Readmission risk prediction method based on trajectory learning long-short memory network

Abstract

The invention belongs to the field of medical health, and provides a re-admission risk prediction method based on a trajectory learning long-short memory network, which is used for improving the accuracy and reliability of re-admission risk prediction of patients. The method comprises the following steps: firstly, acquiring admission record information of a patient, including demographic information, medical coding information and clinical text information; then, a trajectory learning long-short-term memory network unit is adopted to extract characterization values of multi-source data features, a multi-source data feature fusion network is adopted to perform feature fusion, and a re-admission prediction value is obtained through a re-admission prediction network; and finally, obtaining a complete re-admission risk prediction model through loss optimization training, and realizing re-admission risk prediction by the model. According to the invention, the re-admission risk prediction model based on the trajectory learning long-short term memory network is creatively provided, the multi-source information of previous admission treatment of the patient is fused, the health condition description of the patient is comprehensively and dynamically described, and the accuracy and reliability of the re-admission prediction of the patient are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical health, relates to the technology of readmission risk prediction, and specifically provides a readmission risk prediction method based on a trajectory learning long short-term memory network. Background Art

[0002] Readmission is an important issue in the field of medical health. Multiple factors such as doctors failing to prescribe necessary medications for discharged patients, patients failing to follow treatment instructions, and the progression of the disease may all lead to the need for patients to be readmitted; readmission prediction is to identify the potential risks of patients being readmitted after discharge and is also an important indicator for evaluating the quality of hospital care; efficiently and accurately predicting the readmission risk of patients can not only prevent inappropriate discharges and potential dangers of patients, but also reduce related medical costs and relieve the economic burden on patients' families.

[0003] In traditional methods, a scoring system based on domain knowledge is usually adopted to estimate the impact of clinical outcomes on readmission risk prediction; for example, a patent document with the publication number CN112002425A discloses a data-driven readmission risk and admission time risk assessment model for classifying and predicting whether a case is an "avoidable admission" or an "unavoidable readmission". This model is trained using traditional machine learning algorithms (logistic regression and decision tree) to establish a classification prediction model. With the popularization of hospital informatization technology and the development of artificial intelligence technology, by using electronic health record (EHR) data, such as patients' demographic information, medical codes, treatment and clinical records, etc., a readmission risk prediction model can be constructed; for example, a patent document with the publication number CN113223711A discloses a readmission prediction model based on multi-modal data. The structured data and unstructured data are preprocessed to obtain the text-level embedding vector of the unstructured data and the structured data vector. The text-level embedding vector and the structured data vector are fused through a data fusion module to obtain the final representation data of the patient, which is input into the Adaboost-LR model for classification prediction to obtain the prediction result.

[0004] However, data-driven readmission prediction models are usually considered black-box models, that is, the prediction results are difficult to interpret or understand, which may lead to a decrease in the trust of doctors and patients in the model; the prediction method based on electronic health record data usually relies on using time-series data from a single data source and ignores the impact of heterogeneous information of multi-source data characteristics of patients on the readmission risk prediction result; moreover, the development of diseases is often affected by historical conditions, and existing readmission prediction models often only consider static patient information and ignore the disease course development trajectory of patients, making it difficult to capture the temporal relationship between historical medical information and affecting the accuracy of the prediction result. Summary of the Invention

[0005] The object of the present invention is to provide a re - admission risk prediction method based on a trajectory - learning long short - term memory network, so as to improve the accuracy and reliability of patient re - admission risk prediction. The present invention creatively proposes a re - admission risk prediction model Tra - LSTM based on a trajectory - learning long short - term memory network, which integrates multi - source information of a patient's previous hospital admissions, comprehensively and dynamically describes the patient's health status, and significantly improves the accuracy and reliability of patient re - admission prediction.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A re - admission risk prediction method based on a trajectory - learning long short - term memory network, characterized by comprising the following steps:

[0008] Step 1. Obtain the patient's admission record information and form a training data set through data pre - processing; the admission record information includes: demographic information, medical coding information, and clinical text information;

[0009] Step 2. Use a trajectory - learning long short - term memory network unit to extract multi - source data features and obtain the representation value of each feature data;

[0010] Step 3. Use a multi - source data feature fusion network to perform feature fusion on the representation values of multi - source data and obtain a multi - source data feature representation;

[0011] Step 4. According to the multi - source data feature representation, obtain a re - admission prediction value through a re - admission prediction network;

[0012] Step 5. Calculate the loss, and use the stochastic gradient descent method to complete the training of the trajectory - learning long short - term memory network, the multi - source data feature fusion network, and the re - admission prediction network to obtain a complete re - admission risk prediction model;

[0013] Step 6. Send the admission data of the patient to be predicted into the re - admission risk prediction model, and the model outputs a re - admission prediction value.

