Medical diagnosis prediction method and system based on box embedding unified medical concept structure and semantics

Abstract

The invention provides a medical diagnosis prediction method and system based on a box embedding unified medical concept structure and semantics. The method specifically comprises the following steps: modeling a hierarchical structure driven by a medical ontology; modeling a patient medical record driven hierarchical structure; performing box modeling on the patient based on the information of the previous medical records; and quantifying the similarity of the patient box structure and the clinical classification code box structure by using the volume of the intersection part of the patient box structure and the clinical classification code box structure, and performing diagnostic prediction according to the similarity. According to the invention, more accurate diagnosis and prediction can be realized.

Description

Technical Field

[0001] The present invention relates to the medical field, and particularly to a medical diagnosis prediction method and system based on box embedding for unifying medical concept structures and semantics. Background Art

[0002] In order to provide personalized treatment plans for patients and improve their conditions, accurate diagnosis prediction based on electronic medical records has become a crucial task in modern healthcare. However, the widespread application of diagnosis prediction methods based on electronic medical records has raised the problem of potential leakage of patients' private data. Under the limitation of only being able to access specific patient datasets for few-shot learning, existing technologies are required to use only limited patient electronic medical record data to maintain prediction performance.

[0003] Existing language model-based methods solve the few-shot learning problem by incorporating semantic understanding of medical concepts. After pre-training on a large-scale medical corpus, medical entity embeddings are introduced to improve semantic representation and context understanding of medical concepts. However, existing language model-based methods usually cannot capture the structures of medical concepts, such as ontology-driven hierarchical structures and electronic medical record-driven hierarchical structures, which are crucial for accurate diagnosis prediction. For example, consider a patient with a history of common cold and chronic sinusitis. Language model-based methods may predict the next diagnosis as common cold due to the symptom similarity of common cold and its high frequency of occurrence in the training data. If semantics and structures can be unified, hierarchical relationships can be utilized to improve the accuracy of prediction. In the ontology hierarchy, both allergic rhinitis and chronic sinusitis are classified under other diseases of the upper respiratory tract. Additionally, chronic sinusitis is usually a downstream complication of allergic rhinitis in the electronic medical record hierarchy. If these relationships can be recognized, the model can more accurately predict allergic rhinitis as the next diagnosis. However, it is not easy to effectively model complex structures within the framework of language model-based diagnosis prediction. Standard embedding techniques represent concepts as single points in a vector space. Although these vector embeddings can effectively capture similarity relationships, they cannot encode complex relationships, such as the inclusion relationships inherent in hierarchical structures. Box embedding can represent entities as high-dimensional hyperrectangles to capture complex relationships.

[0004] In view of this, the present invention aims at the problem that existing language model-based diagnosis prediction methods cannot effectively model complex structures in medical concepts, and proposes a medical diagnosis prediction method and system based on box embedding for unifying medical concept structures and semantics. Summary of the Invention

[0005] The objective of the present invention is to propose a medical diagnosis prediction method and system based on box embedding to unify the medical concept structure and semantics, including: modeling a hierarchical structure driven by ontology and medical records, using box embedding to fuse the structure semantics, and integrating the semantic embeddings obtained from a pre-trained language model with the hierarchical structure; performing box modeling on patients based on the information of past medical records, and using the volume of the intersection of the patient box structure and the Clinical Classification Codes (CCS) box structure to quantify the similarity between them, so as to achieve more accurate diagnosis prediction.

[0006] To achieve the above objective, the technical solution of the present invention is as follows:

[0007] The present invention proposes a medical diagnosis prediction method based on box embedding to unify the medical concept structure and semantics, specifically including the following steps:

[0008] S1. Model a hierarchical structure driven by medical ontology; box embedding models medical concepts as high-dimensional hyperrectangles, and the medical concepts include diagnoses and Clinical Classification Codes. The center and offset of each box structure respectively represent the semantic meaning and hierarchical relationship of the corresponding medical concept.

[0009] S2. Model a hierarchical structure driven by patient medical records; based on the modeling of the hierarchical structure driven by medical ontology, represent the patient's visit information as a set of medical concepts in the medical record data, including diagnoses and Clinical Classification Codes, and naturally form a hierarchical structure ;

[0010] S3. Perform box modeling on the patient based on the information of past medical records;

[0011] S4. Use the volume of the intersection of the patient box structure and the Clinical Classification Codes (CCS) box structure to quantify the similarity between the two, and perform diagnosis prediction according to the similarity.

