Knowledge graph neural network-based ventricular assist device adaptation degree evaluation method

Abstract

The application discloses a ventricular assist device adaptation degree evaluation method based on a knowledge graph neural network, and realizes intelligent evaluation of the adaptation degree of the ventricular assist device by building a knowledge graph neural network, training and optimizing the knowledge graph neural network, and forming an adaptation degree evaluation model; wherein the knowledge graph neural network comprises a single-modal graph network, a full-modal graph network, and a modal prediction fusion and output module. In the application, the graph attention completion model for missing feature prediction and the single-modal / full-modal fusion model for adaptation degree scoring can provide a unified and reproducible calculation approach for the ventricular assist device adaptation degree evaluation.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to a method for evaluating the fit of ventricular assist devices based on knowledge graph neural networks. Background Technology

[0002] Multimodal data is widely used in engineering and life science scenarios (including assessments of ventricular assist devices (VADs)) and is characterized by heterogeneous types (such as time-series signals, image sequences, numerical scales, and structured text), misaligned acquisition, and missing features.

[0003] Existing methods fall into two categories. One type involves simple statistical imputation or direct concatenation followed by deep model processing, failing to explicitly utilize structural dependencies and domain priors between variables. This leads to instability in handling missing modalities, long-tailed samples, or cross-center migration. Another type relies on graph learning based on static relation matrices, making it difficult to dynamically adjust edge weights and message passing strength according to the missing patterns and confidence levels of individual samples, thus limiting the accuracy and robustness of imputation and fusion. For VAD applications, clinical data often spans multiple sources, including ECG, ultrasound, laboratory, and medical records, resulting in more pronounced missing data and heterogeneity. Traditional scoring or empirical rules struggle to establish a unified and reproducible computational workflow.

[0004] In the cross-modal fusion stage, if attention weights lack prior guidance, they are prone to being assigned to channels that are inconsistent with domain common sense, resulting in structural bias and insufficient interpretability. Simultaneously, many pipelines do not explicitly output uncertainties (such as variance or confidence intervals), lacking a basis for confidence when using thresholding, which is detrimental to risk control in high-cost decisions such as device adaptation. Furthermore, if the training objective is too complex, implementation costs and convergence difficulties increase significantly, hindering its deployment in real-world systems. Summary of the Invention

[0005] To address the aforementioned problems, the purpose of this invention is to provide a method for evaluating the suitability of ventricular assist devices based on knowledge graph neural networks.

[0006] The technical solution for achieving the objective of this invention is as follows:

[0007] A method for evaluating the suitability of ventricular assist devices based on knowledge graph neural networks includes the following steps:

[0008] By constructing a knowledge graph neural network and training and optimizing it to form a fit evaluation model, intelligent evaluation of the fit of ventricular assist devices can be achieved. The knowledge graph neural network includes a single-modal graph network, a full-modal graph network, and a modality prediction fusion and output module.

[0009] The unimodal graph network consists of a cascaded first linear mapping layer, first and second graph attention networks, a second global average pooling layer, a first feedforward neural network, and a first sigmoid function. Its input is the unimodal graph of the corresponding modality, and its output is the fitness evaluation result of the corresponding modality. Each node of each unimodal graph represents a single feature of the corresponding modality, and edge connections between nodes are established based on the knowledge graph.

[0010] The full-modal graph network consists of a cascaded second linear mapping layer, first to third attention map convolutional residual layers, a second feedforward neural network, and a second sigmoid function. Its input is the full-modal graph obtained by concatenating the nodes of the single-modal graphs of all modalities and establishing edge connections between nodes based on the knowledge graph. The output is the full-modal fitness evaluation result.

[0011] The modality prediction fusion and output module consists of cascaded splicing units, multilayer feedforward neural networks, and a third sigmoid activation function. Its inputs are the outputs of all single-modality graph networks and the output of the full-modality graph network, and its output is the final fitness evaluation result.

