Death rate prediction method based on fusion of medical ontology embedding and attention mechanism and computer program product

Abstract

The invention discloses a death rate prediction method based on fusion of medical ontology embedding and an attention mechanism and a computer program product, and relates to the technical field of medical information. The method comprises the following steps: respectively carrying out feature mapping on population statistical information, diagnosis information, treatment information and vital sign information of a patient; constructing medical ontologies for the diagnosis information nodes and the treatment information nodes of the feature mapping; performing nonlinear transformation on the diagnosis ontology embedding matrix and the treatment ontology embedding matrix, and processing diagnosis information and treatment information through a mask attention mechanism; obtaining a corresponding treatment result based on the treatment access record of the patient, learning a causal relationship between diagnosis and treatment, calculating a conditional probability between the diagnosis result and the treatment result of the patient, and obtaining an adjacent matrix; and predicting the death rate of the patient according to the demographic information, the vital sign information, the diagnosis sequence, the treatment sequence and the adjacency matrix of the feature mapping. According to the invention, the accuracy of death rate prediction can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of medical information technology, and in particular to a mortality prediction method and a computer program product based on the fusion of medical ontology embedding and attention mechanism. Background Art

[0002] With the rapid growth of medical data, especially in the context of multimodal data such as electronic health records (EHR), medical imaging, and genomic data, mortality prediction plays an important role in public health, clinical medicine, and disease management. Accurate mortality prediction models can help doctors identify high-risk patients in a timely manner, so that they can take effective intervention measures to reduce mortality.

[0003] However, traditional mortality prediction methods often rely on a single type of medical data, and the performance of the model is limited by the heterogeneity of the data and the lack of comprehensive knowledge representation, resulting in inaccurate predictions. Therefore, a technical solution to improve the accuracy of mortality prediction is urgently needed. Summary of the invention

[0004] The purpose of the present invention is to provide a mortality prediction method and a computer program product based on the fusion of medical ontology embedding and attention mechanism to solve the above technical problems existing in the prior art. The preferred technical solutions among the many technical solutions provided by the present invention can produce many technical effects as described below.

[0005] To achieve the above object, the present invention provides the following technical solutions: The present invention provides a mortality prediction method based on the fusion of medical ontology embedding and attention mechanism, comprising the following steps: The patient's demographic information, diagnosis information, treatment information, and vital signs information are feature mapped as a demographic information node, a diagnosis information node, a treatment information node, and a vital signs information node, respectively; Constructing medical ontologies for the diagnosis information nodes and treatment information nodes of the feature mapping respectively, and obtaining the diagnosis ontology embedding matrix and treatment ontology embedding matrix of the patient; The diagnosis ontology embedding matrix, the treatment ontology embedding matrix, and the diagnosis information and treatment information of the feature map are processed through a masked attention mechanism to obtain a diagnosis sequence and a treatment sequence of the patient; Obtaining corresponding treatment results based on the patient's medical visit records, learning the causal relationship between diagnosis and treatment, calculating the conditional probability between the patient's diagnosis result and treatment result, and obtaining an adjacency matrix; The patient's mortality rate is predicted based on the demographic information, vital sign information, the diagnosis sequence, the treatment sequence and the adjacency matrix of the feature map.

[0006] In some embodiments, the demographic information, diagnosis information, treatment information, and vital sign information of the patient are feature mapped as a demographic information node, a diagnosis information node, a treatment information node, and a vital sign information node, respectively, including the following steps: After converting the patient's demographic information, diagnosis information, treatment information, and vital sign information into continuous vectors, a binary vector is generated through label encoding; and the label-encoded diagnosis information and treatment information are activated through a nonlinear function.

[0007] In some embodiments, the medical ontology is a directed acyclic graph with a hierarchical structure that represents the diagnosis information nodes and treatment information nodes in the form of parent-child relationships, and the output form is the diagnosis ontology embedding matrix and the treatment ontology embedding matrix.

[0008] In some embodiments, the diagnostic ontology embedding matrix and the treatment ontology embedding matrix are activated by a nonlinear function before being processed by a masked attention mechanism.