[0012] Preferably, the modeling of the hierarchical structure driven by medical ontology specifically includes the following steps:

[0013] S1.1. Define the box embedding of each box structure as , where respectively represent the center embedding and offset embedding of the box structure corresponding to the i-th medical concept. The center embedding and offset embedding of the box structure both include diagnosis embedding and Clinical Classification Code embedding; R represents a real number, and dim represents the dimension of the embedding;

[0014] S1.2. Calculate the center of the box structure using semantic embeddings based on a pre-trained language model; use the diagnosis name and Clinical Classification Code name as inputs to obtain the diagnosis embedding and the Clinical Classification Code embedding through the pre-trained language model, and use the diagnosis embedding and the Clinical Classification Code embedding Reduce the dimension to obtain the central embedding of the box structure ;

[0015] S1.3. Calculate the offset of the box structure using medical ontology knowledge through a graph convolution mechanism; in the hierarchical structure diagram driven by the medical ontology where respectively represent the set of medical concepts and the set of edges. After each medical concept (such as a specific diagnosis or a clinical classification code) is initialized using the central embedding of the corresponding box structure the offset embedding is calculated using a relation-aware graph convolutional network :

[0016]

[0017]

[0018] where and respectively represent the diagnostic and clinical classification code offset embeddings of the box structure. is an aggregation function that combines information from adjacent concepts. and respectively represent the set of medical concepts related to the diagnosis and the clinical classification code respectively. The function combines adjacent concepts with the relation combined. represents the central embedding of the diagnosis of the nth box structure corresponding to the adjacent concept of the diagnosis ; represents the central embedding of the diagnosis of the sth box structure corresponding to the adjacent concept of the clinical classification code ; is a learnable weight, and r contains directed edges of two types of adjacent concepts (bidirectional relationships between parent and child).

[0019] Preferably, the pre-trained language model uses the Biobert model.

[0020] Preferably, the diagnostic embedding and the clinical classification code embedding are reduced in dimension through a learnable multi-layer perceptron MLP to obtain the central embedding of the box structure :

[0021]

[0022] where represents the multi-layer perceptron, and Respectively represent the diagnostic and clinical classification code center embedding of the box structure.

[0023] Preferably, the aggregation function includes a summation function or an averaging function.

[0024] Preferably, the hierarchical structure driven by the modeled patient medical records specifically includes the following steps:

[0025] S2.1. Embed and represent a certain medical visit record of the patient as , which respectively represent the center embedding of the medical visit record and the offset embedding of the medical visit record;

[0026] S2.2. Calculate the attention score of each medical concept related to the current medical visit record :

[0027]

[0028] where represents a multi-layer perceptron used to model the adjacent concept relationship, represents the center embedding of the i-th medical concept related to the current medical visit record, represents the set of medical concepts related to the current medical visit record, and j is used to traverse , represents the center embedding of the j-th medical concept related to the current medical visit record;

[0029] S2.3. Calculate the center embedding of the medical visit record by weighted aggregation of the center embeddings of each medical concept related to the current medical visit record :

[0030]

[0031] S2.4. Define the offset embedding of the medical visit record as the maximum offset of the related medical concepts:

[0032]

[0033] where represents the offset embedding of each medical concept in.

[0034] Preferably, the box modeling of the patient based on the information of past medical records specifically includes the following steps:

[0035] S3.1. Capture the time interval relationship in the medical visit record sequence , and calculate the time impact of each past medical visit on the current medical visit:

[0036]

[0037] wherein represents a multi-layer perceptron for modeling temporal relationships represents the number of past medical records, and l is used to traverse each time interval in the medical record sequence in the represents the medical record sequence the l-th time interval in represents the medical record the time interval between the current medical record and the previous medical record;

[0038] S3.2. Dynamically update the central embedding of the patient's current visit through the influence of time :

[0039]

[0040] wherein represents each past medical record represents the central embedding of each past medical record;

[0041] S3.3. Retain the hierarchical structure and semantic relationships by summarizing all the information in the past medical records to obtain the offset embedding of the current visit :