[0012] As a further optimization of the present invention, the first attention map convolutional residual layer is composed of a cascaded first attention convolutional unit, a second LeakyReLU activation function, and a third normalization unit, and a residual connection is introduced between the input of the first attention map convolutional residual layer and the third normalization unit; the second attention map convolutional residual layer is composed of a cascaded second attention convolutional unit, a first Dropout regularization, and a fourth normalization unit, and a residual connection is introduced between the input of the second attention map convolutional residual layer and the fourth normalization unit; the third attention map convolutional residual layer is composed of a cascaded third attention convolutional unit and a third global average pooling, and a residual connection is introduced between the input and output of the third attention map convolutional residual layer.

[0013] As a further optimization of the present invention, the first graph attention network is composed of a cascaded first graph attention convolutional unit, a first LeakyReLU activation function and a first layer normalization unit, and the second graph attention network is composed of a cascaded second graph attention convolutional unit, a first Dropout regularization and a second layer normalization unit.

[0014] As a further optimization of the present invention, the knowledge graph neural network is trained and optimized based on a training database. The steps for constructing the training database include:

[0015] Data related to multimodal ventricular assist devices were collected, and features were extracted and preprocessed. Preprocessing included missing labeling and missing completion. During missing detection, the missing position was recorded and initialized to 0, and the missing label embedding was set.

[0016] The preprocessing results are labeled with fit to form a training database.

[0017] As a further optimization of the present invention, missing data completion is performed using a cross-modal data completion model, which consists of a cascaded knowledge graph guided graph construction layer, a graph neural network encoding layer, a fusion layer, and an output layer.

[0018] The knowledge graph guided graph construction layer is used to build the graph structure. Each node in the graph structure represents a feature, and edge connections between nodes are established based on the knowledge graph.

[0019] The encoding layer of the graph neural network consists of a cascaded first linear transformation layer, first and second graph attention networks, and a first global average pooling layer.

[0020] The fusion layer consists of cascaded first and second fully connected layers;

[0021] In the output layer, the output of the fusion layer is first input into the third fully connected layer, and then the missing prediction values ​​are filtered out by a mask.

[0022] As a further optimization of the present invention, the formula for calculating the edge weight in the graph structure is:

[0023] ,

[0024] In the formula, For adjustable weights, It is a sigmoid compression function. This indicates the strength of the causal / dependency relationship between node i and node j; This represents the absolute value of the statistical correlation coefficient between the corresponding features of node i and node j; , is an adaptive term. Let the observation labels be for nodes i and j. .

[0025] As a further optimization of the present invention, in the training phase of the cross-modal data completion model, a prior-weighted masking strategy is adopted to randomly mask the training data. The masking probability of each feature is:

[0026] ,

[0027] in, For the training dataset, the first Historical missing rate of each feature For the first The importance of each feature; and These represent the lower and upper bounds of the probability of feature masking, respectively; coefficients ,and .

[0028] As a further optimization of the present invention, the output layer uses the MC Dropout method to perform multiple forward propagations, using the mean of the obtained predictions as the missing completion value of the final output, and calculating the 95% confidence interval. , The mean of the predicted values, This represents the variance of the predicted values.

[0029] The present invention also provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the ventricular assist device fit assessment method based on knowledge graph neural network as described above.

[0030] The present invention also provides an electronic device, characterized in that it comprises:

[0031] Memory, used to store computer programs;

[0032] A processor is configured to implement the steps of the ventricular assist device fit assessment method based on knowledge graph neural networks as described above when executing the computer program.

[0033] Compared with the prior art, the significant advantages of this invention are:

[0034] This invention presents a method for assessing ventricular assist device (VAD) fit based on a knowledge graph neural network. The method constructs a relational graph guided by a domain knowledge graph, enabling the graph neural network to explicitly utilize causal / dependency priors and statistical correlations in attention scoring. It introduces an edge weight design that adapts to the sample state to respond to current missing features and credibility. In the fusion stage, an attention mechanism is employed along with prior consistency constraints to suppress unreasonable channels. The training uses a simplified joint loss (primarily based on completion reconstruction and prior consistency) to reduce implementation complexity and improve convergence stability. The graph attention completion model for missing feature prediction and the single-modal / multi-modal fusion model for fit scoring in this invention provide a unified and reproducible computational approach for VAD fit assessment.