[0009] In some embodiments, after the diagnosis ontology embedding matrix and the treatment ontology embedding matrix are processed by the masked attention mechanism, the method further includes: The output processed by the masked attention mechanism is sequentially subjected to linear function and ReLu function activation, residual connection, normalization, linear transformation, residual connection, and normalization to obtain the diagnosis sequence and treatment sequence of the patient.

[0010] In some embodiments, a priori guided attention mechanism is used to predict the mortality rate of the patient, and the output of the priori guided attention mechanism is subjected to an attention pooling operation to obtain the mortality rate of the patient.

[0011] In some embodiments, before the attention pooling operation, the method further includes: The output of the prior-guided attention mechanism is linearly transformed and then processed by the ReLU activation function; the output processed by the ReLU activation function is residually processed and then normalized.

[0012] In some embodiments, the mortality prediction method based on the fusion of medical ontology embedding and attention mechanism also includes weighted processing of the diagnostic ontology embedding matrix, treatment ontology embedding matrix and feature mapped diagnostic information and treatment information processed by the masked attention mechanism, and using the pooled features for the classification task of the medical ontology.

[0013] In some embodiments, the mortality prediction method based on the fusion of medical ontology embedding and attention mechanism also includes using the diagnostic ontology embedding matrix and the treatment ontology embedding matrix for the classification task of the medical ontology.

[0014] On the other hand, the present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the mortality prediction method based on the fusion of medical ontology embedding and attention mechanism as described above.

[0015] Implementing one of the above technical solutions of the present invention has the following advantages or beneficial effects: The present invention obtains a medical ontology by processing structured data and unstructured data in EHR, then obtains different vector representations of the relevant information of the medical ontology through an encoder, and then obtains prediction indicators through a prediction model, thereby improving the accuracy of mortality prediction in the current medical field. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. It is obvious that the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings: Figure 1 is a flowchart of a mortality prediction method based on the fusion of medical ontology embedding and attention mechanism according to an embodiment of the present invention; Figure 2 It is a system framework diagram of a mortality prediction method based on the fusion of medical ontology embedding and attention mechanism in an embodiment of the present invention; Figure 3 It is a medical ontology framework diagram constructed by diagnostic information in an embodiment of the present invention; Figure 4 It is a medical ontology framework diagram constructed by treatment information in an embodiment of the present invention; Figure 5 This is an example diagram of medical consultation-treatment characteristic information created based on medical consultation records according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention clearer, the various exemplary embodiments to be described below will refer to the corresponding drawings, which constitute a part of the exemplary embodiments, wherein various exemplary embodiments that may be used to implement the present invention are described. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation methods described in the following exemplary embodiments do not represent all implementation methods consistent with the present disclosure. It should be understood that they are only examples of processes, methods, devices, etc. that are consistent with some aspects of the present disclosure as detailed in the attached claims, and other embodiments may also be used, or the embodiments listed herein may be modified in structure and function without departing from the scope and essence of the present invention.

[0018] In order to illustrate the technical solution of the present invention, a specific embodiment is used below for description, and only the parts related to the embodiment of the present invention are shown.

[0019] Embodiment 1: like Figure 1-Figure 2 As shown, the present invention provides a mortality prediction method based on the fusion of medical ontology embedding and attention mechanism, comprising the following steps: S100, feature mapping the patient's demographic information, diagnosis information, treatment information, and vital signs information as a demographic information node, a diagnosis information node, a treatment information node, and a vital signs information node, respectively; S200, constructing medical ontologies for the diagnosis information nodes and treatment information nodes of the feature mapping, respectively, to obtain the patient's diagnosis ontology embedding matrix and treatment ontology embedding matrix; S300, processing the diagnosis ontology embedding matrix, the treatment ontology embedding matrix, and the diagnosis information and treatment information of the feature map through a masked attention mechanism to obtain a diagnosis sequence and a treatment sequence of the patient; S400, based on the patient's medical visit records, obtain the corresponding treatment results, learn the causal relationship between diagnosis and treatment, calculate the conditional probability between the patient's diagnosis results and treatment results, and use The function generates conditional probability distribution and finally constructs an adjacency matrix, which represents the relationship between diagnosis and treatment and provides support for subsequent analysis and prediction. S500, predicting the patient's mortality rate based on the demographic information, vital sign information, diagnosis sequence, treatment sequence and adjacency matrix of the feature map.