[0042]

[0043] wherein represents the offset embedding of each past medical record

[0044] Preferably, the similarity between the patient box structure and the clinical classification code box structure is quantified by using the volume of their intersection, and diagnosis prediction is performed according to the similarity, which specifically includes the following steps:

[0045] S4.1. Define the volume of the intersection of the patient box structure and the clinical classification code box structure :

[0046]

[0047] wherein represents the scaling factor, k represents each dimension of the box embedding represents the Euler constant , , and respectively represent the maximum and minimum values of the box embedding p in the k-th dimension and respectively represent the maximum and minimum values of the box embedding c in the k-th dimension; the box embedding p represents the box embedding of the current visit of patient p; the box embedding c represents the box embedding of the clinical classification code c;

[0048] S4.2. Calculate the similarity score based on the intersection volume of the patient box structure and the clinical classification code box structure:

[0049]

[0050] where represents the patient for the clinical classification code the predicted probability of, using function for standardization to ensure representing valid probabilities for multi-label classification.

[0051] Preferably, optimize the model parameters through the binary cross-entropy loss function :

[0052]

[0053] where represents the number of clinical classification codes, represents the true label, represents the predicted label.

[0054] The present invention proposes a medical diagnosis prediction system based on box embedding to unify the medical concept structure and semantics, including a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically executes the steps in any of the above-mentioned medical diagnosis prediction methods based on box embedding to unify the medical concept structure and semantics.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1. Using box embedding unifies the structure and semantics of medical concepts, solving the problem that the existing diagnosis prediction methods based on language models cannot effectively model the complex structure in medical concepts.

[0057] 2. Using the box structure quantifies the relationship between patients and medical concepts, making the obtained diagnosis prediction more interpretable and improving the performance and credibility of medical diagnosis prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is an example diagram showing the semantics and structure of modeling medical concepts given by the present invention;

[0059] Figure 2 is the overall framework diagram of the method of the present invention;

[0060] Figure 3Schematic diagram of the label-level accuracy results of the diagnosis prediction after training with different proportions of training data on the MIMIC-III dataset for the present invention;

[0061] Figure 4 Schematic diagram of the visit-level accuracy results of the diagnosis prediction after training with different proportions of training data on the MIMIC-III dataset for the present invention;

[0062] Figure 5 Schematic diagram of the label-level accuracy results of the diagnosis prediction after training with different proportions of training data on the MIMIC-IV dataset for the present invention;

[0063] Figure 6 Schematic diagram of the visit-level accuracy results of the diagnosis prediction after training with different proportions of training data on the MIMIC-IV dataset for the present invention;

[0064] Figure 7 Illustrative example diagram of the diagnosis prediction for patient p on the MIMIC-III dataset generated by the present invention and the BoxCare model. Detailed implementation manner

[0065] Next, in combination with the attached Figures 1-7 , the technical solution of the present invention will be specifically described.

[0066] Refer to Figure 2 , the present invention proposes a medical diagnosis prediction method based on box embedding to unify the medical concept structure and semantics, specifically including the following steps:

[0067] S1. Model the ontology-driven hierarchical structure; box embedding models medical concepts as high-dimensional hyperrectangles, and the medical concepts include diagnoses and clinical classification codes. The center and offset of each box structure respectively represent the semantic meaning and hierarchical relationship of the corresponding medical concept;

[0068] S2. Model the patient medical record-driven hierarchical structure; based on the modeling of the ontology-driven hierarchical structure, represent the patient's visit information as a set of medical concepts in the medical record data, including diagnoses and clinical classification codes, and naturally form a hierarchical structure ;

[0069] S3. Perform box modeling on the patient based on the information of past medical records;

[0070] S4. Quantify the similarity between the patient box structure and the clinical classification code box structure using the volume of the intersection part, and perform diagnosis prediction according to the similarity.