[0035] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0036] Figure 1 This is the overall flowchart of the present invention.

[0037] Figure 2 It is a knowledge graph neural network structure diagram.

[0038] Figure 3It is a structural diagram of the attention graph convolutional residual layer in a knowledge graph neural network.

[0039] Figure 4 It is a structure diagram of the attention graph convolutional residual layer 2 in a knowledge graph neural network.

[0040] Figure 5 It is a three-structure graph of attention graph, convolution, residual layer, and knowledge graph neural network.

[0041] Figure 6 It is a cross-modal data completion model structure diagram.

[0042] Figure 7 It is a structure diagram of the attention map convolutional layer in the cross-modal data completion model.

[0043] Figure 8 It is the structure diagram of the attention map convolutional layer 2 in the cross-modal data completion model. Detailed Implementation

[0044] like Figure 1 As shown, the ventricular assist device (VAD) fit evaluation method based on knowledge graph neural network provided by the present invention includes the following steps:

[0045] S1. Structural design of knowledge graph neural networks:

[0046] like Figure 2 As shown, the knowledge graph neural network consists of a single-modal graph network, a full-modal graph network, and a modality prediction fusion and output module.

[0047] S11. Construction of a single-modal graphical network:

[0048] In this embodiment, four modalities (ECG, Echo, biochemistry, and basic information / medical history) are addressed by forming unimodal graphs (each node in each unimodal graph represents a single feature of the corresponding modality, and edge connections between nodes are established based on the knowledge graph). Each unimodal graph network first performs a linear mapping on the input features, expanding them into a fixed-dimensional high-dimensional node embedding representation. Subsequently, two layers of attention-based graph neural networks (GAT) are used for feature propagation and fusion between nodes. After graph-level global average pooling, a modality-level feature vector is generated, representing the global feature expression of that modality. Finally, this vector is input into a feedforward neural network, and the sigmoid function outputs scores for each modality. , where m represents the mode.

[0049] S12, Construction of Full Modal Graph Network:

[0050] The nodes of four unimodal graph networks are concatenated, and edge connections between the nodes are established based on the knowledge graph to construct a full-modal graph. The full-modal graph network first performs a linear mapping on the input features, and then sequentially encodes them through three layers of attention map convolutional residual layers: as follows. Figure 3 As shown, the first layer includes a multi-head attention mechanism, a LeakyReLU activation function, LayerNormalization, and residual connections; as Figure 4 As shown, the second layer introduces Dropout regularization to enhance generalization capabilities and also includes LayerNorm and residual paths; as Figure 5 As shown, the third layer performs graph-level average pooling after graph attention convolution and combines it with residual connections to form the final full-modality graph-level feature representation. This full-modality graph-level feature representation is then input into a feedforward neural network (FFN) and outputs a full-modality score via a sigmoid function. The formula for calculating the attention weights in a full-modal graph network is as follows:

[0051]

[0052] in, These are modal interaction weights used to model the differences in importance between modalities. They can be initialized by experts and learned during training.

[0053] S13, Modal Prediction Fusion and Output:

[0054] The following four pieces of information were collected from each of the four modal and full-modal branches: scores and The features include the proportion of missing modalities, the confidence score for missing completion, and the prior bias for each modality. Each modality constitutes a 4-dimensional feature vector, which is then concatenated in modal order to form a unified fusion input vector. The confidence score for missing completion is represented by the width of the 95% confidence interval (CI) estimated through multiple sampling, used to quantify prediction uncertainty. The prior bias for each modality is labeled by clinical experts to reflect its importance in the fit score. Finally, this fusion vector is sequentially input into a multi-layer feedforward neural network (FFN) to achieve feature fusion and nonlinear transformation, and finally outputs the final score after passing through a sigmoid activation function. The range of the result is [0,1].

