Cardiothoracic surgery postoperative complication prediction method based on deep learning algorithm

Through the fusion of deep learning algorithms and multimodal data, a graph structure model is constructed, which solves the problems of single modal data dependence and high technical threshold in the existing technology, and achieves more accurate postoperative complication prediction and personalized risk assessment. Improved the quality of patient care.

CN119028567AInactive Publication Date: 2024-11-26HENAN CANCER HOSPITAL

Patent Information

Application Number
CN202411038774.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing cardiothoracic surgery postoperative complication prediction methods rely on single modal data (CT image data), lack multimodal data fusion, does not consider individual patient factors, high technical threshold, limited equipment availability, and data processing security and privacy protection is complex, resulting in limitations and biases in predicted outcomes.

Method used

Deep learning algorithms are adopted, combined with optical character recognition, natural language processing, graph neural networks and transfer learning technology to build a graph structure model of multi-source data, and use multimodal data (including X-ray films, real-time physiological monitoring data and historical clinical data. ) conduct complication risk assessment and dynamically adjust model parameters through reinforcement learning to provide personalized prediction results.

Benefits of technology

More accurate prediction of postoperative complications is achieved, customized risk assessment is provided, unnecessary medical intervention risks are reduced, patient care quality is improved, and technical threshold is lowered to adapt to changes in different medical environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028567A_ABST
    Figure CN119028567A_ABST
Patent Text Reader

Abstract

The invention relates to a cardiothoracic surgery postoperative complication prediction method based on a deep learning algorithm. Extracting text data from the electronic medical record by using an optical character recognition technology, and performing entity recognition on the extracted text by using a natural language processing technology; secondly, a statistical method and a machine learning algorithm are combined, valuable biomarkers and clinical parameters are identified and predicted through feature selection, and a graph structure containing multi-source data is constructed; finely adjusting the graph neural network model by using a transfer learning technology, and optimizing the complication prediction performance of the graph neural network model in a specific cardiothoracic surgical operation type; and finally, deploying the model to an actual clinical environment, collecting postoperative patient data in real time, adjusting model parameters by applying a dynamic prediction adjustment mechanism based on reinforcement learning, and combining with a multi-modal data fusion technology, including postoperative X-ray films, real-time physiological monitoring data and historical clinical data, to obtain a multi-modal data fusion model. A complication risk assessment is provided through an integrated decision support system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for predicting postoperative complications of cardiothoracic surgery, and specifically to a method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm. Background Art

[0002] Among the methods currently used for predicting postoperative complications in cardiothoracic surgery, Chinese invention patent No.: 2020110300084 proposes a method, device and equipment for predicting postoperative complications of transcatheter aortic valve replacement (TAVI). The method collects characteristic data of the target patient and inputs these data into a preset TAVI postoperative complication prediction model to obtain prediction results of postoperative complications, including prediction results of pacemaker implantation after TAVI and prediction results of paravalvular leak after TAVI. Although the above invention has obvious advantages, there are still some shortcomings and disadvantages. First, the method mainly relies on the accuracy of anatomical indicators in CT image data, which means that any error in data acquisition, processing or interpretation directly affects the accuracy of the prediction results. Secondly, although it is relatively easy to collect CT data and is easily adopted by most medical institutions, this method requires a high technical threshold in data processing and model training. Steps such as extraction and selection of feature data, design and optimization of models require expertise in data science and machine learning, which limits the application of this prediction method in resource-limited environments. In addition, although the invention takes into account the key factor of the implantation depth of the artificial valve, it does not explicitly consider other factors that affect postoperative complications, including the patient's age, gender, medical history, and medication history. These factors also have an important impact on postoperative clinical outcomes, and failure to include these factors in the prediction model may lead to limitations and deviations in the prediction results.

[0003] In addition, the invention focuses on two complications after TAVI: pacemaker implantation and paravalvular leak. However, patients after TAVI face more types of complications, including vascular complications and neurological complications. The invention does not provide a comprehensive risk assessment of complications, which makes doctors lack sufficient information support when making comprehensive clinical decisions. Furthermore, although relying on single-modality data for prediction is effective enough in some cases, combining multimodal data including electrocardiogram and laboratory test results will provide more comprehensive disease monitoring and more accurate risk assessment. Therefore, the prediction model that relies solely on CT image data cannot fully utilize other valuable clinical information. In addition, the technology and equipment on which the invention relies are of limited availability in different regions, and high-quality CT scans and advanced data processing software are not widely available. The limitations of this technology and equipment hinder the widespread application of this prediction method. Finally, from the perspective of ethics and privacy, the processing and analysis of patients' sensitive medical data requires a high degree of security and compliance. The collection, storage and processing of data must comply with relevant laws and regulations to ensure the security and privacy of patient information. This requires medical institutions to invest additional resources and technologies to ensure data security, increasing the cost and complexity of implementing this prediction method. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:

[0006] S1. First, use optical character recognition technology to extract text data from electronic medical records, and apply natural language processing technology including BERT or GPT model to perform entity recognition on the extracted text to extract medical terms and parameters related to postoperative cardiothoracic surgery;

[0007] S2. Secondly, statistical methods and machine learning algorithms are combined to identify and predict valuable biomarkers and clinical parameters through feature selection, and a graph structure containing multi-source data is constructed, including the patient's historical medical records, physiological parameters, and surgical details, in order to control the training of the graph neural network model to capture the interaction between the data;

[0008] The construction of a graph structure containing multi-source data includes:

[0009] S2.1, where patients are represented as nodes in the graph, node attributes consist of the selected key features, and edges between nodes represent medical interactions or common medical conditions between patients;

[0010] S2.2, then apply the graph neural network GNN model, use the node and edge information to capture and learn the interaction between patient data, and predict the postoperative complications of cardiothoracic surgery based on the graph structure through model training control;

[0011] S2.3. Use real-time data to dynamically adjust the graph structure and GNN model parameters to reflect new clinical data and changes in patient status, control the continuous optimization of the prediction model and adapt to changes in the clinical environment;

[0012] S3. Use transfer learning techniques again to fine-tune the graph neural network model to optimize its performance in predicting complications in specific cardiothoracic surgical procedures.

[0013] S4. Finally, the model is deployed to the actual clinical environment to collect postoperative patient data in real time, and the dynamic prediction adjustment mechanism based on reinforcement learning is applied to adjust the model parameters. At the same time, multimodal data fusion technology, including postoperative X-rays, real-time physiological monitoring data and historical clinical data, is combined to provide complication risk assessment through an integrated decision support system.