[0014] Further, in step 1, the data pre - processing process is as follows:

[0015] First, take the demographic information, medical coding information, and clinical text information as feature data. For outliers and missing values in numerical feature data, use the method of filling with the average value; for outliers and missing values in non - numerical feature data, use the value with the highest frequency of occurrence in the patient's data to fill the outliers and missing values;

[0016] Then, represent the patient's admission record as an admission sequence arranged in chronological order (m1,…,m N), where N is the total number of hospital admissions of the patient; the admission data m for each admission t Specifically expressed as (C t , A t , H t , W t ), t = 1, 2,... N, C t is the medical coding record of the current admission, A t is the medical coding record of the previous historical admission before the current admission. When t = 1, it is denoted as H t is the demographic information of the current admission, W t is the clinical text record of the current admission.

[0017] Furthermore, in step 2, the Trajectory Learning Long Short-Term Memory Network (Tra-LSTM) includes: an input gate, a forget gate, an output gate, and a memory unit, where:

[0018] The input gate is expressed as:

[0019]

[0020] The forget gate is expressed as:

[0021]

[0022] The output gate is expressed as:

[0023]

[0024] The memory unit is expressed as:

[0025] q t = i t × tanh(W u mm t + U u d t-1 + b u ) + f t × q t-1

[0026] The hidden state is expressed as:

[0027] d t = o t × tanh(q t )

[0028] where σ(·) is the sigmoid activation function, mm t is the feature data of the current admission; W i , U i , b i , W f , Uf , b f , W o , U o , b o , W u , U u and b u are all learnable weight parameters, and their value ranges are from 0 to 1; × represents the multiplication operation, tanh represents the hyperbolic tangent function, d t-1 represents the hidden state vector of the previous admission, q t-1 represents the memory cell state vector of the previous admission; Δ t is the time interval between the current admission and the previous admission, e is the natural logarithm; 1(readmitted) represents the indicator function, and when the current admission is a readmission, the indicator value is 1, otherwise it is 0; w is the weight factor;

[0029] Transfer the hidden state vector d t learned by the trajectory to the MLP to obtain the representation value mm t of mm′ t :

[0030] mm′ t = ReLU(Linear(d t ))

[0031] where d t is the hidden state vector learned by the trajectory, Linear represents the linear function, and ReLU represents the activation function.

[0032] Furthermore, in step 3, the multi-source data feature fusion network includes: a cross-data feature Transformer model (CA), a self-attention Transformer model (SA), and a multi-source data feature fusion model; where:

[0033] Send the representation value of the feature data into the time non-linear mapper to obtain the corresponding result

[0034]

[0035] where mm′ t is the representation value of the feature data mm t in the multi-source data; Conv1D represents a one-dimensional convolutional neural network, is the result representation of the feature data mm t after passing through the time non-linear mapper;

[0036] For any two types of feature data mma and mmb, the cross-data feature attention between them is expressed as:

[0037]

[0038] Among them, represents the cross-data feature representation from feature data mma to feature data mmb, represents the cross-data feature representation from feature data mmb to feature data mma; Softmax represents the Softmax function, which is used to calculate the attention weights; Q mma , K mma , V mma respectively represent the query, key, and value of feature data mma, Q mmb , K mmb , V mmb respectively represent the query, key, and value of feature data mmb, T represents matrix transpose, and d k represents the dimension of the trainable weight;

[0039] Set mma to medical coding information (C t ), set mmb to demographic information (H t ) or clinical text information (W t ), and obtain and represents the cross-data feature representation from medical coding information to demographic information, represents the cross-data feature representation from medical coding information to clinical text information;

[0040] The self-attention Transformer module is specifically represented as;

[0041]

[0042] Among them, represents the self-attention feature representation of medical coding information, represents the self-attention feature representation of the cross-data feature representation , represents the self-attention feature representation of the cross-data feature representation ; Q C , K C , and V C respectively represent the query, key, and value of the medical coding information feature, Q C→H , K C→H , V C→H respectively represent the query, key, and value of the cross-data feature , Q C→W , K C→W , V C→W respectively represent the query, key, and value of the cross-data feature ;

[0043] The multi-source data feature fusion model is expressed as:

[0044]

[0045] where α is a scaling factor, and MO is the multi-source data feature representation; denotes the concatenation operation, and Linear denotes a linear function, is the result representation of the medical coding information through the time non-linear mapper;

[0046] The multi-source data feature representation MS is obtained by successively performing layer normalization operation and dropout operation on the multi-source data feature representation MO.