[0055] S2. Construction of the training dataset for the knowledge graph neural network:

[0056] S21. Multimodal data acquisition:

[0057] Multimodal data of the sample subject are collected. In this embodiment, the data includes electrocardiogram (ECG) data, echocardiography (Echo) data, biochemical indicators, basic information and medical history information.

[0058] S22. Perform corresponding data processing for different modalities to form structured feature vectors. This includes:

[0059] S221, ECG Feature Extraction: Noise reduction and heartbeat detection are performed on the ECG signal to extract heart rate, QT / QTc interval, QRS width, QT interval dispersion, and heart rate variability. Atrial fibrillation (AF), atrial flutter (AFL), ST segment elevation / depression, PVC (premature ventricular contractions) frequency, and electrical axis deviation are identified using deep learning methods or expert annotation.

[0060] S222, Echo Feature Extraction: The following echocardiographic features are extracted using deep learning methods or expert annotation: Left ventricular ejection fraction (LVEF), left ventricular end-diastolic diameter (LVEDD), left ventricular end-systolic diameter (LVESD), interventricular septal thickness (IVS), left ventricular posterior wall thickness (LVPW), left atrial diameter (LA diameter), and wallmotion score, etc.; dynamic cardiac function parameters such as tricuspid annular systolic displacement (TAPSE), tricuspid annular S-peak velocity (TDI-S), E / A ratio, E / e ratio, and pulmonary artery systolic pressure; valvular and structural heart disease features such as moderate to severe mitral regurgitation (MR), moderate to severe tricuspid regurgitation (TR), and right atrial diameter and area; right ventricular ejection fraction (RVEF) and hepatic venous echocardiography signs (regurgitation, systolic absence, etc.), etc.; and hemodynamic parameters such as estimated cardiac output (CO).

[0061] S223. Biochemical Indicator Feature Extraction: Convert the units of biochemical indicators to form a numerical feature vector. The data includes: lactate, brain natriuretic peptide (BNP or NT-proBNP), creatinine (Cr), sodium (Na⁺), potassium (K⁺), blood pH, partial pressure of oxygen (PaO2), partial pressure of carbon dioxide (PaCO2), albumin (Alb), total bilirubin (TBIL), hemoglobin (Hb), blood urea nitrogen (BUN), and high-sensitivity troponin (hs-TnI).

[0062] S224. Extraction of basic information and medical history: The basic information and medical history records are structured and organized to extract basic information such as patient age, gender, and BMI. It is also marked whether there is a history of diabetes, hypertension, chronic kidney disease, myocardial infarction, COPD or respiratory disease, liver disease, malignant tumor, or cerebrovascular disease, which are represented by 0 / 1 codes respectively.

[0063] S23, Data Preprocessing

[0064] S231, Missing Labeling: Perform missing feature detection on the feature vector of the main body of the input sample, record the position of the missing feature and initialize it to 0, and set the missing label embedding.

[0065] S232. Construction of Cross-Modal Data Completion Model: After detecting missing ECG, Echo, and biochemical indicators, a cross-modal data completion model based on graph neural networks is invoked to complete the data. For example... Figure 6 The construction of the cross-modal data completion model shown mainly includes the following steps:

[0066] S2321. Knowledge Graph Guidance Graph Construction (The construction of the guidance graph here is the same as the modal graph method in knowledge graph neural networks): Construct the multimodal features into a graph structure 𝐺=(𝑉,𝐸), where each node in the node set 𝑉 represents a clinical feature, including 52 features such as cardiac function indicators (EF), biochemical indicators (lactate, creatinine), electrocardiogram features (HR, QRS), and medical history information. The edge set E is connected based on the domain knowledge graph (medical knowledge graph in this embodiment) and statistical relevance. The weight calculation formula for each edge is:

[0067]

[0068] in, This is the Sigmoid compression function; Representation of features With features The strength of the causal / dependency relationship between them is set by expert annotation and the domain's public knowledge graph; It represents the absolute value of the statistical correlation coefficient between two features calculated based on historical sample data (Pearson or Spearman). For sample adaptation term, This represents its parameter set (m, q, etc., which can also be empty). The contribution ratio of the three types of information to the edge weights is controlled, and is usually set as adjustable weights for automatic optimization during model training.