[0014] Furthermore, the steps of combining statistical methods and machine learning algorithms are as follows: first, a hybrid feature selection framework is applied to jointly optimize features through statistical methods and machine learning algorithms, and the formula is:

[0015]

[0016] where X = {x1, x2, …, x n} represents the feature set, α i is feature x i The importance coefficient, β i A coefficient to adjust the intensity of feature influence;

[0017] Then, using multimodal data fusion technology, through the double-layer integration formula:

[0018]

[0019] Where φ(x,y,z)=k·sin(x·y·z) is the three-dimensional data fusion function, X,Y,Z represent the feature matrices from different data sources, and k,a,b,c,d are parameters for adjusting the integration range and intensity;

[0020] Finally, we build the model training of the end-to-end system using the formula:

[0021]

[0022] where Θ = {θ1,θ2,…,θ m} represents the model parameters, X={x1,x2,…,x m} is the input feature set.

[0023] Furthermore, the patient is represented as a node in a graph, and the construction method is as follows:

[0024] S1. Define each patient as a node in a graph. The node attributes are composed of the selected key features. The selection of features is based on the newly written formula:

[0025]

[0026] where x i represents the ith biomarker or clinical parameter, γ i It is a dynamically adjusted weight coefficient to evaluate the importance of each feature;

[0027] S2. Then construct the edges between nodes, representing the medical interactions or common medical conditions between patients. The weight of the edge is determined by the formula:

[0028]

[0029] where x i and x j are the feature vectors of connected nodes, and σ is a parameter to measure the sensitivity of feature similarity;

[0030] S3. Use graph neural network GNN to process structured data and use the formula to update nodes:

[0031]

[0032] Where X is the feature set of all nodes, is the neighborhood of node i, θ ij are the learned parameters, represents the merging operation of features, and σ is a nonlinear activation function.

[0033] Furthermore, the method of applying the graph neural network (GNN) model includes: first, in the graph construction stage, it is defined that each node in the graph represents a patient, the node attributes are composed of key biomarkers and clinical parameters, and the construction of the edge is defined by the formula:

[0034]

[0035] where x i (t) and x j (t) represents the state vectors of patients i and j at time t, respectively, simulating the dynamic medical relationship between patients;

[0036] Next, in the model training phase, the learning formula is used:

[0037]

[0038] Optimize, where Θ represents the model parameters, y i is the complication status of patient i, x k is the feature of the neighbor node, θ ik is the weight parameter.

[0039] Furthermore, the steps of dynamically adjusting the graph structure and GNN model parameters using real-time data are as follows:

[0040] S1. Use real-time data integration technology to dynamically update nodes and edges in the graph, and use the formula to calculate changes in the graph structure:

[0041]

[0042] ΔG(t) represents the change of the graph at time t, λ is the attenuation coefficient, δx i (s) represents the state change of node i at time s;

[0043] S2. Secondly, apply the online learning strategy to dynamically adjust the parameters of the GNN model and update the parameters using the formula:

[0044]

[0045] where θ(t) is the model parameter at time t, η is the learning rate, It is the gradient of the loss function L with respect to the parameter θ, and X represents the input feature data.

[0046] Furthermore, the transfer learning technology implementation method includes: first, using a pre-trained GNN model, and using a newly written fine-tuning formula:

[0047]

[0048] Adjust the model, where η is the learning rate, l is the loss function, and y i is the true label, f(x i ; θ) is the model prediction, x i It is patient data for a specific type of surgery;

[0049] Then, the performance of the fine-tuned model is evaluated using the newly written evaluation formula:

[0050]

[0051] The prediction accuracy in the actual clinical setting was measured, where y is the actual occurrence of complications, f(x) is the model prediction, the probability of complication occurrence, σ represents the standard deviation of the prediction error, and a and b are the upper and lower limits of the integral.

[0052] Furthermore, the integrated decision support system construction method is as follows:

[0053] S1. Use reinforcement learning algorithms to perform real-time feedback adjustments on the graph neural network (GNN) model, and continuously optimize model parameters to adapt to real-time changes in the clinical environment.

[0054] S2, realize multimodal data fusion, integrate postoperative X-rays, real-time physiological monitoring data and historical clinical data to provide a comprehensive view of patient health;

[0055] S3. Integrate the prediction model with the hospital’s decision support system to provide real-time, data-driven medical decision support.

[0056] Furthermore, the reinforcement learning algorithm construction method includes: first defining a reward function to evaluate the accuracy of the prediction results of the GNN model, using the formula:

[0057]

[0058] Where R represents the cumulative reward, λ is the discount factor, P(t) is the predicted probability, A(t) is the actual probability of occurrence, and α is the smoothing parameter;

[0059] Secondly, the GNN model is embedded in the reinforcement learning environment, predicting actions based on the current strategy and receiving rewards based on the results. The model update formula is:

[0060]

[0061] Where θ(t) is the model parameter at time t, η is the learning rate, is the loss function, y(t) is the true label, and f(x(t);θ) is the model prediction value.

[0062] Furthermore, the multimodal data fusion implementation method is as follows:

[0063] S1. Preprocess and standardize the data of different modalities using the formula:

[0064]

[0065] D(x) represents the standardized data, I(x) represents the original data, N(x) represents the noise, γ is the smoothing parameter, and sin(ωt) is the periodic noise correction term;

[0066] S2. Use convolutional neural network (CNN) to extract high-level features from postoperative X-rays, use recurrent neural network (RNN) or long short-term memory (LSTM) to extract time series features from real-time physiological monitoring data, and use multi-layer perceptron (MLP) to extract key features from historical clinical data, using the formula:

[0067]

[0068] Where T(x) represents the extracted features, δ i is the weight, φ i (x) represents the features extracted by different networks, is the gradient term of the feature, and α is the attenuation parameter;

[0069] S3, feature fusion, using the formula:

[0070]

[0071] where f i (x) represents the i-th feature, α i is the weight coefficient, g i (y) represents the characteristics of the additional data source, β i is the fusion weight, H(y) represents the noise, ∈ is the fusion smoothing parameter, is the gradient term of the feature, and λ is the attenuation parameter.

[0072] Furthermore, the prediction model is integrated with the hospital's decision support system as follows:

[0073] First, system integration and interface design, design and development of application program interface API, to achieve data exchange and processing through formulas:

[0074]

[0075] Where ΔI(t) represents the data change at time t, D(τ) is the data input, M(τ) is the processing model, and β and λ are adjustment parameters;

[0076] Secondly, real-time data processing and analysis uses a streaming data processing framework to process data input in real time from various clinical data sources. The formula is:

[0077]

[0078] Where P(t) represents the processed data, S(τ) is the input data at time τ, and κ and μ are adjustment parameters;

[0079] Finally, the output of the prediction model is integrated into the hospital’s decision support system through the formula:

[0080]

[0081] D(t) represents the output of the decision support system, f i (t) and g i (t) represents the output of the prediction model and the output of the decision suggestion module, respectively, i and βi is the weight coefficient.

[0082] Beneficial effects of the present invention:

[0083] 1. By using advanced deep learning models, the present invention is able to learn complex patterns and relationships from large amounts of clinical data, providing more accurate prediction results than traditional statistical methods. This accuracy is crucial for identifying high-risk patients and taking preventive measures in a timely manner, helping to reduce postoperative complications and improve patient survival rates.

