Medical table data feature semantic enhancement disease prediction method and system based on large language model

Abstract

The invention provides a disease prediction method and system based on medical table data feature semantic enhancement of a large language model, and belongs to the technical field of medical information processing. Processing the obtained medical table data to be processed by using a pre-trained disease prediction model; and obtaining a disease prediction result. According to the method, for feature names, a mixed expert system based on a plurality of medical large models is introduced, and the most suitable expert is dynamically selected through a routing mechanism to generate semantic embedding rich in clinical knowledge; for the characteristic value, generating interpretable embedding by simulating a decision path of a doctor on continuous numerical values; and the feature name and the value are integrated through a feature name-value dynamic interactive fusion module and are input into a multi-dimensional Transform integrated network for prediction. According to the method, the whole process of systematically injecting the deep medical priori knowledge into the table to represent learning is realized, and the accuracy, robustness and interpretability of the model in medical tasks such as disease prediction and the like are improved.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, specifically to a disease prediction method and system based on a large language model that can deeply integrate medical knowledge, effectively enhance feature semantic representation, and possess good interpretability for enhancing the feature semantics of medical tabular data. Background Technology

[0002] In the information society, data has become a key factor of production, and tabular data, as the most basic and ubiquitous carrier of structured and semi-structured information, constitutes the digital foundation of enterprise operations, scientific research, and public administration. Unlike unstructured data such as images and audio, tabular data organizes information in rows and columns, giving it inherent efficiency in storage, retrieval, and analysis. Industry estimates suggest that over 80% of data analysis tasks globally are based on tabular data, making its core position irreplaceable in areas such as financial risk control, supply chain management, and customer relationship management.

[0003] Among all applications of tabular data, medical tabular data, especially data generated by electronic health record (EHR) systems, holds a crucial and irreplaceable position. EHR forms are not only objective records of patient medical information but also dynamic life archives reflecting an individual's entire life process from health to disease, from diagnosis to treatment. Their importance is reflected on three levels: Clinical value: Data in EHRs, such as laboratory test results (e.g., blood glucose, cholesterol), vital signs (e.g., blood pressure, heart rate), imaging report summaries, drug prescriptions, and diagnostic codes, serve as direct evidence for doctors to diagnose diseases, develop treatment plans, and assess prognosis. Research value: Large-scale, cross-time-dimension EHR forms constitute a valuable resource for real-world studies (RWS) and evidence-based medicine, which can be used for disease trend analysis, drug efficacy comparison, and complication risk prediction, greatly promoting the progress of clinical medicine. Public health value: Through the analysis of population EHR form data, infectious disease surveillance, public health event early warning, and optimal allocation of medical resources can be achieved.

[0004] Despite the immense application value of medical tabular data, enabling machine models to deeply "understand" and effectively utilize such data remains a key technical challenge in this field. Existing technical solutions primarily suffer from the following limitations:

[0005] The semantic gap problem in traditional machine learning models: Traditional tree models and ensemble methods, represented by gradient boosting decision trees (such as XGBoost and LightGBM), while performing well in many table tasks, essentially treat each feature in the table (such as the table header "hemoglobin") as an independent, semantically meaningless symbol or index. The model cannot utilize the rich domain knowledge inherent in the feature names themselves (such as the clinical association between "hemoglobin" and anemia), resulting in a significant "semantic gap" between the model input and human expert knowledge. This limits the model's interpretability and its performance ceiling in complex medical logical reasoning.

[0006] Limitations of Deep Tabular Neural Networks: In recent years, deep neural networks such as TabNet and SAINT, specifically designed for tabular data, have emerged. These models have improved performance to some extent through structures such as attention mechanisms. However, in terms of feature initialization, they typically initialize feature names randomly or use only simple embedding techniques, failing to fully explore and utilize the rich semantic information inherent in the feature names. Their representation learning methods remain at the level of fitting surface patterns in the data, lacking deep integration with underlying medical prior knowledge.

[0007] The application of Large Language Models (LLMs) in the tabular domain is insufficient: With the development of Large Language Models (LLMs), techniques such as TableGPT have emerged that attempt to utilize LLMs to process tabular data. Existing methods typically tend to convert the entire table into a natural language description or use a general pre-trained model as the encoder. These methods often fail to handle the structured characteristics of tables with fine precision and have not specifically designed a mechanism for the medical field to dynamically and efficiently inject medical knowledge into the representation learning of feature names and feature values. In particular, for numerical features, existing methods lack an encoding mechanism with explicit medical interpretability that can simulate the decision-making path of clinicians.

[0008] In summary, existing technologies for representation learning of medical tabular data generally suffer from three core shortcomings: insufficient utilization of feature semantics, difficulty in injecting prior medical knowledge, and a lack of clinical interpretability in numerical feature encoding. This makes it difficult for existing models to truly "understand" medical tables, and their accuracy and reliability are limited when dealing with medical tasks that require deep semantic understanding and complex logical reasoning (such as early disease prediction and personalized efficacy assessment). Summary of the Invention

[0009] The purpose of this invention is to provide a disease prediction method and system based on the semantic enhancement of medical tabular data features using a large language model, in order to solve at least one of the technical problems existing in the background art.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] In a first aspect, the present invention provides a disease prediction method based on semantic enhancement of medical tabular data features using a large language model, comprising:

[0012] Retrieve medical form data to be processed;

[0013] The acquired medical table data is processed using a pre-trained disease prediction model to obtain disease prediction results. The pre-trained disease prediction model includes: a feature name semantic enhancement module, used to generate enhanced semantic embeddings of feature names through extraction from a large medical model embedding layer and integration with a hybrid expert system; a feature value embedding enhancement module, used to generate enhanced embeddings of feature values ​​by using embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values; a feature name-value dynamic interaction fusion module, used to adaptively fuse the enhanced semantic embeddings of feature names and the enhanced embeddings of feature values ​​to obtain a fused feature column embedding; and a prediction network, used to concatenate all fused feature column embeddings to form an initial patient representation vector, which is then processed by a multi-dimensional projection and Transformer ensemble network to obtain the final disease prediction result.