[0069] Sample Adaptive Term This is a scoring function for the current sample state, used for "dynamic fine-tuning" of the edge weights. Its parameters... It may contain observation markers. (If feature i / j exists) or (1 if the value is 1, otherwise 0), data quality / reliability, etc. (When there is no quality score, it can be set as follows) And (optionally) the uncertainty estimated by the model and the feature type embedding:

[0070] .

[0071] In this embodiment, only parameter m is selected, then: If both ends of the edge have values, then , (Slightly weighted); as long as one end is missing , (Slight weight reduction).

[0072] S2322, Graph Neural Network Encoding: Employs trainable linear transformation layers to extend the original one-dimensional scalar feature vectors of nodes to high-dimensional feature vectors. Subsequently, two layers of attention-based graph neural networks are used to encode the constructed feature map, followed by graph-level average pooling to achieve the propagation and fusion of feature information between nodes. An improved attention weight calculation method, guided by causal strength and adjusted for missing importance, is combined to enhance the accuracy of the completion prediction.

[0073] like Figure 7 As shown, the first layer of the GAT structure consists of a Graph Attention Convolution unit, a LeakyReLU activation function, and Layer Normalization. This layer performs a neighbor feature aggregation operation based on a multi-head attention mechanism, completing the weighted fusion of each node's information from its first-order neighbors. This enables nodes to capture locally dependent features and mitigates the impact of feature distribution differences on training. Figure 8 As shown, the second-layer GAT structure consists of graph attention convolutional units, Dropout regularization, and Layer Normalization. Building upon the first layer, this layer further aggregates information from higher-order neighbor nodes, enabling deeper interactions across modal features, while simultaneously reducing overfitting risk through Dropout. After computation by this layer, a high-dimensional embedding representation of each node is obtained. The improved attention weight calculation formula is as follows:

[0074]

[0075] in, and Represents the original node and The initial high-dimensional feature vector; It is the parameter matrix to be trained, which contains the original nodes. and Linear mapping to the embedded feature space; and It is the mapped node embedding vector; This indicates a feature concatenation operation, which concatenates two embedding vectors into a single long vector. It is a vector of parameters to be learned for an attention mechanism, which is used to perform a linear mapping on the concatenated vector to generate a scalar similarity score; Indicates the strength of causal / dependency relationships derived from the domain knowledge graph; Indicate the importance of the missing node i; and These are the corresponding hyperparameters, used to control the contribution ratios of causal terms and missing importance terms. Next, attention scores are calculated for all neighboring nodes. Softmax normalization is performed to obtain the final attention weights.

[0076] S2323 Fusion Layer: The obtained global feature vector passes through a fully connected network containing two fully connected (Dense) layers and two ReLU activation functions: the first fully connected (Dense) layer performs a linear transformation on the features and reduces their dimensionality, and then the ReLU function enhances the non-linear expression; then the second fully connected (Dense) layer further compresses the dimensionality, and then the ReLU function outputs the final fused feature vector for subsequent prediction steps.

[0077] S2324 Output Layer (Prediction and Uncertainty): This layer contains a prediction output module and a post-processing module. The prediction output module inputs the feature vectors generated by the fusion layer into the fully connected (Dense) layer, outputs the prediction results of all feature nodes through a linear activation function, and then filters the prediction values ​​corresponding to missing indicators through a mask. The uncertainty estimation module uses the MC Dropout method, keeping Dropout enabled during the inference phase and performing 20 forward propagations to obtain the statistical distribution of the prediction values. Based on this, the mean is calculated as the final estimate, and the standard deviation is used to measure the prediction uncertainty. Furthermore, a 95% confidence interval is calculated. The expression is:

[0078]

[0079] in, The mean, To predict variance.