[0047] Furthermore, in step 4, the readmission prediction network is expressed as:

[0048]

[0049] where σ denotes the Sigmoid function, and Linear denotes a linear function, denotes the readmission prediction value, and MS is the multi-source data feature representation.

[0050] Furthermore, in step 5, the loss function L total is expressed as:

[0051] L total = λ ce L ce + λ bi-con L bi-con + λ hrchy L hrchy

[0052] where L ce denotes the readmission prediction network loss term, L bi-con denotes the multi-source data feature fusion network loss term, L hrchy denotes the trajectory learning long short-term memory network loss term, and λ ce 、λ bi-con and λ hrchy are all weight parameters;

[0053] y represents the ground truth of the patient's readmission situation, denotes the readmission prediction value;

[0054] denotes the bimodal contrast loss between the medical coding information and the demographic information, denotes the bimodal contrast loss between the medical coding information and the clinical text information; Represents the contrastive loss from medical coding information to demographic information, Represents the contrastive loss from demographic information to medical coding information, and β represents the weighting parameter;

[0055] respectively represent the re - admission prediction stratification losses of medical coding information, demographic information, and clinical text information; for the feature data m in multi - source data t the cross - entropy loss is calculated y mm represents the ground truth of the re - admission situation corresponding to the feature data m t represents the re - admission prediction result of the feature data m t

[0056] Based on the above technical solution, the beneficial effects of the present invention are as follows:

[0057] The present invention provides a re - admission risk prediction method based on a trajectory - learning long - short - term memory network. The long - short - term memory network based on trajectory learning (Tra - LSTM) is used to weight the continuous admission time intervals, significantly enhancing the representation ability of historical admission information; at the same time, by combining the cross - data feature Transformer and the self - attention Transformer, a multi - source data feature fusion network with medical coding as the core is constructed, greatly enhancing the influence of medical coding feature representation, enabling the model to consider the patient's state more comprehensively and meticulously; in addition, a bimodal contrastive loss function is innovatively designed to further optimize the accuracy of re - admission risk prediction. Finally, the present invention has the advantage of significantly improving the performance of the re - admission risk prediction model, providing more accurate and reliable decision - making support for clinical medical practice, and is expected to effectively reduce medical costs and improve the treatment effect and quality of life of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a schematic flowchart of the re - admission risk prediction method based on the trajectory - learning long - short - term memory network provided by the present invention.

[0059] Figure 2 is a network model framework diagram of the re - admission risk prediction method based on the trajectory - learning long - short - term memory network provided by the present invention.

[0060] Figure 3 is a network structure diagram of the trajectory - learning long - short - term memory network (Tra - LSTM) unit provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS ​​

[0061] To make the objectives, technical solutions and beneficial effects of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0062] This embodiment provides a readmission risk prediction method based on a trajectory learning long short-term memory network, and its process is as Figure 1 shown, and the network model framework is as Figure 2 shown, and specifically includes the following steps:

[0063] Step 1. Data preprocessing;

[0064] Extract the patient-related demographic information (such as age, gender, ethnicity, etc.), medical codes (such as ICD-9 codes), and clinical text information (such as medical records, inspection reports, etc.) from the patient electronic health record database of a specialized hospital as feature data. For numerical data, perform preprocessing, and fill in outliers and missing values by using the average value; for non-numerical data, according to the mode principle in statistics, use the value with the highest frequency of occurrence in the patient data to fill in outliers and missing values;