[0020] The present invention processes structured data and unstructured data in EHR (electronic health records) to obtain a medical ontology, then encodes the relevant information of the medical ontology to obtain different vector representations, and then obtains prediction indicators through a prediction model, thereby improving the accuracy of patient mortality prediction.

[0021] Based on the above embodiment, the above patient's demographic information, diagnosis information, treatment information, vital signs information and other data are derived from the open source, de-identified MIMIC-III (Medical Information for Intensive Care, MIMIC) database. In step S100, the patient's basic hospitalization information and related data are extracted from the MIMIC-III database, which includes a data fusion module with structured data and unstructured data, the structured data includes static information and time-series medical codes; the unstructured data includes time-series clinical record information.

[0022] It should be noted that the MIMIC-III database is a large-scale, publicly available clinical data set that is widely used in medical research and training of artificial intelligence models. This data set contains anonymous medical data from patients in intensive care units, covering patients' physiological information, medical records, drug treatments, laboratory results, imaging data, etc., and is widely used in medical data analysis, machine learning, predictive model development and other research. Of course, the database of this embodiment can also be a subsequent version of MIMIC-III, such as MIMIC-IV.

[0023] Based on the above embodiment, in the above step S100, the demographic information, diagnostic information, treatment information, and vital signs information of all patients are extracted from the data as demographic information node i, diagnostic information node d, treatment information node p, and vital signs information node s, respectively, and the node data is feature mapped using an encoder. Further, after extracting the demographic information, diagnostic information, treatment information, and vital signs information from the data set, these discrete data are then converted into continuous vectors, which can capture the intrinsic characteristics and relationships of the data. Finally, the demographic information i, vital signs s, diagnostic information d, and treatment information p are converted into encoded information vectors, which are then encoded to generate binary vectors, where 0 or 1 at each position represents the presence or absence of the feature. When these data of all patients are converted into binary vectors, the patient's diagnostic information x d and treatment informationx p Two vectors pass through Function for nonlinear activation, and then as Input to the masked attention mechanism in the model; patient demographic information x iand vital signs x s As The prior in the model guides the input of the attention mechanism model.

[0024] It should be noted that It is a deep learning model based on self-attention. The core of its success lies in the innovative use of efficient self-attention, which enables the model mechanism to capture long-distance dependencies while allowing unitized calculations, greatly improving efficiency and model performance. Its core technology is the self-attention mechanism, which enables the model to dynamically select and aggregate relevant information by focusing on the interdependencies between positions in the input sequence. Its main components include: self-attention mechanism, multi-head attention mechanism, position encoding, encoder-decoder structure. The model was originally used for natural language processing and quickly extended to multiple fields, such as natural language processing, image processing, speech recognition, time series prediction, etc.

[0025] Based on the above embodiment, in step S200, the medical ontology is a directed acyclic graph with a hierarchical structure that represents diagnosis information nodes and treatment information nodes in the form of parent-child relationships, and the output form is a diagnosis ontology embedding matrix and a treatment ontology embedding matrix. Specifically, like Figure 3-Figure 4 As shown, the nodes in the embedding matrix is the basic embedding via the attention mechanism The embedding matrix of all leaf nodes is calculated by computing the embedding of its ancestors. In the medical ontology, each diagnosis (treatment) node is assigned a basic embedding vector, and then is the basic embedding of each layer of nodes. The nodes in the embedding matrix The formula can be expressed as: (1); in, It is calculated Time Embed The attention weight on represents the basic embedding, Represents each layer node and the embedding of its ancestors.

[0026] The above attention weights By following Function calculation: (2); in, Is a scalar value indicating and Compatibility between basic embeddings of . Scalar values is calculated by a feed-forward network with a single hidden layer: (3); in, yes and The concatenated weight matrix, is the bias vector, is the weight vector used to generate the scalar value.