[0084] 2. By analyzing individual patient data, the present invention can provide customized risk assessments for each patient, thereby guiding personalized treatment plans. This approach takes into account the patient's unique physiological characteristics and clinical history, helping to optimize treatment outcomes and reduce the risk of unnecessary medical interventions.

[0085] 3. This invention not only covers data science and machine learning technology, but also closely integrates with clinical medicine. It provides a platform for medical professionals from different backgrounds to work together, promote interdisciplinary collaboration, and jointly improve the quality of patient care.

[0086] 4. Deep learning-based models can be easily integrated into existing hospital information systems without large-scale changes to existing IT infrastructure. In addition, the flexibility and scalability of this approach allow hospitals to add new data sources and functions as needed to meet changing medical needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 This is a flow chart of the method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm of the present invention.

[0088] Figure 2 This is a module diagram of the present invention combining statistical methods and machine learning algorithms.

[0089] Figure 3 A flow chart is constructed for the decision support system integrated with the present invention.

[0090] Figure 4 This is a module diagram of the method for integrating the prediction model of the present invention with the hospital's decision support system. DETAILED DESCRIPTION

[0091] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0092] Combined with Figure 1,The postoperative complication prediction method for cardiothoracic surgery based on deep learning algorithm first extracts text data from the electronic medical record using optical character recognition technology, and then applies natural language processing technology, including BERT or GPT model, to perform entity recognition on the extracted text, which can accurately extract medical terms and parameters related to postoperative cardiothoracic surgery;

[0093] Next, feature selection is performed by combining statistical methods and machine learning algorithms to identify biomarkers and clinical parameters that are valuable for prediction, and then a graph structure containing multi-source data is constructed. This graph structure includes the patient's medical history, physiological parameters, and surgical details, so that the graph neural network GNN model can control and capture the interactions between these data during training; in this graph structure, patients are represented as nodes, and the attributes of the nodes are composed of the selected key features. The edges between the nodes represent the medical interactions or shared medical conditions between patients. The GNN model then uses the information of these nodes and edges to learn the interactions between patient data, and predict complications after cardiothoracic surgery through model training;

[0094] In addition, the system can use real-time data to dynamically adjust the graph structure and GNN model parameters to adapt to new clinical data and changes in patient status, ensuring that the prediction model is continuously optimized and adapts to changes in the clinical environment. Furthermore, the GNN model is fine-tuned through transfer learning technology to optimize the model's performance in predicting complications in specific types of cardiothoracic surgery.

[0095] Finally, the model was deployed in an actual clinical environment to collect postoperative patient data in real time, and the model parameters were adjusted in combination with a dynamic prediction adjustment mechanism based on reinforcement learning. At the same time, multimodal data fusion technology was used to integrate postoperative X-rays, real-time physiological monitoring data, and historical clinical data to provide accurate complication risk assessment through an integrated decision support system.

[0096] Embodiment 1:

[0097] Combined with Figure 2 , the embodiment uses a method based on deep learning algorithm to predict postoperative complications of cardiothoracic surgery. First, a hybrid feature selection framework is applied to optimize features by combining traditional statistical methods and machine learning techniques. The feature weight formula is used to quantify the importance and influence of each feature:

[0098]

[0099] X={x1,x2,…,x n} is the feature set extracted from the patient data, where each x i represents a specific clinical parameter or biomarker. i and β i Represents the feature xi The importance and influence adjustment coefficient.

[0100] Three key features are selected from the patient data: heart rate (x1), white blood cell count (x2), and body temperature (x3). The importance coefficient α is set to {0.3, 0.5, 0.2}, reflecting the relative importance of each feature to the prediction result. The adjustment coefficient β is set to {2, 1, 0.5}, indicating that heart rate is the most sensitive to the prediction result and body temperature is the least sensitive.

[0101] Assume that the patient's heart rate x1 = 100 beats / minute, white blood cell count x2 = 12,000 cells / ml, and body temperature x3 = 38.5°C. Calculate as follows:

[0102] exp(β1·x1)=exp(2·100)=e 200

[0103] exp(β2·x2)=exp(1·12000)=e 12000

[0104] exp(β3·x3)=exp(0.5·38.5)=e 19.25

[0105] Sumofexp=e 200 +e 12000 +e 19.25

[0106]

[0107] Due to e 12000 It is very large relative to the other two items and almost determines the value of the denominator, so the weight of the white blood cell count dominates this calculation and has the greatest impact on the F(X) value.

[0108] The embodiment continues to use multimodal data fusion technology and selects the postoperative data of patient A for analysis. The postoperative data of patient A includes postoperative X-rays, real-time physiological monitoring data of heart rate and blood pressure, and the patient's historical medical records. The double-layer integral formula is used to fuse these data:

[0109]

[0110] Among them, φ(x,y,z)=k·sin(x·y·z) is the three-dimensional data fusion function, X, Y, Z represent the feature matrices from different data sources, and k, a, b, c, d are parameters for adjusting the integration range and intensity.

[0111] Assume that the postoperative X-ray film (X) of patient A has been extracted through the convolutional neural network CNN to obtain a feature matrix containing the main features (x1, x2, x3). At the same time, the heart rate and blood pressure physiological monitoring data (Y) are processed to obtain a feature matrix (y1, y2, y3). The historical clinical data (Z) is extracted through a multi-layer perceptron (MLP) to obtain a feature matrix (z1, z2, z3).

[0112] Set the parameters as follows:

[0113] k=1.5,a=0,b=1,c=0,d=1

[0114] In actual calculations, the team substituted these characteristics into the double-layer integral formula:

[0115] φ(x,y,z)=k·sin(x·y·z)

[0116]

[0117] The specific values ​​are:

[0118] x1=0.8,y1=0.9,z1=0.7

[0119] x2=0.5,y2=0.4,z2=0.6

[0120] x3=0.3,y3=0.7,z3=0.2

[0121] The specific calculation is as follows:

[0122]

[0123] After calculation, we get:

[0124]

[0125] The integral result is assumed to be approximately:

[0126] G(X,Y,Z)≈1.5·(0.504·1·1+0.12·1·1+0.042·1·1)

[0127] G(X,Y,Z)≈1.5·(0.504+0.12+0.042)

[0128] G(X,Y,Z)≈1.5·0.666

[0129] G(X,Y,Z)≈0.999

[0130] The fused eigenvalue G(X,Y,Z)≈0.999 is obtained.

[0131] The final embodiment trains a deep learning model to predict complications after cardiothoracic surgery based on the fused and optimized features. A nonlinear model is used to process the obtained features and output the probability of complications. The model expression is as follows:

[0132]

[0133] In this formula, θ={θ1,θ2,…,θ m} represents the model parameters, and X={x1,x2,…,x m} is the input feature set, including features fused from multimodal data.

