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

Abstract

The application provides a medical diagnosis prediction method and system based on box embedding unified medical concept structure and semantics, and the method specifically comprises the following steps: modeling a medical ontology driven hierarchical structure; modeling a patient medical record driven hierarchical structure; box modeling of a patient based on information of past medical records; using a volume of an intersection part of a patient box structure and a clinical classification code box structure to quantify similarity of the two, and making a diagnosis prediction according to the similarity. The application can realize more accurate diagnosis prediction.

Description

Technical Field

[0001] The present invention relates to the medical field, and in particular to a medical diagnosis prediction method and system based on box embedding to unify medical concept structure and semantics. Background Art

[0002] To provide patients with personalized treatment plans and improve their conditions, accurate diagnosis and prediction based on electronic medical records (EMRs) has become a critical task in modern healthcare. However, the widespread use of EMR-based diagnosis and prediction methods raises the concern of potential leakage of patient privacy data. Limited by access to specific patient datasets for few-shot learning, existing technologies are forced to use only limited patient EMR data to maintain predictive performance.

[0003] Existing language model-based methods address the few-shot learning problem by incorporating semantic understanding of medical concepts. After pre-training on a large medical corpus, they introduce medical entity embeddings to improve semantic representation and contextual understanding of medical concepts. However, existing language model-based methods often fail to capture the structure of medical concepts, such as the hierarchical structure driven by medical ontologies and electronic medical records, which are crucial for accurate diagnosis prediction. For example, consider a patient with a history of both the common cold and chronic sinusitis. A language model-based method might predict the common cold as the next diagnosis due to the similarity of common cold symptoms and their high frequency in the training data. However, unifying semantics and structure can leverage hierarchical relationships to improve prediction accuracy. In the ontology hierarchy, allergic rhinitis and chronic sinusitis are both categorized under other diseases of the upper respiratory tract. Furthermore, chronic sinusitis is often a downstream complication of allergic rhinitis in the electronic medical record hierarchy. Recognizing these relationships allows the model to more accurately predict allergic rhinitis as the next diagnosis. However, effectively modeling complex structures within a language model-based diagnostic prediction framework is challenging. Standard embedding techniques represent concepts as single points in a vector space. While these vector embeddings effectively capture similarity relationships, they fail to encode complex relationships, such as the inclusion relationships inherent in hierarchical structures. Box embeddings can capture complex relationships by representing entities as high-dimensional hyperrectangles.

[0004] In view of this, the present invention addresses the problem that existing diagnosis prediction methods based on language models cannot effectively model the complex structure of medical concepts, and proposes a medical diagnosis prediction method and system based on box embedding to unify the medical concept structure and semantics. Summary of the Invention

[0005] The purpose 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 ontology-driven and medical record-driven hierarchical structures, using box embedding to fuse structural semantics, and integrating the semantic embedding obtained from a pre-trained language model with the hierarchical structure; box modeling of patients based on information from past medical records, and using the volume of the intersection of the patient box structure and the clinical classification code (CCS) box structure to quantify the similarity between them, thereby achieving more accurate diagnosis prediction.

[0006] To achieve the above object, 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 medical concept structure and semantics, which specifically includes the following steps:

[0008] S1. Modeling a hierarchical structure driven by medical ontology; box embedding models medical concepts as high-dimensional hyperrectangles. The medical concepts include diagnosis and clinical classification codes. The center and offset of each box structure represent the semantic meaning and hierarchical relationship of the corresponding medical concept.

[0009] S2. Modeling a hierarchical structure driven by patient medical records; Based on the hierarchical structure modeling driven by medical ontology, the patient's medical information is represented as a set of medical concepts in the medical record data, including diagnosis and clinical classification codes, and naturally forms a hierarchical structure ;

[0010] S3, box modeling of patients based on information from past medical records;

[0011] S4. The volume of the intersection of the patient box structure and the clinical classification code (CCS) box structure is used to quantify the similarity between the two, and diagnosis prediction is made based on the similarity.