[0014] As a further limitation of the first aspect of the present invention, generating enhanced semantic embeddings for feature names includes: inputting each feature name in the medical table as a natural language phrase into the embedding layer or the underlying Transformer encoder of a pre-trained medical large language model, and extracting the initial semantic embedding vector of the feature name through forward propagation; using a routing network composed of a lightweight feedforward neural network as input, outputting a relevance weight probability distribution vector between the feature name and each expert; each expert independently processes the feature name to generate an expert embedding, and weighting and summing all expert embeddings using the weight probability distribution vector to obtain the enhanced semantic embedding of the feature name.

[0015] As a further limitation of the first aspect of the present invention, the enhanced embedding for generating feature values ​​includes: categorical feature value encoding: for each categorical feature column Maintain independent embedding matrices ,in This represents the number of all possible values ​​for this feature column. The embedding dimension for the categorical values ​​is obtained by looking up a table to retrieve the embedding vector corresponding to the categorical values.

[0016] As a further limitation of the first aspect of the present invention, the enhanced embedding for generating feature values ​​further includes: binary semantic tree encoding of numerical feature values: for each numerical feature column The construction depth is The binary semantic tree determines the initial split point based on the numerical statistical distribution, and the path of the value in the tree is encoded as a binary sequence. Calculate the relative position of the value within the leaf node interval. Numerical base coding is obtained by fusing path coding and position scores; hard label vectors are generated based on feature-based clinical normal ranges, and a clinical state prediction head is constructed through a multilayer perceptron to minimize the cross-entropy loss between the prediction results and the self-supervised labels; after joint training, the numerical base coding... This is the final numerical embedding.

[0017] As a further limitation of the first aspect of the present invention, adaptive fusion of feature name-enhanced semantic embedding and feature value-enhanced embedding to obtain feature column fusion embedding includes: dynamically adjusting the contribution ratio of feature name-enhanced semantic embedding and feature value-enhanced embedding through a gating network to generate feature column fusion embedding; and including All feature columns of a patient sample with a given feature are fused, embedded, and concatenated to form an initial representation vector.

[0018] As a further limitation of the first aspect of the present invention, the multi-dimensional projection and Transformer integrated network processing includes: multi-dimensional projection: through Each independent projection layer will initialize the representation vector. Mapped to Different semantic subspaces; Transformer ensemble prediction: representation after each projection The input is a separate Transformer encoder for deep feature interaction. After pooling, the results are fed into the classification head to generate preliminary predictions. All preliminary predictions are then integrated through mode voting to obtain the final prediction.

[0019] Secondly, the present invention provides a disease prediction system based on semantic enhancement of medical tabular data features using a large language model, comprising:

[0020] The acquisition module is used to acquire medical form data to be processed;

[0021] The processing module is used to process the acquired medical table data to be processed using a pre-trained disease prediction model to obtain disease prediction results. The pre-trained disease prediction model includes: a feature name semantic enhancement module, used to generate enhanced semantic embeddings of feature names through extraction from a large medical model embedding layer and integration with a hybrid expert system; a feature value embedding enhancement module, used to generate enhanced embeddings of feature values ​​by using embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values; a feature name-value dynamic interaction fusion module, used to adaptively fuse the enhanced semantic embeddings of feature names and the enhanced embeddings of feature values ​​to obtain a fused feature column embedding; and a prediction network, used to concatenate all fused feature column embeddings to form an initial patient representation vector, which is then processed by a multi-dimensional projection and Transformer ensemble network to obtain the final disease prediction result.

[0022] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the disease prediction method based on the semantic enhancement of medical tabular data features according to the first aspect.

[0023] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the disease prediction method based on large language model-based medical tabular data feature semantic enhancement as described in the first aspect.

[0024] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the disease prediction method based on large language model-based medical tabular data feature semantic enhancement as described in the first aspect.

[0025] The beneficial effects of this invention are as follows: By using a hybrid expert system, stable injection of medical knowledge and task adaptation are achieved, thus solving the semantic gap problem; the binary semantic tree encoding of numerical features enables the representation to have clear clinical interpretability and conforms to the doctor's decision-making logic; and the multi-dimensional integration strategy improves the robustness of the model.

[0026] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, 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 effort.

[0028] Figure 1 This is a framework diagram of a disease prediction model based on large language model for enhancing the semantic features of medical tabular data, as described in an embodiment of the present invention.

[0029] Figure 2 This is a flowchart of the numerical feature encoding process described in an embodiment of the present invention.

[0030] Figure 3 This is a flowchart of the disease prediction method based on large language model for semantic enhancement of medical tabular data features, as described in an embodiment of the present invention. Detailed Implementation

[0031] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0032] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0033] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0034] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0035] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0036] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0037] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0038] This invention proposes a disease prediction method based on large language models for semantic enhancement of medical tabular data features, addressing the semantic gap and insufficient utilization of medical knowledge in existing technologies. The core of this invention lies in constructing a dual-branch semantic enhancement framework, which performs deep encoding on feature names and feature values ​​separately. For feature names, a hybrid expert system based on multiple large medical models is innovatively introduced, dynamically selecting the most suitable expert through a routing mechanism to generate semantic embeddings rich in clinical knowledge. For feature values, especially numerical features, a binary semantic tree-based encoding method is proposed, generating interpretable embeddings by simulating the decision-making path of doctors on continuous numerical values. Finally, the two are integrated through a dynamic interaction fusion module of feature name and value, and input into a multi-dimensional Transformer ensemble network for prediction. This invention is the first to systematically inject deep medical prior knowledge into the entire process of tabular representation learning, significantly improving the accuracy, robustness, and interpretability of the model in medical tasks such as disease prediction. As shown in Table 1 (experimental results on the Alzheimer's disease dataset) and Table 2 (experimental results on the ABIDE dataset), in the early disease prediction task, the model provided by this invention improves the accuracy by an average of 0.79 percentage points compared to CatBoost, an average of 0.57 percentage points compared to FT-Transformer, and an average of 4.38 percentage points compared to ModernNCA on the two medical datasets. At the same time, it can clearly trace the semantic association of features, assisting clinical decision-making.