[0027] When the medical ontology embedding matrix of diagnosis information and treatment information is obtained and Afterwards, they will pass through Function for nonlinear activation, and then as Input to the model's masked attention module.

[0028] It should be noted that the above medical ontology It is a formal description of concepts in the medical field and the relationships between them. It provides important support for data structuring by defining the hierarchy and relationships between medical terms and concepts. Medical ontology can effectively integrate knowledge from different data sources (such as clinical texts, medical images, genetic data, etc.) and map these multimodal information into a unified framework. Embedding based on medical ontology The method provides a unified semantic representation by converting medical knowledge into a low-dimensional vector space, which makes it possible to integrate and efficiently process multimodal data.

[0029] Medical ontology information embedding It is an ontology in the medical field The technology of converting information into machine-processable numerical representations (i.e., embedding vectors). Its purpose is to use the medical concepts and terms defined in the ontology and convert this information into a form that can be understood by computers through embedding learning methods. The core goal of medical ontology information embedding is to efficiently map the concepts and relationships in the medical ontology into a low-dimensional processing space through machine learning technology, so that these semantic information can be processed by computers. Medical ontology information embedding is usually achieved with the help of deep learning, graph neural networks, etc. , natural language processing and other technologies to achieve this.

[0030] Based on the above embodiment, before the diagnosis ontology embedding matrix and the treatment ontology embedding matrix are processed by the masked attention mechanism, they are processed by a nonlinear function (such as The activation function is used to activate the diagnostic ontology embedding matrix and the treatment ontology embedding matrix to obtain enhanced feature representation. Furthermore, after the diagnostic ontology embedding matrix and the treatment ontology embedding matrix are processed by the masked attention mechanism, they also include: The output processed by the masked attention mechanism is sequentially subjected to linear function and ReLu function activation, residual connection, normalization, linear transformation, residual connection, and normalization to obtain the patient's diagnosis sequence and treatment sequence.

[0031] It is understandable that in step S300, The model uses a masked attention mechanism to process diagnosis sequences and treatment sequences, thereby considering the dependencies between elements in the sequence. The diagnosis and treatment ontology embedding matrices after nonlinear activation and the two binary vectors of diagnosis information and treatment information are used as The model masks the input to the attention mechanism, which allows the model to take into account dependencies between elements in a sequence when processing sequence data.

[0032] The above residual connection adds the input directly to the (Mask Attention Mechanism) processed output, which helps alleviate the gradient vanishing problem in deep networks.

[0033] The above normalization is implemented by layer normalization, which normalizes the data to improve the stability of training.

[0034] The above linear function is activated with the ReLu function ( ) represents a feedforward neural network Layer, which consists of a linear transformation followed by a ReLU activation function. This step performs a nonlinear transformation on the data, increasing the expressive power of the model.

[0035] Then residual connection and layer normalization (normalization) are performed to process the output of the feedforward neural network layer. This also helps the stability and efficiency of model training.

[0036] The above linear transformation is achieved through the linear layer, which transforms the output of the feedforward network into the final output. The linear layer is usually used to project the data back to the original dimension or transform it into a specific dimension required by the model.

[0037] Based on the above embodiment, in step S400, based on the patient's visit record visit, the corresponding treatment result is obtained, and the causal relationship between diagnosis and treatment is learned, and the conditional probability of the diagnosis result and the treatment result is calculated, and the Function generates conditional probability distribution and creates adjacency matrix , as a guide for the attention mechanism, indicating the relationship between diagnosis and treatment.

[0038] It is understandable that Learn the causal relationship between diagnosis and treatment, and calculate the conditional probability of diagnosis and treatment results. It is a network that combines graph topology and A novel deep learning model with a novel mechanism for processing complex data containing graph structures and sequence information. Mainly for graph structure data, it can find node neighbors and their relationships. It excels in natural language processing, can handle insanely long sequences and capture global dependencies, Fusion Powerful representation capabilities in graph structured data processing and Advantages in capturing long-range dependencies. The core technology of the graph includes node features and adjacency information in graph data. Its main principles include layer , Multi-head self-attention mechanism , Positional encoding , Graph Attention Layer .in It is widely used in the fields of social network analysis, molecular and drug discovery, knowledge graphs, question-answering systems, and recommendation systems.