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

[0013] S1.1. Define the box embedding of each box structure as ,in They represent the center embedding and offset embedding of the box structure corresponding to the i-th medical concept, respectively. 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. Use semantic embedding based on pre-trained language model to calculate the center of the box structure; take the diagnosis name and clinical classification code name as input and obtain the diagnosis embedding through the pre-trained language model and clinical classification code embedding , embedding diagnostics into and clinical classification code embedding Reduce the dimension to get the center embedding of the box structure ;

[0015] S1.3. Calculate the offset of the box structure using medical ontology knowledge through graph convolution mechanism; In the hierarchical structure graph driven by medical ontology middle, Represents the medical concept set and edge set respectively, each medical concept (e.g. specific diagnosis or clinical classification code) using the center embedding of the corresponding box structure After initialization, a relation-aware graph convolutional network is used to calculate the offset embedding :

[0016]

[0017]

[0018] in, and represent the diagnostic and clinical classification code offset embeddings of the box structure, is an aggregation function that combines information from adjacent concepts, and Representation and diagnosis and clinical classification codes Related medical concept set, Function combines adjacent concepts with relations Combined, Indicates the nth diagnosis Diagnostic center embedding of adjacent concepts corresponding to box structures, Indicates the sth clinical classification code Diagnostic center embedding of adjacent concepts corresponding to box structures, are learnable weights, and r contains directed edges between two adjacent concept types (bidirectional relationships between parent and child).

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

[0020] Preferably, the diagnosis is embedded in and clinical classification code embedding The center embedding of the box structure is obtained by reducing the dimension through the learnable multi-layer perceptron MLP :

[0021]

[0022] in, represents a multilayer perceptron, and respectively represent the diagnosis and clinical classification code central embedding of the box structure.

[0023] Preferably, the aggregation function comprises a summation function or an average function.

[0024] Preferably, the modeling patient medical record driven hierarchy comprises the following steps:

[0025] S2.1, embedding a certain visit record of a patient into , respectively represent the visit record central embedding and the visit record offset embedding;

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

[0027]

[0028] wherein represents a multi-layer perceptron for modeling the adjacent concept relationship, 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, j is used for iteration , represents the central embedding of the j-th medical concept related to the current visit record;

[0029] S2.3, calculating the visit record central embedding by weighted aggregation of the central embedding of each medical concept related to the current visit record :

[0030]

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

[0032]

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

[0034] Preferably, the box modeling of the patient based on the past medical record information comprises the following steps:

[0035] S3.1, capturing the time interval relationship in the visit record sequence , calculating the time influence of each past visit record on the current visit :

[0036]

[0037] in represents a multilayer perceptron for modeling temporal relationships, Indicates the number of past medical records, l is used to traverse the medical record sequence For each time interval, Represents a sequence of medical records The lth time interval in Indicates medical records the time interval since the last visit record;

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

[0039]

[0040] in Indicates each past medical record, represents the central embedding of each past medical visit record;

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

[0042]

[0043] in Represents the offset embedding of each past medical visit record.

[0044] Preferably, the method of quantifying the similarity between the patient box structure and the clinical classification code box structure by using the volume of the intersection of the two and making a diagnosis prediction based on the similarity specifically comprises the following steps:

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

[0046]

[0047] in represents the scaling factor, k represents each dimension of the box embedding, represents Euler's constant, , , and denote the maximum and minimum values ​​of the box embedding p in the kth dimension, respectively. and denote the maximum and minimum values ​​of the box embedding c on the kth dimension respectively; the box embedding p denotes the box embedding of the current visit of patient p; the box embedding c denotes the box embedding of the clinical classification code c;

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

[0049]

[0050] in Indicates patient Clinical classification codes The predicted probability of The function is normalized to ensure that it represents valid probabilities for multi-label classification.

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

[0052]

[0053] in 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 a unified medical concept structure and semantics, comprising a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically performs any step in the above-mentioned medical diagnosis prediction method based on box embedding a unified medical concept structure and semantics.

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

[0056] 1. Box embedding is used to unify the structure and semantics of medical concepts, solving the problem that existing language model-based diagnosis prediction methods cannot effectively model the complex structure of medical concepts.