[0134] Use the data features obtained from the previous step, including heart rate, white blood cell count, body temperature, and image features obtained through X-rays. The model contains three parameters and three input features, and the parameter value range is set to -2 to 2. Set θ1 = 1.5, θ2 = -0.5, θ3 = 1.0.

[0135] Set the patient's characteristic values ​​to x1 = 0.8, x2 = 0.6, x3 = 0.9. Calculate:

[0136]

[0137]

[0138] The calculated result H(Θ,X)≈0.73 indicates that the probability of the patient developing complications is approximately 73%.

[0139] Embodiment 2:

[0140] Embodiment In the method for predicting complications after cardiothoracic surgery based on a deep learning algorithm, each patient is defined as a node in a graph, and the attributes of the node are composed of the selected key features. The selection of features is based on the newly written formula:

[0141]

[0142] x i represents the i-th biomarker or clinical parameter, γ i It is a dynamically adjusted weight coefficient used to evaluate the importance of each feature.

[0143] Suppose a patient data set after cardiothoracic surgery includes the following biomarkers and clinical parameters: heart rate (x1), white blood cell count (x2), body temperature (x3). Set the weight coefficient γ = {0.5, 1.0, 0.7}, heart rate x1 = 80 bpm (beats / minute), white blood cell count x2 = 12000 cells / μL (cells / microliter), body temperature x3 = 38.5℃. Substitute these values ​​into the formula for calculation:

[0144] f(x1)=0.5·log(1+e 80 )

[0145] f(x2)=1.0·log(1+e 12000 )

[0146] f(x3)=0.7·log(1+e 38.5 )

[0147] Because e x grows rapidly when x is very large, resulting in log(1+e x ) is close to the value of x, so:

[0148] f(x1)≈0.5·80

[0149] f(x2)≈1.0·12000

[0150] f(x3)≈0.7·38.5

[0151] The calculation results show that the white blood cell count has the greatest weight in this case because its absolute value is very high.

[0152] The embodiment further constructs edges in the graph structure to represent medical interactions or shared medical conditions between patients. This is done by calculating the weights between each pair of nodes to express the mutual relationships and similarities between nodes (patients), using the following formula:

[0153]

[0154] x i and x j are the feature vectors of the connected nodes, respectively, containing the feature values ​​extracted from the data of each patient. The parameter σ is a parameter that measures the sensitivity of feature similarity and controls the degree of influence of feature differences on edge weights.

[0155] Assume that the feature vectors of patients A and B are x A =[0.8,0.6,0.9] and x B = [0.85, 0.55, 0.95], these values ​​represent the normalized features including heart rate, white blood cell count and body temperature. Set σ = 0.1.

[0156] Compute the weight of the edge between two patients:

[0157]

[0158] First, calculate the sum of the squares of the differences between the eigenvectors:

[0159] |xA -x B | 2 =(0.8-0.85) 2 +(0.6-0.55) 2 +(0.9-0.95) 2

[0160] =0.0025+0.0025+0.0025=0.0075

[0161] Then substitute this value into the formula:

[0162]

[0163] Calculation result w AB ≈0.687 indicates that there is a relatively strong similarity or medical connection between patients A and B.

[0164] The last example uses GNN to update each node (patient) by representing the node's own features and the features of its neighboring nodes. The update formula is as follows:

[0165]

[0166] X is the feature set of all nodes, is the neighborhood of node i (i.e., the set of nodes directly connected to node i), θ ij is the learned parameter used to adjust the contribution of node j to the update of node i, Represents the feature merging operation, σ is the nonlinear activation function, which is the Sigmoid function.

[0167] Set the feature vectors of the three patient nodes as follows, as well as the connections between them:

[0168] x1=[0.8,0.6]

[0169] x2=[0.85,0.55]

[0170] x3=[0.75,0.65]

[0171] Assume that the neighbors of node 1 include nodes 2 and 3, and the parameter θ 12 and θ 13 Set to 0.5 and 0.3, respectively, to represent the influence of node 2 and node 3 on node 1. The nonlinear activation function is selected as Sigmoid.

[0172] According to the GNN update formula, the update process of node 1 is as follows:

[0173]

[0174] First, perform feature merging (here we simply use addition as the merging operation):

[0175]

[0176] Then do the weighted sum:

[0177] 0.5 · [1.65, 1.15] + 0.3 · [1.55, 1.25] = [0.825, 0.575] + [0.465, 0.375] = [1.29, 0.95]

[0178] Finally, the Sigmoid activation function is applied:

[0179]

[0180] The result is [0.784, 0.721], indicating that after GNN updating, the feature vector of node 1 becomes [0.784, 0.721], representing the new feature state of patient 1, which integrates its own information and the information of other patients directly connected to it.

[0181] Embodiment 3:

[0182] In the embodiment, when predicting complications after cardiothoracic surgery, the graph construction stage is first performed. Each patient is defined as a node in the graph, and the attributes of the node include key biomarkers and clinical parameters. The edges in the graph are defined by formulas to simulate the dynamic medical relationship between patients:

[0183]

[0184] x i (t) and x j (t) represent the state vectors of patients i and j at time t, respectively.

[0185] The status of the two patients at several points in the day, including heart rate and body temperature, are recorded as follows:

[0186] Patient A’s state vector: x A (0) = [80, 37.5], x A (1) = [78, 37.0]

[0187] Patient B’s state vector: x B (0) = [76, 36.5], x B (1) = [75, 36.0]

[0188] Compute the weight of the edge between two patients:

[0189]

[0190] Simplify the calculation process and set the internal result of the integral to the average value:

[0191] |x A (t)-x B (t)| 2 =[4 2 +1 2 +3 2 +1 2 ] / 2=10.5

[0192] e AB =exp(-10.5)≈0.0000253

[0193] Calculation result e AB ≈0.0000253 shows that the status difference between patients A and B is large, so the mutual influence between them is small.

[0194] The embodiment further adopts the following learning formula to optimize the model parameters:

[0195]

[0196] Θ represents the set of model parameters, y i is the complication status of patient i (1 indicates complication, -1 indicates no complication), x k is the feature vector of the neighbor node of patient i, θ ik is the weight parameter connecting patient i and its neighbor k.

[0197] Assume that the hospital network contains 3 patients:

[0198] Patient 1 has complications, y1 = 1;

[0199] Patient 2 had no complications, y2 = -1;

[0200] Patient 3 has complications, y3=1.

[0201] The feature vectors of each patient and the relationship between them are as follows:

[0202] x1=[0.8,0.6],x2=[0.85,0.55],x3=[0.75,0.65]

[0203] Setting θ 12 =0.5,θ 13 =0.3,θ 23 =0.4

[0204] The learning formula portion for patient 1 was calculated as follows:

[0205]

[0206] L(Θ)=-log(1+exp(-1·(0.65+0.47)))≈-log(1+exp(-1.12))

[0207] ≈-log(1+0.326)=-log(1.326)≈-0.282

[0208] The calculation results show that the model successfully reflects that patient 1 has a higher probability of complication risk, indicating that the model effectively uses the information of neighbor nodes and the patient's own status to make predictions.

