map the preoperative historical test feature baseline to the postoperative dynamic space to obtain the baseline compensation vector. The graph attention risk classification component fuses the baseline compensation vector with node features and calculates the probability distribution of complication risk.

[0075] The preoperative baseline alignment compensation component extracts historical laboratory data from Crohn's disease patients within the three months prior to surgery from the electronic medical record system. This historical data includes 12 physiological indicators identical to those used in postoperative examinations. The component calculates the average value of each physiological indicator over the three months prior to surgery, forming a 12-dimensional preoperative historical laboratory characteristic baseline. The preoperative history test characteristics baseline reflects the patient's basic physiological state before surgery. There are significant differences in the basic physiological state among different patients, and the risk of postoperative complications is closely related to the patient's preoperative basic physiological state.

[0076] The preoperative baseline alignment compensation component calculates the covariance matrix between the preoperative historical test characteristic baseline and the multimodal temporal embedding vector. The multimodal temporal embedding vector has a dimension of 768, therefore the covariance matrix... The dimension is 12×768. The elements of the covariance matrix... The first characteristic representing preoperative features The dimension of the postoperative temporal embedding vector The covariance between dimensions is calculated using the following formula:

[0077] in, It is the number of training samples. It is the first The first preoperative characteristics of the sample dimension, It is the first of all preoperative features of the samples The mean of the dimension, It is the first The postoperative temporal embedding vector of the nth sample dimension, It is the postoperative temporal embedding vector of all samples. The mean of the dimension.

[0078] Preoperative baseline alignment compensation component for covariance matrix Perform singular value decomposition to obtain the orthogonal transformation matrix. Orthogonal transformation matrix The column vectors are the covariance matrix. The feature vector has a dimension of 12×12. An orthogonal transformation matrix is ​​used to map the preoperative historical test feature baseline to the postoperative dynamic space to obtain the baseline compensation vector:

[0079] in, It is a 12-dimensional baseline compensation vector. This is a matrix transpose operation. Through orthogonal transformation, the preoperative feature baseline is mapped from the preoperative space to the postoperative dynamic space, enabling effective fusion of the preoperative baseline with postoperative temporal features.

[0080] The graph attention risk classification component fuses the baseline compensation vector with the updated node features output by the dynamic graph construction component. Specifically, the 12-dimensional baseline compensation vector is mapped to a 256-dimensional vector through a 12×256 linear transformation matrix, and then element-wise added to the 256-dimensional feature vector of each node to obtain the fused node features. Based on the fused node features, the graph attention risk classification component performs subsequent spatiotemporal feature pooling and classification operations to calculate the probability distribution of complication risk.

[0081] This embodiment integrates the patient's preoperative basic physiological status information into the postoperative risk assessment process through a preoperative baseline alignment compensation component. By using covariance matrix eigenvalue decomposition and orthogonal transformation, spatial alignment between preoperative and postoperative features is achieved, providing individualized reference benchmarks for early postoperative data and improving the individualized accuracy of risk assessment.

[0082] Reference Figure 7 In a preferred embodiment, the system further includes an intervention strategy generation and feedback correction component, which generates a corresponding nursing intervention strategy vector based on the complication risk probability distribution. After executing the nursing intervention strategy vector, the actual test indicator time series at the next time stamp is obtained. The root mean square error (RMSE) between the actual test indicator time series and the predicted indicator series is calculated. If the RMSE exceeds a preset correction threshold, the gradient is calculated based on the RMSE and backpropagated to the fully connected classification sub-component in the graph attention risk classification component, and the weight parameters of the fully connected classification sub-component are fine-tuned.

[0083] The intervention strategy generation and feedback correction component receives the complication risk probability distribution output by the attention risk classification component and generates a corresponding 10-dimensional nursing intervention strategy vector. Each dimension of the nursing intervention strategy vector corresponds to a specific nursing operation, including prolonged fasting time, increased frequency of gastrointestinal decompression, adjustment of antibiotic dosage, adjustment of intravenous nutritional support intensity, increased frequency of drainage tube flushing, increased frequency of vital sign monitoring, increased frequency of abdominal sign observation, activity restriction, adjustment of pain management intensity, and adjustment of psychological intervention intensity. The value of each dimension ranges from 0 to 1, with a larger value indicating a greater intensity of the corresponding nursing operation. The intervention strategy generation and feedback correction component determines the values ​​of each dimension of the nursing intervention strategy vector based on the different dimension values ​​of the complication risk probability distribution. For example, when the anastomotic leakage risk probability is greater than 0.7, the values ​​of four dimensions—prolonged fasting time, increased frequency of gastrointestinal decompression, increased frequency of drainage tube flushing, and increased frequency of abdominal sign observation—are set to 1.