[0039] like Figure 5 As shown, starting from the top visit node, the visit is composed of multiple types of feature information, and the situations are divided into: demographic information, visit time series, cough, headache, and diabetes. Among them, dextromethorphan, ibuprofen, aspirin, and insulin reflect the doctor's decision-making process.

[0040] The above has created an adjacency matrix for the patient's visit record visit , that is, the The modification aims to introduce the adjacency matrix and prior knowledge (in the form of attention mask to learn about electronic health records ( ). Among them, the adjacency matrix Middle The conditional probability of a node By the following Function calculation: (4); in, and They are Node query ( ) and keys ( )vector, yes The column dimensions of the trainable parameters in the module (i.e. the vector dimensions of the query and key), is a mask matrix used to introduce prior knowledge (such as some irrelevant nodes).

[0041] Further, Calculated by the following formula: (5); in, Indicates The node embedding of convolutions, Indicates indivual The trainable parameter matrices of the module are used to generate query and key vectors respectively.

[0042] Further, Calculated by the following formula: (6); in, Indicates The node embedding of convolutions, Indicates indivual The trainable parameter matrices of the module are used to generate query and key vectors respectively.

[0043] Further, Each convolution node in Calculated by the following formula: (7); (8); in, Indicates The node embedding of convolutions, is the first A convolutional multilayer perceptron is used to further transform the node embeddings. Indicates convolution conditional probability matrix, Indicates The node embedding of convolutions, Indicates indivual The module's trainable parameter matrix, used to generate values ​​( )vector.

[0044] Based on the above embodiment, in step S500, the patient's mortality rate is predicted by using the prior-guided attention mechanism, and the output of the prior-guided attention mechanism is subjected to an attention pooling operation to obtain the patient's mortality rate.

[0045] Furthermore, before the attention pooling operation, it also includes: The output of the prior-guided attention mechanism is linearly transformed and then processed by the ReLU activation function; the output processed by the ReLU activation function is residually processed and then normalized.

[0046] Based on the above embodiment, it also includes weighted processing of the diagnostic ontology embedding matrix, the treatment ontology embedding matrix and the diagnostic information and treatment information of the feature map processed by the masked attention mechanism, and the features after pooling are used for the classification task of the medical ontology.

[0047] Based on the above embodiment, it also includes the classification task of using the diagnosis ontology embedding matrix and the treatment ontology embedding matrix for the medical ontology.

[0048] In summary, this embodiment obtains a medical ontology by processing structured data and unstructured data in EHR, and then obtains different vector representations of the relevant information of the medical ontology through an encoder, and then obtains prediction indicators through a prediction model, thereby improving the accuracy of mortality prediction in the current medical field. This method has a high prediction accuracy, and at the same time, the model expression ability of this method is strong, the training stability is high, and the gradient vanishing problem in the deep network is effectively avoided.

[0049] Embodiment 2: This embodiment provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the mortality prediction method based on the fusion of medical ontology embedding and attention mechanism as described in Example 1.

[0050] Those skilled in the art can understand that all or part of the features / steps of the above-mentioned method embodiments can be implemented by methods, data processing systems or computer programs, and these features can be implemented without hardware, entirely by software, or by a combination of hardware and software. The aforementioned computer program can be stored in one or more computer-readable storage media, and the storage medium stores a computer program. When the computer program is executed (such as by a processor), the steps of the above-mentioned method embodiment of mortality prediction based on the fusion of medical ontology embedding and attention mechanism are executed.

[0051] The aforementioned storage media that can store computer programs / instruction codes include: static hard disks, solid-state hard disks, random access memories (SRAM), electrically erasable programmable read-only memories (EEPROM), erasable programmable read-only memories (EPROM), programmable read-only memories (PROM), read-only memories (ROM), optical storage devices, magnetic storage devices, flash memories, magnetic disks or optical disks and / or combinations of the above devices, that is, they can be implemented by any type of volatile or non-volatile storage devices or combinations thereof.