[0065] Represent the patient's admission records as an admission sequence arranged in chronological order (m1,…,m N ), where N is the total number of admissions of patient P; the admission data m t for each admission is specifically represented as (C t ,A t ,H t ,W t ), t = 1, 2,... N, C t is the medical code record of the current admission, A t is the medical code record of the historical admission before the current admission. When t = 1, it is recorded as H t is the demographic information of the current admission, and W t is the clinical text record of the current admission; more specifically, the medical code information of the t-th admission is represented as C t ={c t；1 ,c t；2 ,…,c t；|C|}, |C| represents the number of medical codes, the medical code information record of the historical admission is represented as A t ={a t；1 ,a t；2 ,…,a t；|A|}, |A| represents the number of medical codes of the historical admission, and the demographic information is represented as H t ={h t；1 ,h t；2 ,…,h t；|H|, |H| is the number of demographic characteristics, and the clinical text information is represented as W t ={w t；1 , w t；2 ,…, w t；|W|}, |W| is the maximum number of words in the clinical text information;

[0066] Step 2. Use the trajectory learning long short-term memory network unit to extract the multi-source data features and obtain the representation value of each feature data;

[0067] In the admission sequence, the time interval between two consecutive admissions may vary from a few days to several months. This changing time interval indicates certain upcoming disease conditions; for example, frequent admissions indicate serious health problems and poor health status. The records of these frequent patterns provide a source for studying the progression of the disease. Therefore, in determining the readmission risk, recent admission experiences should have a more significant impact than past admission experiences;

[0068] On this basis, the present invention proposes a trajectory learning long short-term memory network (Tra-LSTM), as Figure 3 shown. This unit weights the time interval between two admissions and enhances the information of previous readmission records at the same time; the Tra-LSTM includes: an input gate, a forget gate, an output gate, and a memory unit with a hidden state. The three gates inherit the information of historical admission records from the hidden state and receive the current admission record at the same time, as follows:

[0069] Input gate:

[0070]

[0071] Forget gate:

[0072]

[0073] Output gate:

[0074]

[0075] Memory unit:

[0076] q t = i t ×tanh(W u mm t + U u d t-1 + b u ) + f t × q t-1 (4)

[0077] Hidden state:

[0078] dt = o t × tanh(q t ) (5)

[0079] Wherein, σ(·) is the sigmoid activation function, mm t is the feature data of the current hospitalization, that is, mm t is a certain feature in the multi-source data at the t-th hospitalization, mm t ∈ (C, H, W); W i , U i , b i , W f , U f , b f , W o , U o , b o , W u , U u and b u are all learnable weight parameters, and the value range is 0 to 1; × represents the multiplication operation, tanh represents the hyperbolic tangent function, q t-1 represents the memory cell state vector of the previous hospitalization, d t-1 represents the hidden state vector of the previous hospitalization;

[0080] The present invention introduces as the attenuation coefficient, and introduces log(e + w × 1(readmitted)) as the amplification coefficient; Δ t is the time interval between the current hospitalization and the previous hospitalization, e is the natural logarithm; 1(readmitted) represents the indicator function, and the indicator value is 1 when the current hospitalization is a readmission, otherwise it is 0; w is the weight factor, which is used to represent the influence degree of the previous readmission on the future readmission risk;

[0081] The hidden state vector d t obtained by trajectory learning is passed to the MLP to obtain the representation value mm t ' of the feature data mm t :

[0082] mm' t = ReLU(Linear(d t )) (6)

[0083] Wherein, d t is the hidden state vector of trajectory learning, Linear represents the linear function and is used to represent the operation of a fully connected layer, and ReLU represents the activation function;

[0084] Step 3. Use the multi-source data feature fusion network to perform feature fusion on the multi-source data to obtain the multi-source data feature representation MS;

[0085] The multi-source data feature fusion network includes: a cross-data feature Transformer model (CA), a self-attention Transformer model (SA), and a multi-source data feature fusion model centered on medical coding; the cross-data feature Transformer can fuse demographic information, medical coding, and clinical text information together, which is beneficial to mining the associations and connections between them, so as to better solve the readmission prediction problem; the self-attention Transformer model (SA) module can effectively extract multi-source data feature representations and enhance the dependence between the two cross-data feature representations and medical code representations; since different data types have different characteristics, using different Transformer models can better adapt to their respective characteristics and improve the performance and effect of the readmission prediction model; specifically:

[0086] 1) Cross-data feature Transformer model (CA);

[0087] Use two cross-data feature Transformer models to fuse demographic information, medical coding, and clinical text information together, and put the representation values of the feature data into a time non-linear mapper (one-dimensional convolutional neural network, 1DCNN) to obtain the corresponding results

[0088]

[0089] Among them, mm′ t is the representation value of the feature data mm in the multi-source data t mm t ∈(C, H, W); Conv1D represents a one-dimensional convolutional neural network, is the result representation of the feature data mm t after passing through the time non-linear mapper;