[0209] Embodiment 4:

[0210] In the cardiothoracic surgery postoperative complication prediction method of the embodiment, real-time data integration technology is used to realize the dynamic update of nodes and edges in the graph, and the following formula is used to calculate the change of the graph structure:

[0211]

[0212] ΔG(t) represents the change in the graph at time t, λ is the attenuation coefficient, which is used to simulate the decreasing effect of the past state on the current state; δx i (s) represents the state change of node i at time s.

[0213] Assume there are 3 patients in the hospital network, and their status is recorded over time as follows:

[0214] The state change δx1(t) of patient 1 changes from time t = 0 to t = 1 to [0.1, 0.2, 0.1];

[0215] The state change δx2(t) of patient 2 changes from time t = 0 to t = 1 to [0.05, 0.1, 0.05];

[0216] The state change δx3(t) of patient 3 changes to [0.2, 0.3, 0.2] between time t=0 and t=1.

[0217] The attenuation coefficient λ is set to 0.1.

[0218] Calculate the integrated value of the graph structure change:

[0219]

[0220] Set the simplified calculation, use the numerical integration method, and set δx i (s) remains constant over the entire interval:

[0221] ΔG(1)=e -0.1(1-0) [0.1+0.2+0.1]+e -0.1(1-0)[0.05+0.1+0.05]+e -0.1(1-0) [0.2+0.3+0.2]

[0222] =e -0.1 [0.4+0.2+0.7]

[0223] =e -0.1 1.3

[0224] ≈0.886·1.3=1.152

[0225] The calculated result ΔG(1)≈1.152 shows that the graph structure has changed significantly at time t=1, indicating that the patient's status has been significantly updated.

[0226] The embodiment further introduces an online learning strategy to dynamically adjust the parameters of the graph neural network (GNN) model and update it in real time according to new data, thereby improving its adaptability and prediction accuracy. The following update formula is used:

[0227]

[0228] θ(t) is the model parameter at time t, η is the learning rate, It is the gradient of the loss function L with respect to the parameter θ, and X represents the input feature data.

[0229] The model processes data from a patient population, including:

[0230] The patient's characteristic data X includes physiological indicators such as heart rate and blood pressure;

[0231] Initial parameters θ(0) = [0.5, -0.3];

[0232] Set the learning rate η = 0.1.

[0233] Assume that at time t, the gradient of the loss function L with respect to θ Calculated as [0.2,-0.1].

[0234] According to the parameter update formula of online learning, the parameters are updated as follows:

[0235] θ(1)=[0.5,-0.3]-0.1×[0.2,-0.1]

[0236] =[0.5-0.02,-0.3+0.01]

[0237] =[0.48,-0.29]

[0238] The updated parameters θ(1) = [0.48, -0.29] reflect that the model has adjusted its learning direction according to the latest data.

[0239] Embodiment 5:

[0240] In the cardiothoracic surgery postoperative complication prediction method based on deep learning algorithm, transfer learning technology is used to improve the prediction performance of the model. First, the pre-trained GNN model is used and the model is adjusted by the following fine-tuning formula:

[0241]

[0242] η is the learning rate, l is the loss function, y i is the true label, f(x i ; θ) is the model prediction, x i It is the patient data of a specific surgery type.

[0243] A GNN model was pre-trained on postoperative cardiothoracic surgery data with the following initial parameters:

[0244] Initial model parameters θ = [0.4, -0.6, 0.8];

[0245] The learning rate η=0.01 is used.

[0246] Collect a set of patient data of a specific type of surgery. The dataset contains the following groups of samples: patient 1's data x1 = [0.9, 0.8, 0.7], true label y1 = 1;

[0247] Patient 2’s data x2=[0.6, 0.7, 0.8], and the true label y2=0.

[0248] The loss function l uses the mean square error (MSE):

[0249] l(y i ,f(x i ; θ))=(y i -f(x i ;θ)) 2

[0250] First, calculate the model's predicted output f(x i ; θ), set the model prediction function to:

[0251] f(x i ; θ) = θ1x i1 +θ2x i2 +θ3x i3

[0252] For patient 1:

[0253] f(x1;θ)=0.4·0.9-0.6·0.8+0.8·0.7

[0254] =0.36-0.48+0.56=0.44

[0255] For patient 2:

[0256] f(x2;θ)=0.4·0.6-0.6·0.7+0.8·0.8

[0257] =0.24-0.42+0.64=0.46

[0258] Then calculate the gradient of the loss function:

[0259] For patient 1:

[0260]

[0261] For patient 2:

[0262]

[0263] Combine these gradients together to perform a parameter update:

[0264]

[0265] Update model parameters:

[0266] θ new =θ-Δθ

[0267] =[0.4,-0.6,0.8]-[0.00456,0.00252,0.00048]

[0268] =[0.39544,-0.60252,0.79952]

[0269] Through transfer learning techniques, the model parameters are updated to θ new =[0.39544,-0.60252,0.79952], which optimizes the model on patient data of a specific surgical type.

[0270] Next, the embodiment performs a performance evaluation on the fine-tuned model to ensure its prediction accuracy in an actual clinical setting, using the evaluation formula:

[0271]

[0272] y is the actual occurrence of complications, f(x) is the probability of complication occurrence predicted by the model, σ represents the standard deviation of the prediction error, and a and b are the upper and lower limits of the integral.

[0273] The setting team collected the following patient data for model performance evaluation:

[0274] The actual complication of patient 1 is y1=1, and the model prediction is f(x1)=0.85;

[0275] The actual complication occurrence of patient 2 is y2=0, and the model predicts f(x2)=0.15.

[0276] The standard deviation of the prediction error is set as σ = 0.1; the upper and lower limits of the integral are set as a = 0, b = 1.

[0277] For patient 1:

[0278]

[0279] For patient 2:

[0280]

[0281]

[0282] The calculated results P1≈0.3247 and P2≈0.3247 indicate that the fine-tuned model has high prediction accuracy in actual clinical settings.

[0283] Embodiment 6:

[0284] In the development of a deep learning-based method for predicting postoperative complications in cardiothoracic surgery, a reinforcement learning algorithm is applied to perform real-time feedback adjustment on the graph neural network (GNN) model. A reward function is defined to evaluate the accuracy of the prediction results of the GNN model at a specific time point:

[0285]

[0286] R represents the cumulative reward, λ is the discount factor, which adjusts the impact of time on the reward; P(t) is the predicted probability of the model at time t; A(t) is the actual probability of occurrence at the same time; α is a smoothing parameter, which is used to adjust the impact of incorrect predictions on the reward.