[0039] Table 1

[0040] Model Method AUC Acc F1 The method of this invention (multiple models) 0.9513 0.9534 0.9489 The method of this invention (single large model) 0.9522 0.9488 0.9440 FT-Transformer 0.9543 0.9465 0.9412 catboost 0.9384 0.9511 0.9464 xgboost 0.9393 0.9395 0.9338 TabM 0.9118 0.8581 0.8450 Excelformer 0.9386 0.8488 0.8488 ModernNCA 0.9210 0.8883 0.8400 MLP 0.8765 0.8326 0.7677 MLP_PLR 0.9344 0.9233 0.8896 TabTransformer 0.8240 0.7791 0.6965

[0041] Table 2

[0042] Model Method AUC Acc F1 Our approach (how large a model) 0.9929 0.9641 0.9640 Our approach (single large model) 0.9785 0.9551 0.9550 FT-Transformer 0.9887 0.9596 0.9595 catboost 0.9833 0.9506 0.9505 xgboost 0.9890 0.9372 0.9369 TabM 0.9853 0.9506 0.9506 ExcelFormer 0.9905 0.7803 0.7803 ModernNCA 0.9808 0.9417 0.9422 MLP 0.9912 0.9506 0.9528 MLP_PLR 0.9957 0.9596 0.9614 TabTransformer 0.9845 0.9372 0.9421

[0043] Example 1

[0044] In this embodiment 1, a disease prediction system based on a large language model for enhancing the semantic features of medical table data is first provided. The system includes: an acquisition module for acquiring medical table data to be processed; a processing module for processing the acquired medical table data using a pre-trained disease prediction model; and obtaining a disease prediction result. The pre-trained disease prediction model includes: a feature name semantic enhancement module for generating enhanced semantic embeddings of feature names through extraction from a large medical model embedding layer and integration with a hybrid expert system; a feature value embedding enhancement module for generating enhanced embeddings of feature values ​​by using embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values; a feature name-value dynamic interaction fusion module for adaptively fusing the enhanced semantic embeddings of feature names and the enhanced embeddings of feature values ​​to obtain a fused feature column embedding; and a prediction network for concatenating all fused feature column embeddings to form an initial patient representation vector, which is then processed by a multi-dimensional projection and Transformer ensemble network to obtain the final disease prediction result.

[0045] like Figure 3 As shown, in this embodiment, the above-described system is used to implement a disease prediction method based on the semantic enhancement of medical table data features using a large language model. The method includes: acquiring medical table data to be processed; processing the acquired medical table data using a pre-trained disease prediction model; and obtaining a disease prediction result. The pre-trained disease prediction model includes: a feature name semantic enhancement module, used to generate enhanced semantic embeddings of feature names through extraction from a large medical model embedding layer and integration with a hybrid expert system; a feature value embedding enhancement module, used to generate enhanced embeddings of feature values ​​by using embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values; a feature name-value dynamic interaction fusion module, used to adaptively fuse the enhanced semantic embeddings of feature names and the enhanced embeddings of feature values ​​to obtain a fused feature column embedding; and a prediction network, used to concatenate all fused feature column embeddings to form an initial patient representation vector, which is then processed by a multi-dimensional projection and Transformer ensemble network to obtain the final disease prediction result.

[0046] The enhanced semantic embedding of feature names includes: inputting each feature name in the medical table as a natural language phrase into the embedding layer or the underlying Transformer encoder of a pre-trained medical large language model, and extracting the initial semantic embedding vector of the feature name through forward propagation; using a routing network composed of a lightweight feedforward neural network as input, taking the original text or initial embedding of the feature name as input, and outputting the relevance weight probability distribution vector between the feature name and each expert; each expert independently processes the feature name to generate expert embeddings, and the weighted sum of all expert embeddings is obtained by weighting the weight probability distribution vector.

[0047] Specifically, the steps for learning semantically enhanced representations of feature names are as follows:

[0048] (1) Medical Large Model Embedding Layer Extraction: Each feature name in the medical table is treated as a natural language phrase and input into the embedding layer or the bottom Transformer encoder of the pre-trained medical large language model. The initial semantic embedding vector of the feature name is extracted through forward propagation, which is formally represented as:

[0049] ;

[0050] in, For feature name, For embedding functions of a medical large language model, This is the extracted initial semantic embedding vector;

[0051] (2) Construction and semantic integration of hybrid expert systems:

[0052] ① Expert pool construction: Maintain an expert pool containing four pre-trained medical large language models (BioBert, ClinicalBert, BioGPT, MedQwen2), which is divided into a frozen expert group (with fixed parameters) and a trainable expert group (with fine-tunable parameters).