[0052] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0053] The above description is only the preferred embodiment of the present invention. It is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the protection scope of the present invention.

Claims

1. A mortality prediction method based on the fusion of medical ontology embedding and attention mechanism, characterized in that: The steps include: The patient's demographic information, diagnosis information, treatment information, and vital signs information are feature mapped as a demographic information node, a diagnosis information node, a treatment information node, and a vital signs information node, respectively; Constructing medical ontologies for the diagnosis information nodes and treatment information nodes of the feature mapping respectively, and obtaining the diagnosis ontology embedding matrix and treatment ontology embedding matrix of the patient; The diagnosis ontology embedding matrix, the treatment ontology embedding matrix, and the diagnosis information and treatment information of the feature map are processed through a masked attention mechanism to obtain a diagnosis sequence and a treatment sequence of the patient; Obtaining corresponding treatment results based on the patient's medical visit records, learning the causal relationship between diagnosis and treatment, calculating the conditional probability between the patient's diagnosis results and treatment results, and obtaining an adjacency matrix; The patient's mortality rate is predicted based on the demographic information, vital sign information, the diagnosis sequence, the treatment sequence and the adjacency matrix of the feature map.

2. The mortality prediction method based on the fusion of medical ontology embedding and attention mechanism according to claim 1 is characterized in that: The patient's demographic information, diagnosis information, treatment information, and vital signs information are respectively used as demographic information nodes, diagnosis information nodes, treatment information nodes, and vital signs information nodes for feature mapping, including the following steps: After converting the patient's demographic information, diagnosis information, treatment information, and vital sign information into a continuous vector, a binary vector is generated by label encoding; The diagnostic information and treatment information encoded by the label are activated through a nonlinear function.

3. The mortality prediction method based on the fusion of medical ontology embedding and attention mechanism according to claim 1 is characterized in that: The medical ontology is a directed acyclic graph with a hierarchical structure that represents the diagnosis information nodes and the treatment information nodes in the form of a parent-child relationship, and the output form is the diagnosis ontology embedding matrix and the treatment ontology embedding matrix.

4. The mortality prediction method based on the fusion of medical ontology embedding and attention mechanism according to claim 1 is characterized in that: The diagnostic ontology embedding matrix and the treatment ontology embedding matrix are activated by a nonlinear function before being processed by the masked attention mechanism.

5. The mortality prediction method based on the fusion of medical ontology embedding and attention mechanism according to claim 4 is characterized in that: After the diagnosis ontology embedding matrix and the treatment ontology embedding matrix are processed by the masked attention mechanism, the following further includes: The output processed by the masked attention mechanism is sequentially subjected to linear function and ReLu function activation, residual connection, normalization, linear transformation, residual connection, and normalization to obtain the diagnosis sequence and treatment sequence of the patient.

6. The mortality prediction method based on the fusion of medical ontology embedding and attention mechanism according to claim 1 is characterized in that: The mortality rate of the patient is predicted by using a priori guided attention mechanism, and the output of the prior guided attention mechanism is subjected to an attention pooling operation to obtain the mortality rate of the patient.

7. The mortality prediction method based on the fusion of medical ontology embedding and attention mechanism according to claim 6 is characterized in that: Before the attention pooling operation, it also includes: The output of the prior-guided attention mechanism is linearly transformed and then processed by the ReLU activation function; the output processed by the ReLU activation function is residually processed and then normalized.

8. The mortality prediction method based on the fusion of medical ontology embedding and attention mechanism according to any one of claims 1 to 7, characterized in that: It also includes weighted processing of the diagnostic ontology embedding matrix, the treatment ontology embedding matrix and the diagnostic information and treatment information of the feature map processed by the masked attention mechanism, and using the features after pooling for the classification task of the medical ontology.

9. The mortality prediction method based on the fusion of medical ontology embedding and attention mechanism according to any one of claims 1 to 7, characterized in that: It also includes a classification task for the medical ontology by using the diagnostic ontology embedding matrix and the therapeutic ontology embedding matrix.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the mortality prediction method based on the fusion of medical ontology embedding and attention mechanism as described in any one of claims 1 to 9 are implemented.

Images