[0090] For any two types of feature data mma and mmb, the query, key, and value are respectively defined as and W (·) represents the trainable weight with a dimension of d k ; then the cross-data feature representation between the feature data mma and the feature data mmb is:

[0091]

[0092] Among them, represents the cross-data feature representation from the feature data mma to the feature data mmb, Represents the cross - data feature representation from feature data mmb to feature data mma, and are the result representations of feature data mma and feature data mmb passing through the time non - linear mapper respectively; Softmax represents the Softmax function, which is used to calculate the attention weights; Q mma 、K mma 、V mma represent the query, key, and value of feature data mma respectively, Q mmb 、K mmb 、V mmb represent the query, key, and value of feature data mmb respectively, T represents matrix transpose, d k represents the dimension of the trainable weight;

[0093] The fusion is carried out in a medical - coding - centered manner, setting mma as medical - coding information (C), and mmb as demographic information (H) or clinical text information (W); Therefore, two cross - data feature Transformer models are used to obtain and represents the cross - data feature representation from medical - coding information to demographic information, represents the cross - data feature representation from medical - coding information to clinical text information;

[0094] 2) Self - attention Transformer model (SA);

[0095] In order to effectively extract medical - coding information and the two cross - data feature representations and enhance the dependence of the above - mentioned two cross - data feature representations and medical - coding information, the present invention adopts a self - attention Transformer module, specifically represented as;

[0096]

[0097] Among them, represents the self - attention feature representation of medical - coding information C, represents the self - attention feature representation of the cross - data feature , represents the self - attention feature representation of the cross - data feature ; Q C 、K C 、V C represent the query, key, and value of medical - coding information feature C respectively, Q C→H 、K C→H 、V C→H represent the query, key, and value of the cross - data feature respectively, Q C→W 、K C→W, V C→W respectively represent the query, key, and value across data features ;

[0098] 3) A multi-source data feature fusion module centered on medical coding;

[0099] Centered on medical coding, fuse the self-attention feature representations of demographic information, medical coding information, and clinical text information; calculate the feature fusion vector g and displacement vector G of the three types of feature representations through concatenation:

[0100]

[0101] where represents the concatenation operation, Linear represents a linear function, used to represent the operation of a fully connected layer;

[0102] Through the above multi-source data feature fusion, the influence of medical coding representation is maximized; then, the medical coding information feature representation is weighted and summed with the displacement vector G to obtain the multi-source data feature representation MO:

[0103]

[0104] where α represents a scaling factor used to adjust the influence of the displacement vector, ρ is a trainable parameter initialized randomly, ‖‖2 represents the L2 norm, represents the result representation of medical coding information through the temporal non-linear mapper, Finally, the multi-source data feature representation MO undergoes layer normalization operation and dropout operation in sequence to obtain the multi-source data feature representation MS;

[0105] Step 4. According to the multi-source data feature representation MS, obtain the readmission prediction value through the readmission prediction network;

[0106] To predict the readmission situation of patients, input the multi-source data feature representation MS into a single linear layer with a Sigmoid activation function to calculate the readmission prediction value Specifically expressed as:

[0107]

[0108] where σ represents the Sigmoid function, and Linear represents the linear function;

[0109] Step 5. Calculate the loss, and use the stochastic gradient descent method to train the trajectory learning long short-term memory network, multi-source data feature fusion network, and readmission prediction network to obtain the readmission risk prediction model;

[0110] 1) For the multi-source data feature fusion network, a bimodal contrast loss is adopted;

[0111] Taking medical coding information and demographic information as an example, calculate the bimodal contrast loss under these two types of feature data; in the current admission, the results of the medical coding information and demographic information after passing through the time non-linear mapper for the i-th batch are expressed as NP is the total number of patients in a batch training process. The bimodal contrast loss contains two asymmetric losses, which are respectively expressed as:

[0112]

[0113] Among them, represents the contrast loss from medical coding information to demographic information, represents the contrast loss from demographic information to medical coding information; and represent the results of the medical coding information and demographic information after passing through the time non-linear mapper; exp represents the exponential function, <,> represents the cosine similarity, τ represents the gradient of the Softmax function, and τ ∈ R + ; i and r represent the batch numbers in the model training and learning process, and R represents the total number of batches in the model training and learning process;

[0114] Using the weighted parameter β ∈ [0, 1], and are weighted and combined to obtain the bimodal contrast loss of medical coding information and demographic information