[0057] 2. The box structure is used to quantify the relationship between patients and medical concepts, making the resulting diagnostic predictions more interpretable and improving the performance and credibility of medical diagnostic predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 An example diagram of the semantics and structure of modeling medical concepts given in the present invention;

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

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

[0061] Figure 4 This is a schematic diagram of the accuracy results of the diagnosis prediction of the visit level after training with different proportions of training data on the MIMIC-III dataset;

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

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

[0064] Figure 7 Illustrative example graphs of diagnosis predictions for patient p on the MIMIC-III dataset generated for the present invention and the BoxCare model. DETAILED DESCRIPTION

[0065] The following is combined with Figure 1-7 , the technical solution of the present invention is described in detail.

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

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

[0068] S2. Modeling a hierarchical structure driven by patient medical records; Based on the hierarchical structure modeling driven by medical ontology, the patient's medical information is represented as a set of medical concepts in the medical record data, including diagnosis and clinical classification codes, and naturally forms a hierarchical structure ;

[0069] S3, box modeling of patients based on information from past medical records;

[0070] 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.

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

[0072] S1.1. Define the box embedding of each box structure as ,in They represent the center embedding and offset embedding of the box structure corresponding to the i-th medical concept, respectively. 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;

[0073] S1.2. Use semantic embedding based on pre-trained language model to calculate the center of the box structure; take the diagnosis name and clinical classification code name as input and obtain the diagnosis embedding through the pre-trained language model and clinical classification code embedding , embedding diagnostics into and clinical classification code embedding Reduce the dimension to get the center embedding of the box structure ;

[0074] S1.3. Calculate the offset of the box structure using medical ontology knowledge through graph convolution mechanism; In the hierarchical structure graph driven by medical ontology middle, Represents the medical concept set and edge set respectively, each medical concept Use the center embedding of the corresponding box structure After initialization, a relation-aware graph convolutional network is used to calculate the offset embedding :

[0075]

[0076]

[0077] in, and represent the diagnostic and clinical classification code offset embeddings of the box structure, is an aggregation function that combines information from adjacent concepts, and Respective expression and diagnosis and clinical classification codes Related medical concept set, Function combines adjacent concepts with relations Combined, Indicates the nth diagnosis Diagnostic center embedding of adjacent concepts corresponding to box structures, Indicates the sth clinical classification code Diagnostic center embedding of adjacent concepts corresponding to box structures, are learnable weights, and r contains directed edges between two adjacent concept types.

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

[0079] In this embodiment, the diagnosis embeddings are and clinical classification code embeddings are reduced by a learnable multi-layer perceptron (MLP) to obtain the center embeddings of the box structure :

[0080]

[0081] wherein, denotes a multi-layer perceptron, and denote the diagnosis and clinical classification code center embeddings of the box structure, respectively.

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

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

[0084] S2.1, embedding a certain visit record of a patient into , denote the visit record center embedding and the visit record offset embedding, respectively;

[0085] S2.2, calculating the attention score of each medical concept related to the current visit record :

[0086]

[0087] wherein, denotes a multi-layer perceptron for modeling the adjacent concept relationship, denotes the center embedding of the i-th medical concept related to the current visit record, denotes the set of medical concepts related to the current visit record, and j is used for traversal , denotes the center embedding of the j-th medical concept related to the current visit record; here, the medical concepts related to the current visit record are determined according to the existing relationship in the medical record, i.e., the relationship between the patient and the diagnosis / clinical classification code

[0088] S2.3, calculating the visit record center embedding by weighted aggregation of the center embedding of each medical concept related to the current visit record :

[0089] ​

[0090] S2.4. Embedding the medical record offset Defined as the maximum offset of the related medical concept:

[0091]

[0092] in express The offset embedding of each medical concept in .