[0071] In this embodiment, the modeling of the ontology-driven hierarchical structure specifically includes the following steps:

[0072] S1.1. Define the box embedding of each box structure as , where respectively represent the central embedding and offset embedding of the box structure corresponding to the i-th medical concept. Both the central embedding and offset embedding of the box structure include diagnostic embedding and clinical classification code embedding; R represents a real number, and dim represents the dimension of the embedding;

[0073] S1.2. Use semantic embedding based on a pre-trained language model to calculate the center of the box structure; take the diagnostic name and clinical classification code name as inputs and obtain the diagnostic embedding and clinical classification code embedding through the pre-trained language model. Reduce the dimensions of the diagnostic embedding and clinical classification code embedding to obtain the central embedding of the box structure;

[0074] S1.3. Calculate the offset of the box structure using medical ontology knowledge through a graph convolution mechanism; in the hierarchical structure graph driven by medical ontology, respectively represent the set of medical concepts and the set of edges. After each medical concept is initialized using the corresponding central embedding of the box structure, a relation-aware graph convolutional network is used to calculate the offset embedding :

[0075]

[0076]

[0077] where, and respectively represent the diagnostic and clinical classification code offset embeddings of the box structure. is an aggregation function that combines information from adjacent concepts. and respectively represent the sets of medical concepts related to the diagnosis and the clinical classification code . The function combines adjacent concepts with the relation . represents the diagnostic central embedding of the box structure corresponding to the n-th adjacent concept to the diagnosis . represents the diagnostic central embedding of the box structure corresponding to the s-th adjacent concept to the clinical classification code . is a learnable weight, and r contains directed edges of two types of adjacent concepts.

[0078] In this embodiment, the pre-trained language model uses the Biobert model.

[0079] In this embodiment, the diagnostic embedding and the clinical classification code embedding are reduced in dimension through a learnable multi-layer perceptron MLP to obtain the central embedding of the box structure :

[0080]

[0081] Among them, represents the multi-layer perceptron, and respectively represent the central embeddings of the diagnosis and clinical classification code of the box structure.

[0082] In this embodiment, the aggregation function includes a summation function or an averaging function.

[0083] In this embodiment, the modeling of the patient's medical record-driven hierarchy specifically includes the following steps:

[0084] S2.1. Embed a certain medical visit record of the patient as , respectively represent the central embedding of the medical visit record and the offset embedding of the medical visit record;

[0085] S2.2. Calculate the attention score :

[0086]

[0087] Among them represents the multi-layer perceptron for modeling the relationship between adjacent concepts, represents the central embedding of the i-th medical concept related to the current medical visit record, represents the set of medical concepts related to the current medical visit record, j is used to traverse , represents the central embedding of the j-th medical concept related to the current medical visit record; here, according to the existing relationships in the medical record, that is, the relationship between the patient and the diagnosis / clinical classification code to determine the medical concepts related to the current medical visit record;

[0088] S2.3. Calculate the central embedding of the medical visit record by weighted aggregation of the central embeddings of each medical concept related to the current medical visit record :

[0089]

[0090] S2.4. Embed the offset of the medical record Define it as the maximum offset of relevant medical concepts:

[0091]

[0092] Wherein represents the offset embedding of each medical concept in

[0093] In this embodiment, the box modeling of the patient based on the information in the past medical records specifically includes the following steps:

[0094] S3.1. Capture the time interval relationship in and calculate the time impact of each past medical record on the current medical visit :

[0095]

[0096] Wherein represents the multi-layer perceptron for modeling the time relationship, represents the number of past medical records, and l is used to traverse each time interval in the medical record sequence in represents the medical record sequence the l-th time interval in represents the medical record the time interval between and the previous medical record;

[0097] S3.2. Dynamically update the central embedding of the patient's current medical visit through the time impact :

[0098]

[0099] Wherein represents each past medical record, represents the central embedding of each past medical record;

[0100] S3.3. Retain the hierarchical structure and semantic relationship by summarizing all the information in the past medical records to obtain the offset embedding of the current medical visit :

[0101]

[0102] Wherein represents the offset embedding of each past medical record.

[0103] In this embodiment, the volume of the intersection of the patient box structure and the clinical classification code box structure is used to quantify the similarity between the two, and diagnosis prediction is performed according to the similarity, which specifically includes the following steps:

[0104] S4.1. Define the volume of the intersection of the patient box structure and the clinical classification code box structure :

[0105]

[0106] where represents the scaling coefficient, k represents each dimension of the box embedding, represents the Euler constant, , , and respectively represent the maximum and minimum values of the box embedding p in the k-th dimension, and respectively represent the maximum and minimum values of the box embedding c in the k-th dimension; the box embedding p represents the box embedding of the current visit of patient p; the box embedding c represents the box embedding of the clinical classification code c;

[0107] S4.2. Calculate the similarity score according to the volume of the intersection of the patient box structure and the clinical classification code box structure:

[0108]

[0109] where represents the predicted probability of patient for the clinical classification code , and the function is used for standardization to ensure valid probabilities representing multi-label classification.