[0115]

[0116] Similarly, calculate the bimodal contrast loss of medical coding information and clinical text information

[0117] Thus, the bimodal contrast loss of the multi-source data feature fusion network for clinical text can be obtained:

[0118]

[0119] 2) For the trajectory learning long short-term memory network, a cross-entropy loss is adopted;

[0120] For the feature data m in the multi-source data, where m ∈ (C, H, W), calculate the cross-entropy loss

[0121]

[0122] Among them, r represents the batch in the model training and learning process, and R represents the total number of batches in the model training and learning process; y mm represents the ground truth of the readmission situation corresponding to the feature data mm, and mm′ represents the representation value obtained by the feature data mm through the non-linear mapper MLP. represents the readmission prediction result of the feature data mm;

[0123] From this, the readmission prediction stratification losses of medical coding information, demographic information, and clinical text information are calculated The sum of the three is used as the cross-entropy loss L of the trajectory learning long short-term memory network hrchy :

[0124]

[0125] 3) For the readmission prediction network, the cross-entropy loss is adopted;

[0126] The cross-entropy loss L of the readmission prediction network ce is expressed as:

[0127]

[0128] Among them, y represents the ground truth of the patient's readmission situation, represents the readmission prediction value;

[0129] Finally, the loss function L total is the weighted sum of the above three loss terms, specifically expressed as:

[0130] L total = λ ce L ce + λ bi-con L bi-con + λ hrchy L hrchy (26)

[0131] Among them, λ ce 、λ bi-con and λ hrchy are all weight parameters;

[0132] According to the calculated loss value, the model parameters are updated using the stochastic gradient descent (SGD) method to complete the training;

[0133] Step 6. Send the admission data of the patient to be predicted into the readmission risk prediction model, and the model outputs the readmission prediction value.

[0134] The above are only specific embodiments of the present invention. Any feature disclosed in this specification, unless specifically stated, can be replaced by other equivalent or alternative features with similar purposes; all the disclosed features, or all the steps in any method or process, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A method for predicting hospital readmission risk based on trajectory learning long short-term memory network, characterized in that: The following steps are involved: Step 1. Obtain the patient's admission record information and form a training data set through data preprocessing; The admission record information includes: demographic information, medical coding information and clinical text information; Step 2. Use trajectory learning long short-term memory network units to extract multi-source data features and obtain the representation value of each feature data; Step 3. Use a multi-source data feature fusion network to perform feature fusion on the representation values ​​of multi-source data to obtain multi-source data feature representation; Step 4. Obtain the readmission prediction value through the readmission prediction network according to the multi-source data feature representation; Step 5. Calculate the loss and use the stochastic gradient descent method to complete the training of the trajectory learning long short-term memory network, the multi-source data feature fusion network and the readmission prediction network to obtain a complete readmission risk prediction model; Step 6. Send the admission data of the patient to be predicted into the readmission risk prediction model, and the model outputs the readmission prediction value.

2. The method for predicting hospital readmission risk based on trajectory learning long short-term memory network according to claim 1, characterized in that: In step 1, the data preprocessing process is: First, demographic information, medical coding information and clinical text information are used as feature data. For abnormal values ​​and missing values ​​in numerical feature data, the average value is used to fill them. For abnormal values ​​and missing values ​​in non-numerical feature data, the most frequent value in the patient data is used to fill them. Then, the patient's admission record is represented as a chronological sequence of admissions (m1,…,m N ), where N is the total number of hospitalizations of the patient; the admission data for each hospitalization is m t Specifically expressed as (C t ,A t ,H t ,W t ), t=1,2,...N,C t For the medical coding record of the current admission, A t is the medical coding record of the previous hospitalization before the current hospitalization. When t = 1, it is recorded as H t is the demographic information of the current admission, W t A clinical text record of the current admission.