[0084] After nursing staff execute the nursing operations corresponding to the nursing intervention strategy vector, the intervention strategy generation and feedback correction component obtains the actual time series of test indicators for the next time stamp. Simultaneously, based on the node features and nursing intervention strategy vector at the current time, the intervention strategy generation and feedback correction component predicts the test indicator sequence for the next time stamp. The prediction model employs a simple multilayer perceptron, with the node features and nursing intervention strategy vector at the current time as input, and the predicted values ​​of 12 test indicators for the next time stamp as output.

[0085] The intervention strategy generation and feedback correction component calculates the root mean square error (RMSE) between the actual test index time series and the predicted index series. The formula for calculating the RMSE is:

[0086] in, This is the number of test indicators, with a value of 12. It is the first The actual value of the test index, It is the first The predicted value of the test indicator.

[0087] If the calculated root mean square error exceeds a preset correction threshold (0.5), the current risk classification model is considered to have prediction bias and requires fine-tuning. The intervention strategy generation and feedback correction component calculates the gradient based on the root mean square error and backpropagates the gradient to the fully connected classification sub-component in the graph attention risk classification component. A stochastic gradient descent algorithm is used to fine-tune the weight parameters of the fully connected classification sub-component, with a learning rate of 0.001 and 10 iterations. The fine-tuning process only updates the weight parameters of the fully connected classification sub-component, without modifying the parameters of the multimodal temporal feature extraction component and the dynamic graph construction component, thus avoiding damage to the overall feature extraction capability of the model.

[0088] This embodiment utilizes an intervention strategy generation and feedback correction component to achieve a closed-loop adaptive optimization mechanism from risk warning to nursing intervention and then to model correction. By fine-tuning the classification model based on the error between actual and predicted indicators after intervention, the accumulation of prediction bias can be suppressed, enabling the model to adapt to individual differences and changes in the condition of different patients, and consistently maintain high warning accuracy.

[0089] This invention further enhances the robustness and reliability of the system's early warning by preprocessing time-series data, dynamically evolving edge weights, preoperative baseline alignment compensation, and intervention feedback correction. Addressing the common issues of missing and outlier time series indicators, the invention cleanses the original data using missing value imputation and outlier smoothing operations, based on cubic spline interpolation and sliding window local statistics, avoiding interference from dirty data in the feature extraction process. The edge weights of the dynamic spatiotemporal graph undergo dynamic decay and state transition updates based on time decay factors and state transition matrices as the timestamp progresses. Combined with the half-life of postoperative inflammatory markers and the duration of physiological conduction delays in Crohn's disease, the evolution of the graph structure closely reflects the actual physical laws of pathological evolution. The preoperative baseline alignment compensation component extracts the preoperative historical test feature baseline, obtains an orthogonal transformation matrix through covariance matrix eigenvalue decomposition, and maps the preoperative baseline to the postoperative dynamic space to obtain the baseline compensation vector, providing individualized reference benchmarks for early postoperative data. The intervention strategy generation and feedback correction component generates nursing intervention strategy vectors based on the probability distribution of complication risk, and performs fine-tuning correction on the weight parameters of the fully connected classification sub-component based on the root mean square error of the actual test indicators and predicted indicator sequences after intervention, forming a closed-loop adaptive optimization mechanism that suppresses the accumulation of prediction bias.

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A postoperative care risk warning system for Crohn's disease, characterized in that, The system includes: Multimodal temporal feature extraction component: used to encode the time series of laboratory indicators, image pixel distribution and nursing text after Crohn's disease surgery, and output temporal embedding vectors of each modality; Dynamic graph construction component: used to construct a dynamic spatiotemporal graph of patient status that evolves over time, with timestamps as nodes and the pathological association strength between modalities as edge weights. The pathological association strength refers to the pathological association strength between different modalities at the same timestamp. The node features are updated through a graph attention mechanism to aggregate cross-modal pathological information. Graph Attention Risk Classification Component: Used to calculate the probability distribution of complication risk based on the node features of the updated dynamic spatiotemporal graph.

2. The postoperative nursing risk warning system for Crohn's disease according to claim 1, characterized in that, The multimodal temporal feature extraction component includes a temporal convolution sub-component, a visual coding sub-component, and a textual semantic sub-component; The temporal convolutional sub-component performs dilated causal convolution on the time series of the test index to extract temporal dependency features; The visual coding sub-component performs spatial feature mapping on the image pixel distribution to obtain image spatial features; The text semantic sub-component performs word embedding transformation on the nursing text to extract text semantic features; The multimodal temporal feature extraction component performs heterogeneous feature alignment and concatenation of the temporal dependency features, the image spatial features, and the text semantic features to generate the temporal embedding vectors of each modality.

3. The postoperative nursing risk warning system for Crohn's disease according to claim 2, characterized in that, Before performing the dilated causal convolution operation, the temporal convolution sub-component also performs missing value imputation and outlier smoothing operations. The missing value imputation operation is based on the known data points before and after the missing position in the time series of the test index, which are constructed using a cubic spline interpolation function to fill the missing value. The outlier smoothing operation calculates the local mean and local standard deviation of the sliding window of the test index time series. Data points that deviate from the local mean of the sliding window by more than a preset multiple of the local standard deviation are identified as outliers, and the outliers are replaced with the local mean of the sliding window. The processed sequence is then input into the dilated causal convolution operation.