[0287] Setting The following sample data were collected in a cardiothoracic surgery clinical setting for evaluation:

[0288] The predicted probability P(t) is recorded as 0.9, 0.85, and 0.95 in a time period;

[0289] The actual occurrence probability A(t) is 1, 0, 1.

[0290] Setting parameters:

[0291] Discount factor λ = 0.1;

[0292] Smoothing parameter α = 0.05;

[0293] The time interval T is 3, corresponding to the data at three time points.

[0294] Calculate the integral of the reward function, using a simplified method of numerical integration, setting the time points to be equally spaced:

[0295]

[0296] Calculate the specific value:

[0297]

[0298] The cumulative reward R≈0.086+0.693+0.077=0.856. This result shows that the model performs well in a specific evaluation period, taking into account time discount and prediction error.

[0299] In the next step, the graph neural network GNN model is embedded in the reinforcement learning environment, so that the model can perform predictive operations according to the current strategy, receive rewards and perform parameter updates based on the results, and use the following model update formula:

[0300]

[0301] θ(t) is the model parameter at time t, η is the learning rate, is the loss function, y(t) is the true label, and f(x(t);θ) is the model prediction value.

[0302] Time series data of patient 1: x1(t) = [0.8, 0.7, 0.9], true label: y1(t) = [1, 0, 1];

[0303] Time series data of patient 2: x2(t)=[0.6,0.5,0.7], true label: y2(t)=[0,0,1].

[0304] Set θ(t=0)=[0.3,-0.2,0.4]; learning rate η=0.01

[0305] Setting the loss function Use mean square error (MSE):

[0306]

[0307] First, calculate the model's predicted value at time t: f(x(t);θ):

[0308] Set f(x(t);θ)=θ1x t1 +θ2x t2 +θ3x t3 , for patient 1 at time t=1:

[0309] f(x1(1);θ)=0.3·0.8+(-0.2)·0.7+0.4·0.9=0.24-0.14+0.36=0.46

[0310] Calculate the gradient of the loss function:

[0311]

[0312] Update the parameters according to the update formula:

[0313]

[0314] Set the integral to simplify to the current gradient:

[0315] θ(t+1)=[0.3,-0.2,0.4]-0.01·[-0.432,-0.378,-0.486]

[0316] =[0.3+0.00432,-0.2+0.00378,0.4+0.00486]

[0317] =[0.30432,-0.19622,0.40486]

[0318] The model parameters are updated to θ(t+1)=[0.30432,-0.19622,0.40486]. This update mechanism continuously adjusts the model parameters to optimize its predictive ability in a constantly changing clinical data environment.

[0319] Embodiment 7:

[0320] View attached Figure 3 In the embodiment, in the method for predicting complications after cardiothoracic surgery based on deep learning algorithm, multimodal data fusion is realized, and data of different modalities are preprocessed and standardized. The following formula is used for preprocessing and standardization:

[0321]

[0322] D(x) represents the standardized data, I(x) represents the original data, N(x) represents the noise, γ is the smoothing parameter, and sin(ωt) is the periodic noise correction term.

[0323] The following three modes of data are set:

[0324] 1. Postoperative X-ray film data (image data): set the original data I(x1) = 100, noise N(x1) = 10;

[0325] 2. Real-time physiological monitoring data (heart rate): set the original data I(x2) = 80, noise N(x2) = 5;

[0326] 3. Historical clinical data (blood pressure): Set the original data I(x3)=120 and the noise N(x3)=8.

[0327] Set the parameters as follows:

[0328] Smoothing parameter γ = 0.1;

[0329] Periodic noise correction term parameter ω = 2;

[0330] The upper and lower limits of the integral are a=0, b=1;

[0331] The internal integral upper and lower limits are c=0, d=π.

[0332] Preprocess and normalize the data for each modality:

[0333] 1. For postoperative X-ray data:

[0334]

[0335] 2. For real-time physiological monitoring data (heart rate):

[0336]

[0337] 3. For historical clinical data (blood pressure):

[0338]

[0339] The data after preprocessing and standardization are:

[0340] Postoperative X-ray data: D(x1)≈33.33;

[0341] Real-time physiological monitoring data (heart rate): D(x2)=32;

[0342] Historical clinical data (blood pressure): D(x3)≈42.86.

[0343] The embodiment further extracts high-level features from data of different modalities using the formula:

[0344]

[0345] T(x) represents the extracted features, δ i is the weight, φ i (x) represents the features extracted by different networks, is the gradient term of the feature, and α is the attenuation parameter.

[0346] The setting team collected the following data in a clinical setting:

[0347] 1. Postoperative X-ray data (image data), using CNN to extract features;

[0348] 2. Real-time physiological monitoring data (heart rate, blood pressure), using LSTM to extract time series features;

[0349] 3. Historical clinical data (patient medical history, treatment records), using MLP to extract key features.

[0350] Set the weight δ = [0.4, 0.3, 0.3]; the attenuation parameter α = 0.5.

[0351] Postoperative X-ray data feature extraction:

[0352] Use CNN to extract features, and set the extracted features to φ1(x)=[10,20,30];

[0353] Real-time physiological monitoring data feature extraction:

[0354] Use LSTM to extract time series features, and set the extracted features to φ2(x)=[0.8,0.7,0.9];

[0355] Historical clinical data feature extraction:

[0356] Use MLP to extract key features, and set the extracted features to φ3(x) = [120, 130, 140].

[0357] Apply the formula to calculate the comprehensive characteristics:

[0358]

[0359] First, the weighted sum of the features extracted by each network is calculated:

[0360] δ1·φ1(x)=0.4·[10,20,30]=[4,8,12]

[0361] δ2·φ2(x)=0.3·[0.8,0.7,0.9]=[0.24,0.21,0.27]

[0362] δ3·φ3(x)=0.3·[120,130,140]=[36,39,42]

[0363] Set the feature gradient to:

[0364]

[0365] Points calculation:

[0366]

[0367] Similarly:

[0368]

[0369] Combining all the features:

[0370] T(x)=[4,8,12]+[0.24,0.21,0.27]+[36,39,42]+[1.27,1.27,1.27]+[0.127,0.127,0.127]

[0371] +[0.635,0.635,0.635]

[0372] =[42.272,49.242,56.302]

[0373] The comprehensive feature T(x) = [42.272, 49.242, 56.302] reflects the high-level features extracted from multimodal data.

[0374] The final embodiment uses a feature fusion formula to integrate different types of data, using the following formula:

[0375]

[0376] f i (x) represents the i-th feature, α i is the weight coefficient, g i (y) represents the characteristics of the additional data source, β i is the fusion weight, H(y) represents the noise, ∈ is the fusion smoothing parameter, is the gradient term of the feature, and λ is the attenuation parameter.

[0377] The features extracted from different modalities are set as follows:

[0378] Postoperative X-ray feature f1(x) = [42.272, 49.242, 56.302];

[0379] Real-time physiological monitoring features f2(x) = [32, 29, 35];

[0380] Historical clinical data features f3(x) = [42.86, 41.5, 43.75].