3. The method for predicting hospital readmission risk based on trajectory learning long short-term memory network according to claim 1, characterized in that: In step 2, the Trajectory Learning Long Short Term Memory Network (Tra-LSTM) includes: an input gate, a forget gate, an output gate, and a memory unit, where: The input gate is represented as: The forget gate is expressed as: The output gate is represented as: The memory unit is represented as: q t =i t ×tanh(W u mm t +U u d t-1 +b u )+f t ×q t-1 The hidden state is represented as: d t =o t ×tanh(q t ) Among them, σ(·) is the sigmoid activation function, mm t is the characteristic data of the current admission; W i , U i 、b i , W f , U f 、b f , W o , U o 、b o , W u , U u With b u are all learnable weight parameters, ranging from 0 to 1; × represents the product operation, tanh represents the hyperbolic tangent function, and d t-1 represents the hidden state vector of the last admission, q t-1 The memory cell state vector representing the last hospitalization; Δ t is the time interval between the current hospitalization and the previous hospitalization, e is the natural logarithm; 1 (readmitted) represents the indicator function, when the current hospitalization is a readmission, the indicator value is 1, otherwise it is 0; w is the weight factor; The hidden state vector d of the trajectory learning t Pass it to MLP to get the feature data mm t Characteristic value mm ′ t : mm ′ t =ReLU(Linear(d t )) Among them, d t The hidden state vector is learned for the trajectory, Linear represents the linear function, and ReLU represents the activation function.

4. The method for predicting hospital readmission risk based on trajectory learning long short-term memory network according to claim 1, characterized in that: In step 3, the multi-source data feature fusion network includes: a cross-data feature Transformer model (CA), a self-attention Transformer model (SA) and a multi-source data feature fusion model; wherein: Send the representation value of the feature data to the time nonlinear mapper to obtain the corresponding result Among them, mm t t is the feature data mm in multi-source data t The representation value of; Conv1D represents a one-dimensional convolutional neural network, is the characteristic data mm t The result representation of the time-lapse nonlinear mapper; For any two types of feature data mma and mmb, the cross-data feature between them is expressed as: in, Represents the cross-data feature representation from feature data mma to feature data mmb, represents the cross-data feature representation from feature data mmb to feature data mma, Softmax represents the Softmax function, which is used to calculate the attention weight; Q mma , K mma 、V mma Respectively represent the query, key, and value of the feature data mma, Q mmb , K mmb 、V mmb They represent the query, key, and value of the feature data mmb respectively, T represents the matrix transpose, and d k Represents the dimension of trainable weights; Set mma to medical coding information (C t ), mmb is set to demographic information (H t ) or clinical text information (W t ), obtained through two cross-data feature Transformer models and Cross-data feature representation from medical coding information to demographic information, Cross-data feature representation from medical coding information to clinical text information; The self-attention Transformer module is specifically represented as follows: in, Self-attention feature representation for medical coding information, Representation of cross-data features The self-attention feature representation is Representation of cross-data features Self-attention feature representation; Q C , K C and V C represents the query, key and value of the medical coding information feature, respectively, C→H , K C→H 、V C→H Represent cross-data features The query, key and value of Q C→W , K C→W 、V C→W Represent cross-data features Query, keys and values ​​of ; The multi-source data feature fusion model is expressed as: Among them, α is the scaling factor, MO is the multi-source data feature representation; ⊕ represents the connection operation, and Linear represents the linear function. The result representation of medical coding information through a temporal nonlinear mapper; The multi-source data feature representation MO is subjected to layer normalization and dropout operations in sequence to obtain the multi-source data feature representation MS.

5. The method for predicting hospital readmission risk based on trajectory learning long short-term memory network according to claim 1, characterized in that: In step 4, the readmission prediction network is expressed as: Among them, σ represents the Sigmoid function, Linear represents the linear function, represents the prediction value of readmission, and MS represents the multi-source data feature.

6. The method for predicting hospital readmission risk based on trajectory learning long short-term memory network according to claim 1, characterized in that: In step 5, the loss function L total It is expressed as: L total =λ ce L ce +λ bi-con L bi-con +λ hrchy L hrchy Among them, L ce represents the readmission prediction network loss term, L bi-con represents the multi-source data feature fusion network loss term, L hrchy represents the loss term of the LSTM network for trajectory learning, λ ce , bi-con With λ hrchy They are all weight parameters; y represents the true value of the patient's readmission. represents the predictive value of readmission; represents the bimodal contrast loss between medical coding information and demographic information, A bimodal contrast loss representing medical coding information and clinical text information; represents the contrast loss from medical coding information to demographic information, represents the contrastive loss from demographic information to medical coding information, and β represents the weighting parameter; Respectively represent the stratified loss of readmission prediction for medical coding information, demographic information, and clinical text information; for the feature data mm in multi-source data t , calculate the cross entropy loss y mm Indicates characteristic data mm t The corresponding ground truth value of readmission, Indicates characteristic data mm t Prediction of hospital readmission.

Images