[0053] ② Dynamic route calculation: A routing network composed of a lightweight feedforward neural network takes the original text or initial embedding of the feature name as input and outputs a probability distribution vector of the relevance weights of the feature name and each expert. The formula is:

[0054] ;

[0055] in, and These are the learnable parameters for the routing network. For pooling operations, The weighted probability distribution vector;

[0056] ③ Weighted ensemble to generate the final embedding: Each expert independently processes the feature names to generate an expert embedding. The weighted sum of all expert embeddings is then performed using the weight probability distribution vector to obtain the feature name-enhanced semantic embedding. The formula is as follows:

[0057] ;

[0058] in, For the number of experts, For the first The weight of each expert, For the first An expert-generated embedding.

[0059] The enhanced embedding for generating feature values ​​includes: categorical feature value encoding: for each categorical feature column Maintain independent embedding matrices ,in This represents the number of all possible values ​​for this feature column. For categorical value embedding, the embedding vector corresponding to the categorical value is obtained through a table lookup operation. The enhanced embedding of feature values ​​also includes: binary semantic tree encoding of numerical feature values: for each numerical feature column... The construction depth is The binary semantic tree determines the initial split point based on the numerical statistical distribution, and the path of the value in the tree is encoded as a binary sequence. Calculate the relative position of the value within the leaf node interval. Numerical base coding is obtained by fusing path coding and position scores; hard label vectors are generated based on feature-based clinical normal ranges, and a clinical state prediction head is constructed through a multilayer perceptron to minimize the cross-entropy loss between the prediction results and the self-supervised labels; after joint training, the numerical base coding... This is the final numerical embedding.

[0060] Specifically, the steps for feature value embedding augmentation representation learning are as follows:

[0061] (1) Categorical feature value encoding: for each categorical feature column Maintain independent embedding matrices ,in This represents the number of all possible values ​​for this feature column. For the categorical value embedding dimension, the embedding vector corresponding to the categorical value is obtained through a table lookup operation, as shown in the formula:

[0062] ;

[0063] in, For the first The sample at the th Category values ​​of the column, This is a mapping function from categorical values ​​to indices in the embedding matrix;

[0064] (2) Binary semantic tree encoding of numerical feature values:

[0065] ① Binary semantic tree construction: for each numerical feature column The construction depth is The binary semantic tree determines the initial split point based on the numerical statistical distribution, and the path of the value in the tree is encoded as a binary sequence. Calculate the relative position of the value within the leaf node interval. The numerical base code is obtained by fusing path coding and position scores:

[0066] ;

[0067] in, and These are learnable parameters;

[0068] ② Self-supervised enhancement of clinical semantics: feature-based clinical normal range Generate hard tag vectors A clinical state prediction head is constructed using a multilayer perceptron, minimizing the cross-entropy loss between the prediction results and the self-supervised labels. The total model loss is: ;in, Loss to the main task For self-monitoring losses, To weigh parameters;

[0069] ③ Final numerical embedding generation: After joint training, the numerical base encoding is generated. That is, the final numerical embedding .

[0070] The adaptive fusion of feature name-enhanced semantic embedding and feature value-enhanced embedding to obtain feature column fusion embedding includes: dynamically adjusting the contribution ratio of feature name-enhanced semantic embedding and feature value-enhanced embedding through a gating network to generate feature column fusion embedding; and including... All feature columns of a patient sample with a given feature are fused, embedded, and concatenated to form an initial representation vector.

[0071] Specifically, the steps for dynamic interaction and fusion of feature names and values ​​are as follows:

[0072] (1) Gated fusion calculation: The contribution ratio of feature name-enhanced semantic embedding and feature value-enhanced embedding is dynamically adjusted through a gating network to generate feature column fusion embedding, the formula of which is:

[0073] ;

[0074] ;

[0075] in, and These are the learnable parameters of the gated network. It is the Sigmoid activation function. For element-wise multiplication, This is a vector concatenation operation. Enhance the embedding for eigenvalues. or , Feature column fusion embedding;

[0076] (2) Initial patient representation construction: will include The initial representation vector is formed by fusing and embedding all feature columns of a patient sample with each feature. .

[0077] Processed via a multi-dimensional projection and Transformer integrated network, including: Multi-dimensional projection: through Each independent projection layer will initialize the representation vector. Mapped to Different semantic subspaces; Transformer ensemble prediction: representation after each projection The input is a separate Transformer encoder for deep feature interaction. After pooling, the results are fed into the classification head to generate preliminary predictions. All preliminary predictions are then integrated through mode voting to obtain the final prediction.

[0078] Specifically, the steps for integrating multi-dimensional projection with Transformer are as follows:

[0079] (1) Multidimensional projection: through Each independent projection layer will initialize the representation vector. Mapped to There are three different semantic subspaces, and the formula is:

[0080] ;

[0081] in, and For the first Learnable parameters for each projection layer;

[0082] (2) Transformer ensemble prediction: Representation after each projection The input is a separate Transformer encoder for deep feature interaction, which is then pooled and fed into the classification head to generate preliminary predictions. All preliminary predictions are then integrated through mode voting to obtain the final prediction. ;in, For the first The prediction results corresponding to each Transformer encoder. It is the mode function.

[0083] Example 2

[0084] In this embodiment 2, a method for enhancing the semantic features of medical table data based on a large language model is provided, which includes the following steps: (1) Preprocessing the input medical table data to clarify the feature names and feature values ​​in the data and distinguish the types of feature values ​​(categorical or numerical) to lay the foundation for subsequent semantic enhancement processing; (2) Constructing a feature name semantic enhancement representation learning module, which extracts the feature name through the medical large model embedding layer and integrates with the hybrid expert system to generate the enhanced semantic embedding of the feature name; (3) Constructing a feature value embedding enhancement representation learning module, which uses embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values ​​to generate the enhanced embedding of the feature value; (4) Using the feature name-value dynamic interaction fusion module, the enhanced semantic embedding of the feature name and the enhanced embedding of the feature value are adaptively fused to obtain the feature column fusion embedding.