4. The postoperative nursing risk warning system for Crohn's disease according to claim 1, characterized in that, The dynamic graph construction component includes an association computation subcomponent and a graph structure generation subcomponent; The correlation calculation subcomponent calculates the Pearson correlation coefficient of any two modes at the same timestamp based on the temporal embedding vector of each mode, maps the Pearson correlation coefficient to a Gaussian kernel function value, and obtains the initial correlation degree between modes. The graph structure generation sub-component compares the initial correlation degree between the modalities with a preset pathological correlation threshold, retains the correlation degree greater than the preset pathological correlation threshold as the edge weight, and constructs edges between the corresponding nodes to form a dynamic spatiotemporal graph of patient status that evolves over time.

5. A postoperative nursing risk warning system for Crohn's disease according to claim 4, characterized in that, The edge weights of the patient's state dynamic spatiotemporal graph undergo dynamic decay and state transition updates as the timestamps progress, specifically including: Based on adjacent timestamps, the edge weight at the current time is equal to the edge weight at the previous time multiplied by the time decay factor, plus the initial correlation degree between the modes calculated at the current time multiplied by the state transition matrix. The time decay factor is set according to the half-life of inflammatory markers after Crohn's disease surgery, and the state transition matrix is ​​set according to the physiological conduction delay between different modes.

6. A postoperative nursing risk warning system for Crohn's disease according to claim 1, characterized in that, The step of updating node features using a graph attention mechanism specifically includes the following steps: Based on any target node in the dynamic spatiotemporal graph of the patient's state, obtain a set of neighboring nodes that are connected to the target node by an edge. The feature vector of the target node is concatenated with the feature vectors of each neighboring node in the set of neighboring nodes and then input into the activation function to calculate the attention coefficient between the target node and each of the neighboring nodes. The attention coefficients are normalized to obtain the attention weights; The feature vectors of each neighboring node are weighted and summed according to the attention weights to update the feature vector of the target node.

7. A postoperative nursing risk warning system for Crohn's disease according to claim 6, characterized in that, The dynamic graph construction component employs a multi-head graph attention mechanism to update the node features, specifically including the following steps: The input features of the target node and the set of neighboring nodes are respectively input into multiple groups of graph attention mechanisms, and the graph attention mechanisms of each group calculate and output multi-head update features in parallel; All the multi-head update features are concatenated along the feature dimension and then linearly transformed to obtain the fused update features; The fused updated features are added element-wise to the original feature vector of the target node using a residual join operation, and the residual join result is subjected to layer normalization processing to output the final feature vector of the target node.

8. A postoperative nursing risk warning system for Crohn's disease according to claim 1, characterized in that, The graph attention risk classification component includes a spatiotemporal feature pooling sub-component and a fully connected classification sub-component. The spatiotemporal feature pooling sub-component performs max pooling operation on the node features of the updated patient state dynamic spatiotemporal graph along the time axis to extract the target global temporal features; The fully connected classification subcomponent maps the target global temporal features to the complication category space after Crohn's disease surgery, and outputs the complication risk probability distribution through a normalized exponential function. The complication risk probability distribution includes the anastomotic leakage risk probability, the abdominal infection risk probability, and the postoperative bleeding risk probability.

9. A postoperative nursing risk warning system for Crohn's disease according to claim 1, characterized in that, It also includes a preoperative baseline alignment compensation component, which is used to extract the historical test feature baseline of Crohn's disease patients before surgery, calculate the covariance matrix between the historical test feature baseline and the temporal embedding vector of each modality, and perform eigenvalue decomposition on the covariance matrix to obtain the orthogonal transformation matrix. The historical test feature baseline is mapped to the postoperative dynamic space using the orthogonal transformation matrix to obtain the baseline compensation vector; The graph attention risk classification component calculates the probability distribution of complication risk after fusing the baseline compensation vector with the node features.

10. A postoperative care risk warning system for Crohn's disease according to claim 1, characterized in that, It also includes an intervention strategy generation and feedback correction component, which generates a corresponding nursing intervention strategy vector based on the complication risk probability distribution; After executing the nursing intervention strategy vector, the actual test index time series of the next time stamp is obtained, and the predicted test index series of the next time stamp is predicted based on the node features of the current time and the nursing intervention strategy vector. The root mean square error between the actual test index time series and the predicted test index series is calculated. If the root mean square error exceeds a preset correction threshold, the gradient is calculated based on the root mean square error and backpropagated to the fully connected classification sub-component in the graph attention risk classification component. The weight parameters of the fully connected classification sub-component are then fine-tuned.

Images