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

[0094] S3.1. Capturing medical record sequences The time interval relationship in Calculate the time impact of each past medical record on the current medical record :

[0095]

[0096] in represents a multilayer perceptron for modeling temporal relationships, Indicates the number of past medical records, l is used to traverse the medical record sequence For each time interval, Represents a sequence of medical records The lth time interval in Indicates medical records the time interval since the last visit record;

[0097] S3.2. Dynamically update the center embedding of the patient's current visit through time influence :

[0098]

[0099] in Indicates each past medical record, represents the central embedding of each past medical visit record;

[0100] S3.3. Obtain the offset embedding of the current visit by aggregating all the information in the past medical records to preserve the hierarchical structure and semantic relationships :

[0101]

[0102] in Represents the offset embedding of each past medical visit record.

[0103] In this embodiment, the method of quantifying the similarity between the patient box structure and the clinical classification code box structure by using the volume of the intersection of the two and making a diagnosis prediction based on the similarity specifically includes the following steps:

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

[0105]

[0106] in represents the scaling factor, k represents each dimension of the box embedding, represents Euler's constant, , , and denote the maximum and minimum values ​​of the box embedding p in the kth dimension, respectively. and denote the maximum and minimum values ​​of the box embedding c on the kth dimension respectively; the box embedding p denotes the box embedding of the current visit of patient p; the box embedding c denotes the box embedding of the clinical classification code c;

[0107] S4.2. Calculate the similarity score based on the intersecting volume of the patient box structure and the clinical classification code box structure:

[0108]

[0109] in Indicates patient Clinical classification codes The predicted probability of The function is normalized to ensure that it represents valid probabilities for multi-label classification.

[0110] In this embodiment, the binary cross entropy loss function Optimize model parameters:

[0111]

[0112] in 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 a unified medical concept structure and semantics, comprising a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically performs any step in the above-mentioned medical diagnosis prediction method based on box embedding a unified medical concept structure and semantics.

[0114] Furthermore, in order to evaluate the performance of the method of the present invention in the medical diagnosis prediction task, the performance effect was verified on the MIMIC-III and MIMIC-IV datasets. The study included patients with at least two medical records. There were 5449 patients, 8692 medical records, 3874 diagnoses and 264 clinical classification codes in the MIMIC-III data set, and 79393 patients, 329597 medical records, 37917 diagnoses and 808 clinical classification codes in the MIMIC-IV data set. The same experiments were performed using datasets with different training ratios. Given a patient A medical record and the medical code collection involved To predict the patient's diagnostic label for the next visit, the label range is the clinical classification code 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 diagnostic prediction. The true label and predicted label are represented as and ,in and Visit-level@K gets first and The smaller of the total number of labels is used as the denominator, and the number of correct labels in the top-K predictions is used as the numerator. This sum is used to measure the prediction performance for a patient's next visit. Code-level@K uses the sum of the correct labels in the top-K predictions for all patients as the numerator and the sum of the true labels for all patients as the denominator, thus evaluating the overall accuracy of predictions for all patients.

[0116] This paper compares 12 state-of-the-art baselines from three main perspectives: (1) time-aware methods: RETAIN, StageNet, TRANS; (2) hierarchy-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 demonstrate the effectiveness of our method in few-shot diagnosis prediction tasks, Table 1 summarizes the diagnosis prediction results for the MIMIC-III and MIMIC-IV datasets using 5% training data. Our method outperforms state-of-the-art baseline methods across all evaluation metrics, demonstrating its superiority in medical diagnosis prediction. Compared with the hierarchical perception model BoxCare, our method achieves an average 16.70% improvement in visit-level accuracy and a 13.96% improvement in label-level accuracy. Compared with the temporal perception model Trans, our method achieves a 46.48% improvement in the Visit-level@10 metric on the MIMIC-IV dataset. Compared with the semantic perception model VecoCare, our method achieves an average 28.76% improvement.