[0085] The specific steps of feature name semantic enhancement representation learning in step (2) are as follows:

[0086] (1) Medical Large Model Embedding Layer Extraction: Each feature name in the medical table is treated as a natural language phrase and input into the embedding layer or the bottom Transformer encoder of the pre-trained medical large language model. The initial semantic embedding vector of the feature name is extracted through forward propagation, which is formally represented as:

[0087]

[0088] in, For feature name, For embedding functions of a medical large language model, This is the extracted initial semantic embedding vector;

[0089] (2) Construction and semantic integration of hybrid expert systems:

[0090] ① Expert pool construction: Maintain an expert pool containing four pre-trained medical large language models (BioBert, ClinicalBert, BioGPT, MedQwen2), which is divided into a frozen expert group (with fixed parameters) and a trainable expert group (with fine-tunable parameters).

[0091] ② Dynamic route calculation: A routing network composed of a lightweight feedforward neural network takes the original text or initial embedding of the feature name as input and outputs a probability distribution vector of the relevance weights of the feature name and each expert. The formula is:

[0092]

[0093] in, and These are the learnable parameters for the routing network. For pooling operations, The weighted probability distribution vector;

[0094] ③ Weighted ensemble to generate the final embedding: Each expert independently processes the feature names to generate an expert embedding. The weighted sum of all expert embeddings is then performed using the weight probability distribution vector to obtain the feature name-enhanced semantic embedding. The formula is as follows:

[0095]

[0096] in, For the number of experts, For the first The weight of each expert, For the first An expert-generated embedding.

[0097] The specific steps of feature embedding enhancement representation learning in step (3) are as follows:

[0098] (1) Categorical feature value encoding: for each categorical feature column Maintain independent embedding matrices ,in This represents the number of all possible values ​​for this feature column. For the categorical value embedding dimension, the embedding vector corresponding to the categorical value is obtained through a table lookup operation, as shown in the formula:

[0099]

[0100] in, For the first The sample at the th Category values ​​of the column, This is a mapping function from categorical values ​​to indices in the embedding matrix;

[0101] (2) Binary semantic tree encoding of numerical feature values:

[0102] ① Binary semantic tree construction: for each numerical feature column The construction depth is The binary semantic tree determines the initial split point based on the numerical statistical distribution, and the path of the value in the tree is encoded as a binary sequence. Calculate the relative position of the value within the leaf node interval. The numerical base code is obtained by fusing path coding and position scores:

[0103]

[0104] in, and These are learnable parameters;

[0105] ② Self-supervised enhancement of clinical semantics: feature-based clinical normal range Generate hard tag vectors A clinical state prediction head is constructed using a multilayer perceptron, minimizing the cross-entropy loss between the prediction results and the self-supervised labels. The total model loss is:

[0106]

[0107] in, Loss to the main task For self-monitoring losses, To weigh parameters;

[0108] ③ Final numerical embedding generation: After joint training, the numerical base encoding is generated. That is, the final numerical embedding .

[0109] The specific steps of the dynamic interaction fusion of feature name and value in step (4) are as follows:

[0110] (1) Gated fusion calculation: The contribution ratio of feature name-enhanced semantic embedding and feature value-enhanced embedding is dynamically adjusted through a gating network to generate feature column fusion embedding, the formula of which is:

[0111]

[0112]

[0113] in, and These are the learnable parameters of the gated network. It is the Sigmoid activation function. For element-wise multiplication, This is a vector concatenation operation. Enhance the embedding for eigenvalues. or , Feature column fusion embedding;

[0114] (2) Initial patient representation construction: will include The initial representation vector is formed by fusing and embedding all feature columns of a patient sample with each feature. .

[0115] The specific steps for integrating multi-dimensional projection with Transformer in step (5) are as follows:

[0116] (1) Multidimensional projection: through Each independent projection layer will initialize the representation vector. Mapped to There are three different semantic subspaces, and the formula is:

[0117]

[0118] in, and For the first Learnable parameters for each projection layer;

[0119] (2) Transformer ensemble prediction: Representation after each projection The input is a separate Transformer encoder for deep feature interaction, which is then pooled and fed into the classification head to generate preliminary predictions. All preliminary predictions are then integrated through mode voting to obtain the final prediction.

[0120]

[0121] in, For the first The prediction results corresponding to each Transformer encoder. It is the mode function.

[0122] In practical applications, the medical form data can be data generated by an electronic health record (EHR) system, including laboratory test results, vital signs, imaging report summaries, drug prescriptions, diagnostic codes, and other information. In downstream medical-related tasks, all feature columns are fused, embedded, and concatenated to form an initial patient representation vector. After processing by a multi-dimensional projection and Transformer ensemble network, the final prediction result is output (which may include disease diagnosis, prognosis prediction, complication risk assessment, etc.).

[0123] Example 3

[0124] like Figure 1 , Figure 2 As shown in this embodiment, a semantic enhancement method for medical tabular data features based on a large language model is proposed. A two-branch semantic enhancement framework is constructed to perform deep encoding on feature names and feature values ​​separately. For feature names, a hybrid expert system based on multiple large medical models is introduced, which dynamically selects the most suitable expert through a routing mechanism to generate semantic embeddings rich in clinical knowledge. For feature values, especially numerical features, an encoding method based on a binary semantic tree is proposed, which generates interpretable embeddings by simulating the decision-making path of doctors on continuous numerical values. Finally, the two are integrated through a feature name-value dynamic interaction fusion module and input into a multi-dimensional Transformer ensemble network for prediction.