[0381] The characteristics of the additional data source laboratory test results are set as:

[0382] g1(y)=[1.2,1.3,1.1]

[0383] g2(y)=[0.9,0.85,0.95]

[0384] Setting parameters: weight coefficient α = [0.4, 0.3, 0.3]; fusion weight β = [0.6, 0.4]; noise H(y) = 0.1; smoothing parameter ∈ = 0.05; attenuation parameter λ = 0.2; integral upper and lower limits c = 0, d = 1.

[0385] Weighted sum:

[0386]

[0387] Integral calculation for additional data sources:

[0388]

[0389] Set the feature gradient to:

[0390]

[0391]

[0392] Combine all features:

[0393] F(x)=[39.3668,40.8468,46.1458]+[0.7164,0.7761,0.6567]+[0.3582,0.3373,0.3781]

[0394] +[0.0454,0.0454,0.0454]+[0.0227,0.0227,0.0227]

[0395] =[40.5095,42.0283,47.2487]

[0396] Through the feature fusion formula, the comprehensive feature F(x) = [40.5095, 42.0283, 47.2487] integrates the important information in the multimodal data.

[0397] Embodiment 8:

[0398] Combined with Figure 4 In the embodiment, in the method for predicting complications after cardiothoracic surgery based on deep learning algorithm, system integration and interface design are performed, including designing and developing an application program interface API, and using the following formula to realize data exchange and processing:

[0399]

[0400] ΔI(t) represents the data change at time t, D(τ) is the data input, M(τ) is the processing model, and β and λ are adjustment parameters.

[0401] Data input D(τ) includes the patient’s real-time heart rate and blood pressure.

[0402] The processing model M(τ) includes a patient health status prediction model that is updated in real time.

[0403] Set the adjustment parameter β=0.5; the attenuation parameter λ=0.1.

[0404] The data input and processing model is as follows:

[0405] Assume that the data input D(τ) increases linearly in the time period [0, t], D(τ) = 2τ;

[0406] The processing model M(τ) is a constant, M(τ)=5.

[0407] Calculate ΔI(t) using the formula:

[0408]

[0409]

[0410] Considering e -0.1(t-τ) The integral is Compute the integral:

[0411]

[0412] For t = 5 seconds, we obtain:

[0413] ΔI(5)=45[1-e -0.5 ]

[0414] ≈45[1-0.6065]

[0415] ≈17.7075

[0416] Through the above calculation, the data change ΔI(5)≈17.7075 indicates that the system successfully processed the data input and model output in real time within 5 seconds.

[0417] The embodiment further utilizes the stream data processing framework to perform real-time data processing and analysis, using the formula:

[0418]

[0419] P(t) represents the processed data at time t, S(τ) is the input data at time τ, and κ and μ are adjustment parameters.

[0420] The data set for real-time input of clinical data sources includes heart rate and blood pressure physiological monitoring data:

[0421] It is assumed that the input data S(t) represents a steady increase in heart rate from 70 beats / min to 90 beats / min over a period of time.

[0422] Set the parameters as follows:

[0423] Adjustment parameter κ = 1.0 (used to amplify the impact of data change rate)

[0424] Attenuation parameter μ = 0.1 (used to adjust the impact of past data on current estimates)

[0425] Assume that the heart rate data S(t) increases linearly:

[0426] S(t) = 70 + 0.333t (Here it is assumed that the heart rate increases from 70 to 90 within one hour)

[0427] calculate That is, the heart rate change rate:

[0428]

[0429] Calculate P(t) using the formula:

[0430]

[0431] For t = 60 minutes, we obtain:

[0432] p(60)=3.33(1-e -6 )

[0433] ≈3.33(1-0.0025)

[0434] ≈3.32

[0435] The processed data p(60)≈3.32 reflects the changing trend of heart rate data within one hour.

[0436] The final example integrates the output of the prediction model into the hospital's decision support system, using the following formula to combine the prediction results and decision recommendations to provide a comprehensive output:

[0437]

[0438] In this formula, D(t) represents the output of the decision support system, f i (t) and g i (t) represents the output of the prediction model and the output of the decision suggestion module, respectively, i and β i are weight coefficients, which adjust the contribution of each component to the final output.

[0439] Suppose a hospital's decision support system is processing the following data:

[0440] The prediction model output (f(t)) represents the probability of postoperative complications;

[0441] The output of the decision-making module (g(t)) represents the specific treatment recommendation based on the patient's condition.

[0442] Setting parameters: weight coefficient α = [0.7, 0.3], the weight of the prediction model output and decision recommendation; the upper and lower limits of the integral a = 0, b = 1, to simplify the integral calculation.

[0443] Settings: f1(t) = 0.9, indicating a high probability of postoperative complications; g1(t) = increased monitoring frequency, indicating specific recommendations based on high risk.

[0444] Use the formula to integrate the output:

[0445]

[0446]

[0447] The calculation result D(t)=0.93 indicates that the decision support system outputs high-probability prediction values ​​and corresponding suggestions.

[0448] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. Prediction method of postoperative complications of cardiothoracic surgery based on deep learning algorithm, Features The following steps are involved: S1. First, use optical character recognition technology to extract text data from electronic medical records, and apply natural language processing technology including BERT or GPT model to perform entity recognition on the extracted text to extract medical terms and parameters related to postoperative cardiothoracic surgery; S2. Secondly, statistical methods and machine learning algorithms are combined to identify and predict valuable biomarkers and clinical parameters through feature selection, and a graph structure containing multi-source data is constructed, including the patient's historical medical records, physiological parameters, and surgical details, in order to control the training of the graph neural network model to capture the interaction between the data; The construction of a graph structure containing multi-source data includes: S2.1, where patients are represented as nodes in the graph, and node attributes are composed of the selected key features. The edges of represent medical interactions or common medical conditions between patients; S2.2, then apply the graph neural network GNN model, use the node and edge information to capture and learn the interaction between patient data, and predict the postoperative complications of cardiothoracic surgery based on the graph structure through model training control; S2.

3. Use real-time data to dynamically adjust the graph structure and GNN model parameters to reflect new clinical data and changes in patient status, control the continuous optimization of the prediction model and adapt to changes in the clinical environment; S3. Use transfer learning techniques again to fine-tune the graph neural network model to optimize its performance in predicting complications in specific cardiothoracic surgical procedures. S4. Finally, the model is deployed to the actual clinical environment to collect postoperative patient data in real time, and the dynamic prediction adjustment mechanism based on reinforcement learning is applied to adjust the model parameters. At the same time, multimodal data fusion technology, including postoperative X-rays, real-time physiological monitoring data and historical clinical data, is combined to provide complication risk assessment through an integrated decision support system.