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

[0119] (Best performance is in bold, second best performance is underlined)

[0120]

[0121] In order to prove that the present invention can remain effective under different training data ratios, experiments are conducted on 1%, 5%, 10%, 15%, 15%, 50% and 100% of the MIMIC-III dataset. Figure 3 Visit-level@10 and Figure 4 As shown in the Code-level@10 histogram, the present invention outperforms all baselines on all proportions of training data. Although the predictive performance of all models increases with the increase of training data, the improvement in baseline performance compared to the present invention is not much. For example, as the training data increases from 1% to 5%, the Visit-level@10 of BoxCare, the suboptimal method, increases from 34.78% to 38.21%. But even with 100% training data, BoxCare's Visit-level@10 is still lower than the performance of the present invention when using 15% training data. Similarly, experiments were conducted on 1%, 5%, 10%, and 15% of the MIMIC-IV dataset. As shown Figure 5 Visit-level@10 and Figure 6 As shown in the Code-level@10 histogram, as the training data increases from 1% to 5%, BoxCare, the suboptimal method, improves its Visit-level@10 from 30.60% to 35.13%. However, even with 15% training data, BoxCare's Visit-level@10 is still lower than the performance of the present invention using 1% training data.

[0122] In addition, this paper also provides a real case study of the MIMIC-III dataset. Figure 7 BoxCare and this invention for patients The predicted results of the medical record 3 are compared. In the first two records, the patient was diagnosed with essential hypertension. BoxCare can only rely on the ontology-driven hierarchical structure and vector embedding, which cannot capture the co-occurrence relationship in the medical record data, resulting in inaccurate diagnosis prediction. In contrast, the present invention uses box embedding to fuse structural semantics and box perception to make diagnosis predictions to reveal the interpretable associations between medical entities, thereby making full use of the ontology and medical record-driven hierarchical structure. Therefore, the present invention is able to calculate the patient's The probability of developing coronary atherosclerosis and other heart diseases (with clinical classification code 101) was determined by the intersection volume between the structures and the clinical classification code box, which is consistent with clinical knowledge that high blood pressure increases the workload of arteries in the heart and thus increases the risk of atherosclerosis.

[0123] The present invention uses box embedding to unify the structure and semantics of medical concepts, solving the problem that existing language model-based diagnosis prediction methods cannot effectively model the complex structure of medical concepts; the box structure is used to quantify the relationship between patients and medical concepts, making the obtained diagnosis prediction more interpretable and improving the performance and credibility of medical diagnosis prediction.

[0124] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions and effects do not exceed the scope of the technical solution of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A medical diagnosis prediction method based on box embedding to unify medical concept structure and semantics, characterized by: The specific steps include: S1. Modeling a hierarchical structure driven by medical ontology; box embedding models medical concepts as high-dimensional hyperrectangles. The medical concepts include diagnosis and clinical classification codes. The center and offset of each box structure represent the semantic meaning and hierarchical relationship of the corresponding medical concept. S2. Modeling a hierarchical structure driven by patient medical records; Based on the hierarchical structure modeling driven by medical ontology, the patient's medical information is represented as a set of medical concepts in the medical record data, including diagnosis and clinical classification codes, and naturally forms a hierarchical structure G v ; S3, box modeling of patients based on information from past medical records; S4. quantify the similarity between the patient box structure and the clinical classification code box structure using the volume of the intersection between the two, and make a diagnosis prediction based on the similarity; The modeling of the medical ontology-driven hierarchical structure specifically includes the following steps: S1.

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

2. Use semantic embedding based on pre-trained language model to calculate the center of the box structure; take the diagnosis name and clinical classification code name as input and obtain the diagnosis embedding e through the pre-trained language model d and clinical classification codes embedded in e c , embed diagnosis into e d and clinical classification codes embedded in e c Reduce the dimension to get the center embedding of the box structure S1.

3. Calculate the offset of the box structure using medical ontology knowledge through graph convolution mechanism; In the hierarchical graph G driven by medical ontology m =(V m ,E m ), V m 、E m Represent the medical concept set and edge set respectively, each medical concept m∈V m Use the center embedding of the corresponding box structure After initialization, a relation-aware graph convolutional network is used to calculate the offset embedding in, and denote the box-structured diagnosis and clinical classification code offset embeddings, AGG(·) is an aggregation function that merges information from adjacent concepts, N(d) and N(c) denote the sets of medical concepts associated with diagnosis d and clinical classification code c, respectively, and φ(·,W r ) function combines adjacent concepts with relation r, represents the diagnosis center embedding of the box structure corresponding to the nth adjacent concept of diagnosis d, represents the diagnostic center embedding of the box structure corresponding to the sth concept adjacent to the clinical classification code c, W r ∈R dim×dim are learnable weights, r contains directed edges of two adjacent concept types; The modeling of the patient medical record-driven hierarchical structure specifically includes the following steps: S2.