[0125] In this embodiment, the large language model includes:

[0126] Feature Name Semantic Augmentation Representation Learning Module: This module aims to inject rich medical prior knowledge into the model's understanding of table feature names. The core of this module lies in leveraging the powerful semantic encoding capabilities of a pre-trained medical large-scale language model and designing an innovative ensemble mechanism to dynamically and multi-dimensionally generate embedded representations of feature names. The specific steps are as follows:

[0127] 1.1 Embedding layer extraction based on large medical models

[0128] First, for any feature name in the input medical form (e.g., "hemoglobin concentration" or "white blood cell count"), we treat it as a natural language phrase. This invention abandons the traditional approach of randomly initializing feature names or using simple one-hot encoding, instead utilizing multiple pre-trained large language models in the medical field as the underlying embedding encoder.

[0129] Specifically, for a feature name We input this into an embedding layer or a low-level Transformer encoder of a pre-trained medical LLM. Through the forward propagation of this model, we can extract a deep semantic representation of the feature name in a specific medical context. This process can be formally represented as:

[0130]

[0131] in, It is a feature name The initial semantic embedding vector extracted from a single medical LLM. This vector encapsulates the statistical semantics and knowledge associations learned by the medical term from the pre-training corpus.

[0132] 1.2 Multi-model Semantic Integration Based on Hybrid Expert Systems

[0133] To overcome the potential limitations of a single model in terms of perspective or knowledge bias, and to capture the clinical semantics of feature names more comprehensively and robustly, this embodiment further proposes a multi-model ensemble mechanism based on a hybrid expert system. This system consists of four distinct "experts," each an independent, pre-trained medical large language model (BioBert, ClinicalBert, BioGPT, MedQwen2).

[0134] The design and workflow of this hybrid expert system are as follows:

[0135] (1) Expert Pool Construction: Maintain an expert pool containing four pre-trained medical LLMs. For computational efficiency and stable knowledge injection considerations, the expert pool is divided into two groups:

[0136] Freeze Expert Group: The model parameters in this group are completely frozen during training and are not updated. They serve as stable "knowledge anchors," ensuring that the model always has access to pure medical prior knowledge untouched by task data.

[0137] Trainable expert groups: Models in this group are allowed to fine-tune some or all of their parameters in downstream tasks. This enables the model to adaptively adjust its knowledge representation to better fit the semantic needs of the current specific task, such as the prediction of a specific disease.

[0138] (2) Dynamic routing and aggregation: For each input feature name Instead of using all experts equally, the system dynamically calculates the relevance weights between the feature name and each expert through a learnable routing network (a lightweight feedforward neural network). The routing network takes the original text or initial embedding of the feature name as input and outputs a K-dimensional probability distribution vector. :

[0139]

[0140] in, This represents the importance weight of the k-th expert to the current feature name.

[0141] (3) Weighted integration and final embedding generation: Each expert Independent processing of feature names Generate its corresponding expert embedding The final feature name-enhanced semantic embedding The weighted sum of all expert embeddings is obtained as follows:

[0142]

[0143] The innovation of this step lies in the fact that, for the first time, a large-scale medical model designed for natural language processing is systematically applied to the semantic understanding of feature names in structured tabular data through a trainable, dynamically selected hybrid expert mechanism. By combining "frozen experts" and "trainable experts," this method maintains knowledge stability while also being task-adaptive, thereby generating feature name representations that are more informative and better suited to the needs of downstream medical tasks, laying a strong semantic foundation for subsequent numerical encoding and feature interaction.

[0144] Feature embedding enhancement module: In view of the heterogeneity (categorical and numerical values) of feature values ​​in medical tabular data, a differentiated encoding strategy is adopted to generate feature value representations that are rich in semantics and clinically interpretable.

[0145] 2.1 Embedding and Coding of Classification Feature Values

[0146] Maintain an independent embedding matrix for each categorical feature column j. ,in It is the number of all possible values ​​for this feature column (i.e., vocabulary size). This is the dimension of the categorical value embedding. When a specific categorical value is input... When (representing the value of the i-th sample in the j-th column), its corresponding embedding vector is obtained through a table lookup operation:

[0147]

[0148] in, It is a function that maps classification values ​​to their corresponding indices in the embedding matrix. This embedding vector... During model training, backpropagation will be used for optimization to learn the distributed representation of the classification value in a specific medical context.

[0149] 2.2 Binary Semantic Tree Coding of Numerical Features

[0150] For numerical features (e.g., age 65 years, blood glucose level 7.2 mmol / L), this paper abandons traditional normalization or simple bucketing methods and proposes a representation learning method based on binary semantic tree encoding. The core of this method lies in capturing the hierarchical semantics of numerical values ​​through a binary tree structure and introducing a self-supervised task, so that the learned embedding representation can automatically reflect the clinical status of the numerical value (e.g., normal or abnormal).

[0151] 2.2.1 Construction of Binary Semantic Tree

[0152] Construct a binary semantic tree of depth L for each numerical feature column j. .

[0153] The initial split point is determined based on the statistical distribution (median) of the feature column values. Given values... Its path encoding in the tree is a binary sequence. :

[0154]

[0155] in, This indicates that at the l-th level of the tree, according to The size relationship with the current node's split point determines whether to enter the left subtree (0) or the right subtree (1).

[0156] When the path reaches the leaf node interval of the Lth level Then, the relative position of the value within this finest granularity interval is further calculated:

[0157]

[0158] Final numerical base encoding It is formed by combining path encoding and position score:

[0159]

[0160] in, and These are learnable parameters.

[0161] 2.2.2 Self-supervised clinical semantic enhancement

[0162] To enable the model to understand the clinical semantics of numerical values, a clinical state self-supervised task is introduced, using known clinical reference ranges as supervisory signals to guide the model in learning to identify whether a numerical value is "normal".

[0163] For each value The known clinical normal range of its associated characteristic j Automatically generate a hard-label vector :

[0164]

[0165] Among them, II ( ) is an indicator function. To enhance robustness, this label can be smoothed.