[0110] In this embodiment, the model parameters are optimized through the binary cross-entropy loss function :

[0111]

[0112] where represents the number of clinical classification codes, represents the true label, represents the predicted label.

[0113] The present invention proposes a medical diagnosis prediction system based on box embedding to unify the medical concept structure and semantics, including a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically executes the steps in any of the above-mentioned medical diagnosis prediction methods based on box embedding to unify the medical concept structure and semantics.

[0114] Furthermore, to evaluate the performance of the method described in the present invention in medical diagnosis prediction tasks, the performance was verified on the MIMIC-III and MIMIC-IV datasets. Patients with at least two visit records were studied. There were a total of 5,449 patients, 8,692 visit records, 3,874 diagnoses, and 264 clinical classification codes in the MIMIC-III data, and a total of 79,393 patients, 329,597 visit records, 37,917 diagnoses, and 808 clinical classification codes in the MIMIC-IV data. The same experiments were conducted using datasets with different training ratios. Given a patient 's one visit record and the set of medical codes involved to predict the diagnosis label at the patient's next visit, and the label range is the clinical classification codes encoded by the clinical classification software.

[0115] The present invention uses Visit-level@K (visit-level accuracy) and Code-level@K (label-level accuracy) to evaluate the performance of diagnosis prediction. The patient true label and predicted label are respectively represented as and , where and . Visit-level@K first obtains and the smaller value of the total number of labels as the denominator, and then uses the number of correct labels in the top-K predictions as the numerator, and sums them up to measure the prediction performance of the patient's next visit. Code-level@K sums up the number of correct labels in the top-K predictions of all patients as the numerator, and sums up the number of true labels of all patients to obtain the denominator, so as to evaluate the overall accuracy of all patients' predictions.

[0116] The present invention compared 12 state-of-the-art baselines and made comparisons from three main aspects: (1) time-aware methods: RETAIN, StageNet, TRANS; (2) hierarchical structure-aware methods: KAME, CGL, HiTANet, BoxCare; (3) semantic-aware methods: BERT, BERT*, BioBERT, BioBERT*, VecoCare, where * indicates that the patient's past medical records were incorporated during training.

[0117] To prove the effectiveness of the present invention in few-shot diagnostic prediction tasks, Table 1 summarizes the diagnostic prediction results on the MIMIC-III and MIMIC-IV datasets with 5% training data. The present invention outperforms the state-of-the-art baseline methods in all evaluation metrics, demonstrating its advantages in medical diagnostic prediction. Compared with the hierarchical perception model BoxCare, the average improvement in the visit-level accuracy metric is 16.70%, and the average improvement in the label-level accuracy metric is 13.96%; compared with the time perception model Trans, the Visit-level@10 metric on the MIMIC-IV dataset is improved by 46.48%; compared with the semantic perception model VecoCare, the average improvement is 28.76%.

[0118] Table 1: Prediction performance (%) on MIMIC-III and MIMIC-IV datasets with 5% training data

[0119] (The best performance is shown in bold, and the second-best performance is underlined.)

[0120]

[0121] To prove that the present invention remains effective under different training data ratios, experiments are conducted on 1%, 5%, 10%, 15%, 15%, 50% and 100% of the MIMIC-III dataset. As shown in the bar charts of Visit-level@10 in Figure 3 and Code-level@10 in Figure 4 , the present invention outperforms all baselines on all ratios of training data. Although the prediction performance of all models increases with the increase of training data, compared with the present invention, the improvement in baseline performance is not much. For example, as the training data increases from 1% to 5%, the Visit-level@10 of BoxCare as the second-best method increases from 34.78% to 38.21%. However, even with 100% training data, the Visit-level@10 of BoxCare is still lower than the performance of the present invention when using 15% training data. Similarly, experiments are conducted on 1%, 5%, 10% and 15% of the MIMIC-IV dataset. As shown in the bar charts of Visit-level@10 in Figure 5 and Code-level@10 in Figure 6 , as the training data increases from 1% to 5%, the Visit-level@10 of BoxCare as the second-best method increases from 30.60% to 35.13%. However, even with 15% training data, the Visit-level@10 of BoxCare is still lower than the performance of the present invention when using 1% training data.