1. Embed a patient's medical record into They represent the center embedding and offset embedding of medical records respectively; S2.

2. Calculate the attention score α for each medical concept related to the current medical record i : Among them, MLP E represents a multilayer perceptron used to model the relationship between adjacent concepts, represents the central embedding of the i-th medical concept related to the current medical record, Represents the medical concept set related to the current medical record, j is used to traverse represents the central embedding of the jth medical concept related to the current medical record; S2.

3. Central embedding of each medical concept related to the current medical record through weighted aggregation Computational medical record center embedding S2.

4. Embedding the medical record offset Defined as the maximum offset of the related medical concept: in express The offset embedding of each medical concept in .

2. The medical diagnosis prediction method based on box embedding unified medical concept structure and semantics according to claim 1 is characterized in that: The pre-trained language model adopts the Biobert model.

3. The medical diagnosis prediction method based on box embedding unified medical concept structure and semantics according to claim 1, characterized in that: Embed diagnostics into e d and clinical classification codes embedded in e c The center embedding of the box structure is obtained by reducing the dimension through the learnable multi-layer perceptron MLP Where MLP(·) represents a multilayer perceptron, and Diagnosis and clinical classification code center embeddings representing the box structure, respectively.

4. The medical diagnosis prediction method based on box embedding unified medical concept structure and semantics according to claim 1, characterized in that: The aggregation function includes a sum function or an average function.

5. The medical diagnosis prediction method based on box embedding unified medical concept structure and semantics according to claim 1, characterized in that: The box modeling of the patient based on the information of the past medical records specifically includes the following steps: S3.

1. Capture medical record sequence H p The time interval relationship ΔT in l Calculate the time impact of each past medical record on the current medical record Among them, MLP T represents a multilayer perceptron used to model temporal relationships, |H p | represents the number of past medical records, l is used to traverse the medical record sequence H p Each time interval in ΔT l Represents the medical record sequence H p The lth time interval in ΔT q Indicates medical record v q the time interval since the last visit record; S3.

2. Dynamically update the center embedding of the patient's current visit through time influence where v q Indicates each past medical record, represents the central embedding of each past medical visit record; S3.

3. Obtain the offset embedding of the current visit by aggregating all the information in the past medical records to preserve the hierarchical structure and semantic relationships in Represents the offset embedding of each past medical visit record.

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

1. Define the intersection volume Vol(b p∩c ): where β represents the scaling factor, k represents each dimension of the box embedding, and γ represents Euler's constant, and denote the maximum and minimum values ​​of the box embedding p in the kth dimension, respectively. and denote the maximum and minimum values ​​of the box embedding c on the kth dimension respectively; the box embedding p denotes the box embedding of the current visit of patient p; the box embedding c denotes the box embedding of the clinical classification code c; S4.

2. Calculate the similarity score based on the intersecting volume of the patient box structure and the clinical classification code box structure: in represents the predicted probability of patient p for clinical classification code c, which is normalized using the Softmax function to ensure that it represents the effective probability of multi-label classification.

7. The medical diagnosis prediction method based on box embedding unified medical concept structure and semantics according to claim 1, characterized in that: The model parameters are optimized through the binary cross entropy loss function L: where |M c | represents the number of clinical classification codes, y p,c represents the true label, Represents the predicted label.

8. A medical diagnosis prediction system based on box embedding to unify medical concept structure and semantics, characterized by: The invention comprises a processor, a memory and a computer program stored in the memory. When the processor executes the computer program, the method specifically performs the steps of the medical diagnosis prediction method based on box embedding unified medical concept structure and semantics as described in any one of claims 1 to 7.

Images