[0166] In basic encoding On top of this, a lightweight multilayer perceptron (MLP) is added as the clinical state prediction head:

[0167]

[0168] The goal of this prediction head is to approximate the generated self-supervised labels. .

[0169] The model's total loss consists of the main task loss. and clinical status self-supervision loss Together they constitute:

[0170] in, , It is a parameter that weighs the importance of two tasks.

[0171] 2.2.3 Final Numerical Embedding Generation

[0172] After the above joint training, the model learns the final numerical embedding. It will simultaneously contain hierarchical information of the original numerical values ​​and their clinical state semantics:

[0173]

[0174] Here Is it being optimized simultaneously? and The data is updated during the backpropagation process, so the geometric relationships in its vector space will spontaneously form meaningful clusters based on clinical status (normal / abnormal) without having to hard-specify clinical thresholds during the tree construction phase.

[0175] 3. Feature name-value dynamic fusion and multi-dimensional integrated prediction

[0176] Enhanced semantic embedding for obtaining feature names Corresponding embeddings of eigenvalues (From the classification or numerical encoding module) This embodiment proposes a dynamic interactive fusion module based on a gating mechanism to generate a fusion representation that can fully characterize the semantics of the "feature-value" combination.

[0177] 3.1 Dynamic Interactive Fusion of Feature Name-Value

[0178] To establish a deep interaction between feature name semantics and feature value information, a lightweight gated fusion network is introduced. This network takes the two aforementioned embeddings as input and dynamically adjusts their contribution ratios through learnable gating weights to generate a fused feature column embedding. :

[0179]

[0180]

[0181] in, and These are the learnable parameters of the gated network. It is the Sigmoid activation function. This indicates element-wise multiplication; This represents a vector concatenation operation. This gating mechanism allows the model to adaptively decide, based on the specific combination of feature names and feature values, whether to emphasize the global semantic information of the feature names or the specific instance information of the feature values ​​in the final representation.

[0182] For a patient sample containing N features, all feature columns are fused, embedded, and concatenated to form the patient's initial representation vector. :

[0183]

[0184] 3.2 Multi-dimensional projection and Transformer integration

[0185] To model the complexity of patient representations from different subspace perspectives, a multi-dimensional ensemble learning strategy is proposed. First, K independent projection layers are used to... Mapped to K distinct semantic subspaces:

[0186]

[0187] Each of them and This corresponds to a projection transformation in a specific dimension.

[0188] The representation after each projection It is fed into a separate Transformer encoder for deep feature interaction and context modeling:

[0189]

[0190] The output of each Transformer encoder is pooled and then fed into the corresponding classification head to produce a preliminary prediction result. :

[0191]

[0192] in, The function returns the mode category. This ensemble strategy effectively mitigates the bias that may exist in a single model, enhancing the stability and reliability of the final prediction results.

[0193] By using a series design of feature name-value dynamic fusion, multi-dimensional projection learning and ensemble prediction, end-to-end modeling from raw medical tabular data to the final clinical prediction task is achieved. While maintaining the semantic interpretability of each module, the performance of the model in downstream tasks such as disease classification and prognosis prediction is significantly improved.

[0194] Example 4

[0195] This embodiment 4 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the disease prediction method based on the semantic enhancement of medical tabular data features according to a large language model, as described above. The method includes:

[0196] Retrieve medical form data to be processed;

[0197] The acquired medical table data is processed using a pre-trained disease prediction model to obtain disease prediction results. The pre-trained disease prediction model includes: a feature name semantic enhancement module, used to generate enhanced semantic embeddings of feature names through extraction from a large medical model embedding layer and integration with a hybrid expert system; a feature value embedding enhancement module, used to generate enhanced embeddings of feature values ​​by using embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values; a feature name-value dynamic interaction fusion module, used to adaptively fuse the enhanced semantic embeddings of feature names and the enhanced embeddings of feature values ​​to obtain a fused feature column embedding; and a prediction network, used to concatenate all fused feature column embeddings to form an initial patient representation vector, which is then processed by a multi-dimensional projection and Transformer ensemble network to obtain the final disease prediction result.

[0198] Example 5

[0199] This embodiment 5 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the disease prediction method based on the semantic enhancement of medical tabular data features according to a large language model as described above. The method includes:

[0200] Retrieve medical form data to be processed;

[0201] The acquired medical table data is processed using a pre-trained disease prediction model to obtain disease prediction results. The pre-trained disease prediction model includes: a feature name semantic enhancement module, used to generate enhanced semantic embeddings of feature names through extraction from a large medical model embedding layer and integration with a hybrid expert system; a feature value embedding enhancement module, used to generate enhanced embeddings of feature values ​​by using embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values; a feature name-value dynamic interaction fusion module, used to adaptively fuse the enhanced semantic embeddings of feature names and the enhanced embeddings of feature values ​​to obtain a fused feature column embedding; and a prediction network, used to concatenate all fused feature column embeddings to form an initial patient representation vector, which is then processed by a multi-dimensional projection and Transformer ensemble network to obtain the final disease prediction result.

[0202] Example 6

[0203] This embodiment 6 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the disease prediction method based on the semantic enhancement of medical tabular data features as described above. The method includes:

[0204] Retrieve medical form data to be processed;

[0205] The acquired medical table data is processed using a pre-trained disease prediction model to obtain disease prediction results. The pre-trained disease prediction model includes: a feature name semantic enhancement module, used to generate enhanced semantic embeddings of feature names through extraction from a large medical model embedding layer and integration with a hybrid expert system; a feature value embedding enhancement module, used to generate enhanced embeddings of feature values ​​by using embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values; a feature name-value dynamic interaction fusion module, used to adaptively fuse the enhanced semantic embeddings of feature names and the enhanced embeddings of feature values ​​to obtain a fused feature column embedding; and a prediction network, used to concatenate all fused feature column embeddings to form an initial patient representation vector, which is then processed by a multi-dimensional projection and Transformer ensemble network to obtain the final disease prediction result.