[0122] In addition, the present invention also provides a real-case study of the MIMIC-III dataset, Figure 7 for the comparison of the prediction results of BoxCare and the present invention for the patient's visit record 3. It is known that the patient was diagnosed with essential hypertension in the first two records. BoxCare can only rely on the ontology-driven hierarchy and vector embedding and cannot capture the co-occurrence relationships existing in the medical record data, resulting in inaccurate diagnostic predictions. In contrast, the present invention reveals the interpretable associations between medical entities by using the box embedding to fuse the structure and semantics and performing diagnostic predictions based on box perception, thus making full use of the ontology and the medical record-driven hierarchy. Therefore, the present invention can judge the possibility of forming coronary atherosclerosis and other heart diseases (with clinical classification code 101) by calculating the intersection volume between the patient and the box structure of the clinical classification code, which is consistent with the clinical knowledge that hypertension increases the workload of the arteries in the heart, thus increasing the risk of atherosclerosis.

[0123] The present invention uses box embedding to unify the structure and semantics of medical concepts, solving the problem that the existing language model-based diagnostic prediction methods cannot effectively model the complex structures in medical concepts; using the box structure to quantify the relationship between the patient and medical concepts makes the obtained diagnostic predictions more interpretable and improves the performance and credibility of medical diagnostic predictions.

[0124] The above are the preferred embodiments of the present invention. All changes made according to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention in terms of the functions and effects produced belong to the protection scope of the present invention.

Claims

1. A medical diagnosis prediction method based on box embedding to unify the structure and semantics of medical concepts, characterized in that, Specifically, it includes the following steps: S1. Model a hierarchical structure driven by a medical ontology; box embedding models medical concepts as high-dimensional hyperrectangles, where the medical concepts include diagnosis and clinical classification codes, and the center and offset of each box structure represent the semantic meaning and hierarchical relationship of the corresponding medical concept respectively; S2. Model the hierarchical structure driven by the patient's medical record; based on the hierarchical structure modeling driven by the medical ontology, represent the patient's visit information as a set of medical concepts in the medical record data, including diagnosis and clinical classification codes, and naturally form a hierarchical structure ; S3. Perform box modeling on the patient based on the information in past medical records; S4. Use the volume of the intersection of the patient box structure and the clinical classification code box structure to quantify the similarity between the two, and make a diagnosis prediction based on the similarity.

2. The medical diagnosis prediction method based on the unified medical concept structure and semantics of box embedding according to claim 1, wherein The modeling of the hierarchical structure driven by the medical ontology specifically includes the following steps: S1.

1. Define the box embedding of each box structure as , where respectively represent the central embedding and the offset embedding of the box structure corresponding to the i-th medical concept. Both the central embedding and the offset embedding of the box structure include a diagnosis embedding and a clinical classification code embedding; R represents a real number, and dim represents the dimension of the embedding; S1.

2. Use the semantic embedding based on the pre-trained language model to calculate the center of the box structure; take the diagnosis name and the clinical classification code name as inputs, and obtain the diagnosis embedding through the pre-trained language model and the clinical classification code embedding , take the diagnosis embedding and the clinical classification code embedding to reduce the dimension to obtain the center embedding of the box structure ; S1.

3. Calculate the offset of the box structure using medical ontology knowledge through the graph convolution mechanism; in the hierarchical structure diagram driven by the medical ontology where respectively represent the set of medical concepts and the set of edges, and each medical concept uses the center embedding of the corresponding box structure After initialization, a relation-aware graph convolutional network is used to calculate the offset embedding : Among them, and respectively represent the diagnostic and clinical classification code offset embeddings of the box structure. is an aggregation function that combines information from adjacent concepts. and respectively represent the medical concept sets related to the diagnosis and the clinical classification code respectively. The function combines adjacent concepts with the relationship combined. represents the diagnostic center embedding of the nth box structure corresponding to the adjacent concept of the diagnosis respectively. represents the diagnostic center embedding of the sth box structure corresponding to the adjacent concept of the clinical classification code respectively. is a learnable weight, and r contains directed edges of two types of adjacent concepts.