2. The method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm according to claim 1, Features The steps of combining statistical methods and machine learning algorithms are as follows: first, a hybrid feature selection framework is applied to optimize features by combining statistical methods and machine learning algorithms, using the formula: Where X = {x 1 , x 2 , ..., x n } represents the feature set, α i is feature x i The importance coefficient, β i A coefficient to adjust the intensity of feature influence; Then, using multimodal data fusion technology, through the double-layer integration formula: Where φ(x, y, z) = k·sin(x·y·z) is the three-dimensional data fusion function, X, Y, Z represent the feature matrices from different data sources, and k, a, b, c, d are parameters for adjusting the integration range and intensity; Finally, we build the model training of the end-to-end system using the formula: where Θ = {θ 1 ,θ 2 , ..., θ m } represents the model parameters, X={x 1 , x 2 , ..., x m } is the input feature set.

3. The method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm according to claim 1, Features The patient is represented as a node in the graph, and the construction method is as follows: S1. Define each patient as a node in a graph. The node attributes are composed of the selected key features. The selection of features is based on the newly written formula: where x i represents the ith biomarker or clinical parameter, γ i It is a dynamically adjusted weight coefficient to evaluate the importance of each feature; S2. Then construct the edges between nodes, representing the medical interactions or common medical conditions between patients. The weight of the edge is determined by the formula: where x i and x j are the feature vectors of connected nodes, and σ is a parameter to measure the sensitivity of feature similarity; S3. Use graph neural network GNN to process structured data and use the formula to update nodes: Where X is the feature set of all nodes, is the neighborhood of node i, θ ij are the learned parameters, represents the merging operation of features, and σ is a nonlinear activation function.

4. The method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm according to claim 1, Features The method for applying the graph neural network (GNN) model includes: first, in the graph construction stage, it is defined that each node in the graph represents a patient, the node attributes are composed of key biomarkers and clinical parameters, and the construction of the edge is defined by the formula: where x i (t) and x j (t) represents the state vectors of patients i and j at time t, respectively, simulating the dynamic medical relationship between patients; Next, in the model training phase, the learning formula is used: Optimize, where Θ represents the model parameters, y i is the complication status of patient i, x k is the feature of the neighbor node, θ ik is the weight parameter.

5. The method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm according to claim 1, Features The steps of dynamically adjusting the graph structure and GNN model parameters with real-time data are as follows: S1. Use real-time data integration technology to dynamically update nodes and edges in the graph, and use the formula to calculate changes in the graph structure: ΔG(t) represents the change of the graph at time t, λ is the attenuation coefficient, δx i (s) represents the state change of node i at time s; S2. Secondly, apply the online learning strategy to dynamically adjust the parameters of the GNN model and update the parameters using the formula: where θ(t) is the model parameter at time t, η is the learning rate, It is the gradient of the loss function L with respect to the parameter θ, and X represents the input feature data.

6. The method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm according to claim 1, Features The transfer learning technology implementation method includes: first, using a pre-trained GNN model and using a newly written fine-tuning formula: Adjust the model, where η is the learning rate, l is the loss function, and y i is the true label, f(x i ; θ) is the model prediction, x i It is patient data for a specific type of surgery; Then, the performance of the fine-tuned model is evaluated using the newly written evaluation formula: The prediction accuracy in the actual clinical setting was measured, where y is the actual occurrence of complications, f(x) is the model prediction, the probability of complication occurrence, σ represents the standard deviation of the prediction error, and a and b are the upper and lower limits of the integral.

7. The method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm according to claim 1, Features The integrated decision support system construction method is as follows: S1. Use reinforcement learning algorithms to perform real-time feedback adjustments on the graph neural network (GNN) model, and continuously optimize model parameters to adapt to real-time changes in the clinical environment. S2, realize multimodal data fusion, integrate postoperative X-rays, real-time physiological monitoring data and historical clinical data to provide a comprehensive view of patient health; S3. Integrate the prediction model with the hospital’s decision support system to provide real-time, data-driven medical decision support.

8. The method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm according to claim 7, Features The reinforcement learning algorithm construction method includes: first defining a reward function to evaluate the accuracy of the GNN model prediction results, using the formula: Where R represents the cumulative reward, λ is the discount factor, P(t) is the predicted probability, A(t) is the actual probability of occurrence, and α is the smoothing parameter; Secondly, the GNN model is embedded in the reinforcement learning environment, predicting actions based on the current strategy and receiving rewards based on the results. The model update formula is: Where θ(t) is the model parameter at time t, η is the learning rate, is the loss function, y(t) is the true label, and f(x(t);θ) is the model prediction value.

9. The method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm according to claim 7, Features The multimodal data fusion implementation method is as follows: S1. Preprocess and standardize the data of different modalities using the formula: D(x) represents the standardized data, I(x) represents the original data, N(x) represents the noise, γ is the smoothing parameter, and sin(ωt) is the periodic noise correction term; S2. Use convolutional neural network (CNN) to extract high-level features from postoperative X-rays, use recurrent neural network (RNN) or long short-term memory (LSTM) to extract time series features from real-time physiological monitoring data, and use multi-layer perceptron (MLP) to extract key features from historical clinical data, using the formula: Where T(x) represents the extracted features, δ i is the weight, φ i (x) represents the features extracted by different networks, is the gradient term of the feature, and α is the attenuation parameter; S3, feature fusion, using the formula: where f i (x) represents the i-th feature, α i is the weight coefficient, g i (y) represents the characteristics of the additional data source, β i is the fusion weight, H(y) represents the noise, ∈ is the fusion smoothing parameter, is the gradient term of the feature, and λ is the attenuation parameter.

10. The method for predicting postoperative complications of cardiothoracic surgery based on a deep learning algorithm according to claim 7, Features The method for integrating the prediction model with the hospital's decision support system is as follows: First, system integration and interface design, design and development of application program interface API, to achieve data exchange and processing through formulas: Where ΔI(t) represents the data change at time t, D(τ) is the data input, M(τ) is the processing model, and β and λ are adjustment parameters; Secondly, real-time data processing and analysis uses a streaming data processing framework to process data input in real time from various clinical data sources. The formula is: Where P(t) represents the processed data, S(τ) is the input data at time τ, and κ and μ are adjustment parameters; Finally, the output of the prediction model is integrated into the hospital’s decision support system through the formula: D(t) represents the output of the decision support system, fi ( t) and g i (t) represents the output of the prediction model and the output of the decision suggestion module, respectively, i and β i is the weight coefficient.

Citation Information

Patent Citations

  • Transcatheter aortic valve replacement postoperative complication prediction method, device and equipment

    CN114305323A

Cited By

  • Thoracic surgery postoperative complication early warning system

    CN120388746A

  • A thoracic surgery postoperative complication warning system

    CN120388746B

  • Thoracic surgery postoperative complication prediction method and system

    CN120636824A

  • A method and system for predicting postoperative complications in thoracic surgery

    CN120636824B