[0206] 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.

[0207] 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.

[0208] 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.

[0209] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed 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.

[0210] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A disease prediction method based on semantic enhancement of medical tabular data features using a large language model, characterized in that, include: Retrieve medical form data to be processed; The acquired medical table data is processed using a pre-trained disease prediction model. The disease prediction result is obtained; wherein, the pre-trained disease prediction model includes: a feature name semantic enhancement module, used to generate enhanced semantic embeddings of feature names by extracting through a medical large model embedding layer and integrating with a hybrid expert system; a feature value embedding enhancement module, used to generate enhanced embeddings of feature values ​​by using embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values; a feature name-value dynamic interaction fusion module, used to adaptively fuse the enhanced semantic embeddings of feature names and the enhanced embeddings of feature values ​​to obtain a feature column fusion embedding; and a prediction network, used to concatenate all feature column fusion embeddings to form an initial patient representation vector, which is then processed by a multi-dimensional projection and Transformer ensemble network to obtain the final disease prediction result.

2. The disease prediction method based on large language model for semantic enhancement of medical tabular data features according to claim 1, characterized in that, The process of generating enhanced semantic embeddings for feature names includes: inputting each feature name in the medical table as a natural language phrase into the embedding layer or the underlying Transformer encoder of a pre-trained medical large language model, and extracting the initial semantic embedding vector of the feature name through forward propagation; using a routing network composed of a lightweight feedforward neural network as input, taking the original text or initial embedding of the feature name as input, and outputting the relevance weight probability distribution vector between the feature name and each expert; each expert independently processes the feature name to generate expert embeddings, and the weighted sum of all expert embeddings is obtained by weighting the weight probability distribution vector.

3. The disease prediction method based on large language model for semantic enhancement of medical tabular data features according to claim 1, characterized in that, Enhanced embeddings for generating feature values ​​include: categorical feature value encoding: for each categorical feature column Maintain independent embedding matrices ,in This represents the number of all possible values ​​for this feature column. The embedding dimension for the categorical values ​​is obtained by looking up a table to retrieve the embedding vector corresponding to the categorical values.

4. The disease prediction method based on large language model for semantic enhancement of medical tabular data features according to claim 3, characterized in that, The enhanced embedding for generating feature values ​​also includes: binary semantic tree encoding of numerical feature values: for each numerical feature column The construction depth is The binary semantic tree determines the initial split point based on the numerical statistical distribution, and the path of the value in the tree is encoded as a binary sequence. Calculate the relative position of the value within the leaf node interval. Numerical base coding is obtained by fusing path coding and position scores; hard label vectors are generated based on feature-based clinical normal ranges, and a clinical state prediction head is constructed through a multilayer perceptron to minimize the cross-entropy loss between the prediction results and the self-supervised labels; after joint training, the numerical base coding... This is the final numerical embedding.

5. The disease prediction method based on large language model for semantic enhancement of medical tabular data features according to claim 1, characterized in that, Adaptive fusion of feature name-enhanced semantic embedding and feature value-enhanced embedding yields a feature column fusion embedding. This includes: dynamically adjusting the contribution ratio of feature name-enhanced semantic embedding and feature value-enhanced embedding through a gating network to generate the feature column fusion embedding; and including... All feature columns of a patient sample with a given feature are fused, embedded, and concatenated to form an initial representation vector.

6. The disease prediction method based on large language model for semantic enhancement of medical tabular data features according to claim 1, characterized in that, Processed via a multi-dimensional projection and Transformer integrated network, including: Multi-dimensional projection: through Each independent projection layer will initialize the representation vector. Mapped to Different semantic subspaces; Transformer ensemble prediction: representation after each projection The input is a separate Transformer encoder for deep feature interaction. After pooling, the results are fed into the classification head to generate preliminary predictions. All preliminary predictions are then integrated through mode voting to obtain the final prediction.

7. A disease prediction system based on semantic enhancement of medical tabular data features using a large language model, characterized in that, include: The acquisition module is used to acquire medical form data to be processed; The processing module is used to process the acquired medical table data using a pre-trained disease prediction model. The disease prediction result is obtained; wherein, the pre-trained disease prediction model includes: a feature name semantic enhancement module, used to generate enhanced semantic embeddings of feature names by extracting through a medical large model embedding layer and integrating with a hybrid expert system; a feature value embedding enhancement module, used to generate enhanced embeddings of feature values ​​by using embedding matrix lookup table encoding for categorical feature values ​​and binary semantic tree encoding combined with self-supervised learning for numerical feature values; a feature name-value dynamic interaction fusion module, used to adaptively fuse the enhanced semantic embeddings of feature names and the enhanced embeddings of feature values ​​to obtain a feature column fusion embedding; and a prediction network, used to concatenate all feature column fusion embeddings to form an initial patient representation vector, which is then processed by a multi-dimensional projection and Transformer ensemble network to obtain the final disease prediction result.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the disease prediction method based on the semantic enhancement of medical tabular data features according to any one of claims 1-6.

9. A computer device, characterized in that, The method includes a memory and a processor, the processor and the memory communicating with each other, the memory storing program instructions executable by the processor, and the processor calling the program instructions to execute the disease prediction method based on the semantic enhancement of medical tabular data features according to any one of claims 1-6.

10. An electronic device, characterized in that, include: The electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the disease prediction method based on the semantic enhancement of medical tabular data features according to any one of claims 1-6.

Images