[0080] S2325 Data Completion Model Training and Evaluation

[0081] S23251 Training Sample Construction and Masking Strategy

[0082] During the training phase, random masking is applied to complete samples to form supervised pairs where "input is visible, target is masked." To better reflect the true missing data distribution and enhance robustness, a prior-weighted masking strategy is adopted: for the first... Each feature is assigned a masking probability.

[0083]

[0084] in, To obtain the historical missing rate of this feature in the training set, The importance of this feature is given based on knowledge graphs or expert experience; , These are the lower and upper bounds of the probability of feature masking, respectively; ,and During the verification / testing phase, only true missing values ​​or preset masking sets are evaluated; no new random masks are introduced.

[0085] S23252 Joint Loss Function

[0086] The loss is calculated only for the set of dimensions that were masked during training, and the objective is to minimize it.

[0087]

[0088] in These are the weighting coefficients; To reconstruct the loss, standard likelihood and cross-entropy forms are used for continuous and discrete features, respectively; The prior consistency loss is used to suppress excessive attention allocation on low prior edges, and its formula is:

[0089]

[0090] in Let be the set of edges of the graph; For the prior strength of the knowledge graph This represents the attention coefficient for the last layer.

[0091] S24, Compatibility Labeling

[0092] The preprocessing results are stored in a database. Each data entry in the database includes a multimodal feature set and a corresponding fitness label. In this embodiment, the fitness value ranges from [0,1]. During the database construction process, the data is divided into training, validation, and test sets according to a certain ratio for subsequent model training and performance evaluation.

[0093] S3. Training and optimization of knowledge graph neural networks:

[0094] S31. Joint Loss and Training Strategy: Input the training set from the training database constructed in step S2 into the designed knowledge graph neural network for supervised training to obtain the fitness evaluation model. Iteratively update the model parameters using backpropagation and gradient descent optimization methods. Adjust hyperparameters such as the model learning rate, batch size, and number of attention heads based on the validation results to continuously optimize the model.

[0095] S32. Interpretable Modeling: Using the fitness assessment model obtained from training, the SHAP (SHapley Additive exPlanations) algorithm is used to perform a dimension-by-dimensional perturbation analysis on the model input features; the marginal contribution of each input feature (such as EF, lactic acid, BNP, AF, Cr, etc.) to the final prediction result is calculated; the contribution values ​​are sorted by absolute value and a "risk factor contribution map" is drawn to show the main basis for the model to judge "high risk".

[0096] S4. Compatibility Assessment:

[0097] S41. Collect multimodal data of the target subject, complete feature generation and missing data labeling, and call the cross-modal data completion model to predict and complete the missing dimensions; call the fitness evaluation model to perform fitness assessment. predict.

[0098] S42. Judgment Rules and Recommendations

[0099] In this embodiment, two threshold levels are set. and The value range is [0,1]. The judgment and suggestions are as follows:

[0100] 1. Low adaptation zone: If The system outputs "low compatibility," indicating that the device implantation process is not required.

[0101] 2. Gray area / follow-up area: If Output dynamic monitoring and periodic review.

[0102] 3. High adaptation zone: If It has high output adaptability and can be included in the device implantation process.

[0103] S43, Standardized Report Generation

[0104] The system automatically generates a standardized calculation report based on the fit output (which can be used for clinical assistance in application-oriented examples). This report includes... It also includes a "Risk Factor Contribution Ranking Map" generated based on the SHAP interpretive algorithm, which intuitively displays the key indicators (such as EF, BNP, Cr, AF, etc.) that affect the system's judgment and their contribution ranking.