3. The medical diagnosis prediction method based on the unified medical concept structure and semantics of box embedding according to claim 2, wherein The pre-trained language model uses the Biobert model.

4. The medical diagnosis prediction method based on box embedding to unify the medical concept structure and semantics according to claim 2, characterized in that Embed the diagnosis and the clinical classification code Reduce the dimension through a learnable multi-layer perceptron MLP to obtain the central embedding of the box structure : Among them, represents a multi-layer perceptron, and respectively represent the diagnostic and clinical classification code center embeddings of the box structure.

5. The medical diagnosis prediction method based on the unified medical concept structure and semantics of box embedding according to claim 2, wherein The aggregation function includes a summation function or an averaging function.

6. The medical diagnosis prediction method based on the unified medical concept structure and semantics embedded in a box according to claim 1, characterized in that The modeling of the hierarchical structure driven by the patient's medical record specifically includes the following steps: S2.

1. Embed a certain medical visit record of the patient as , representing the central embedding and the offset embedding of the medical visit record respectively; S2.

2. Calculate the attention score for each medical concept related to the current medical record : Among them represents a multi-layer perceptron for modeling adjacent concept relationships represents the central embedding of the i-th medical concept related to the current visit record represents the set of medical concepts related to the current visit record, and j is used to iterate through , represents the central embedding of the j-th medical concept related to the current visit record; S2.

3. Calculate the central embedding of the visit record by weighted aggregation of the central embeddings of each medical concept related to the current visit record , and calculate the central embedding of the visit record : S2.

4. Embed the offset of the medical record Define it as the maximum offset of relevant medical concepts: Among them denotes the offset embedding of each medical concept in 7. The medical diagnosis prediction method based on the unified medical concept structure and semantics embedded in a box according to claim 1, characterized in that The performing of box modeling on the patient based on the information in past medical records specifically includes the following steps: S3.

1. Capture the sequence of medical visit records The time interval relationship in it, and calculate the time impact of each previous medical visit record on the current medical visit : Among them represents a multi-layer perceptron for modeling temporal relationships represents the number of past medical records, and l is used to traverse each time interval in the medical record sequence represents the medical record sequence the l-th time interval in represents the medical record and the time interval from the previous medical record; S3.

2. Dynamically update the center embedding of the patient's current visit based on time influence : wherein represents each past medical record, represents the central embedding of each past medical record; S3.

3. Retain the hierarchical and semantic relationships by aggregating all the information in past medical records to obtain the offset embedding for the current visit : Among them represents the offset embedding of each past medical record.

8. The medical diagnosis prediction method based on box embedding to unify the medical concept structure and semantics according to claim 1, characterized in that The using of the volume of the intersection of the patient box structure and the clinical classification code box structure to quantify the similarity between the two, and making a diagnosis prediction based on the similarity specifically includes the following steps: S4.

1. Define the intersecting volume between the patient cassette structure and the clinical classification code cassette structure : where represents the scaling factor, k represents each dimension of the box embedding, represents the Euler's constant, , , and represent the maximum and minimum values of the box embedding p in the k-th dimension respectively, and represent the maximum and minimum values of the box embedding c in the k-th dimension respectively; the box embedding p represents the box embedding of the current visit of patient p; the box embedding c represents the box embedding of the clinical classification code c; S4.

2. Calculate the similarity score according to the volume of the intersection of the patient box structure and the clinical classification code box structure: Among them represents the patient predicted probability for the clinical classification code is normalized using the function to ensure a valid probability representing multi-label classification.

9. The medical diagnosis prediction method based on box embedding to unify the medical concept structure and semantics according to claim 1, characterized in that Through the binary cross-entropy loss function Optimize the model parameters: Among them represents the number of clinical classification codes represents the true label represents the predicted label 10. A medical diagnosis prediction system based on box embedding to unify the structure and semantics of medical concepts, characterized in that, It includes a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically executes the steps in the medical diagnosis prediction method for unifying the medical concept structure and semantics based on box embedding as described in any one of claims 1-9.

Images