[0105] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

[0106] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0110] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A ventricular assist device fitness evaluation method based on a knowledge graph neural network, characterized in that, The application relates to a knowledge graph neural network for intelligent evaluation of the adaptability of a ventricular assist device. The single-mode graph network is composed of a cascaded first linear mapping layer, a first and a second graph attention network, a second global average pooling layer, a first feedforward neural network and a first Sigmoid function, the input of the single-mode graph network is a single-mode graph corresponding to a mode, and the output of the single-mode graph network is an adaptability evaluation result corresponding to the mode; each node of each single-mode graph represents a single feature of a corresponding mode, and an edge connection between nodes is established according to a knowledge graph; The full-mode graph network is composed of a cascaded second linear mapping layer, a first to a third attention graph convolution residual layer, a second feedforward neural network and a second Sigmoid function, the input of the full-mode graph network is a full-mode graph obtained by splicing nodes of single-mode graphs of all modes and establishing an edge connection between the nodes according to the knowledge graph, and the output of the full-mode graph network is a full-mode adaptability evaluation result; The modal prediction fusion and output module is composed of a cascaded splicing unit, a multilayer feedforward neural network and a third Sigmoid activation function, the input of the modal prediction fusion and output module is the output of all single-mode graph networks and the output of the full-mode graph network, and the output of the modal prediction fusion and output module is a final adaptability evaluation result; The first attention graph convolution residual layer is composed of a cascaded first attention convolution unit, a second LeakyReLU activation function and a third layer normalization unit, and a residual connection is introduced between the input of the first attention graph convolution residual layer and the third layer normalization unit; the second attention graph convolution residual layer is composed of a cascaded second attention convolution unit, a first Dropout regularization and a fourth layer normalization unit, and a residual connection is introduced between the input of the second attention graph convolution residual layer and the fourth layer normalization unit; the third attention graph convolution residual layer is composed of a cascaded third attention convolution unit and a third global average pooling, and a residual connection is introduced between the input and the output of the third attention graph convolution residual layer; The first graph attention network is composed of a cascaded first graph attention convolution unit, a first LeakyReLU activation function and a first layer normalization unit, and the second graph attention network is composed of a cascaded second graph attention convolution unit, a first Dropout regularization and a second layer normalization unit; The knowledge graph neural network is trained and optimized based on a training database, and the construction steps of the training database include: Multi-modal ventricular assist device related data are collected, and feature extraction and preprocessing are performed; the preprocessing includes missing label and missing completion, the missing position is recorded and initialized as 0 during missing detection, and a missing label embedding is set; The adaptability of the preprocessing result is labeled to form a training database; The missing completion is performed through a cross-modal data completion model, and the cross-modal data completion model is composed of a cascaded knowledge graph guided graph construction layer, a graph neural network coding layer, a fusion layer and an output layer: ​ The knowledge graph guided graph construction layer is configured to construct a graph structure, each node in the graph structure represents a feature, and edges between the nodes are established according to the knowledge graph; The graph neural network encoding layer is composed of a cascaded first linear transformation layer, first and second graph attention networks, and a first global average pooling layer; The fusion layer is composed of a cascaded first and second fully connected layers; In the output layer, the output of the fusion layer is first input into a third fully connected layer, and then the corresponding prediction value is screened out through a mask.

2. The method of claim 1, wherein, The edge weight calculation formula in the graph structure is: ， wherein, is the adjustable weight, is the Sigmoid compression function, represents the strength of the causal / dependence relationship between node i and node j; represents the absolute value of the statistical correlation coefficient between the corresponding features of node i and node j; is the adaptive term, is the observed label of node i and node j, .

3. The method of claim 1, wherein, The output layer performs multiple forward propagations using the MC Dropout method, and the mean of the obtained prediction values is taken as the final output of the missing completion value, and the 95% confidence interval is calculated , is the mean of the prediction value, is the variance of the prediction value.

4. The method of claim 1, wherein, In the training phase of the cross-modal data completion model, a prior weighted masking strategy is adopted to implement random masking on the training data, and the masking probability of the first feature is: ， wherein, is the historical missing rate of the th feature in the training dataset, is the importance of the th feature; and are the lower and upper bounds of the probability of feature masking, respectively; the coefficient , and .

5. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method in any one of claims 1-4.

6. An electronic device, comprising: Comprise: a memory for storing a computer program; a processor for executing the computer program to realize the steps of the method in any one of claims 1-4.

Images