Method for predicting in-hospital mortality risk based on icu medical record data

Abstract

The application discloses an in-hospital death risk prediction method based on ICU medical record data, and relates to the technical field of medical record data processing. The method comprises the following steps: acquiring historical medical record data sets of ICU patients and processing the same to obtain a training set sequence, a verification set sequence and a test set sequence; constructing an in-hospital death risk prediction preliminary model and performing training, verification and testing to obtain an in-hospital death risk prediction model; and acquiring medical record data of a target to be predicted and inputting the same into the in-hospital death risk prediction model to obtain an in-hospital death risk prediction result of the target to be predicted. The in-hospital death risk prediction method based on ICU medical record data provided by the application models dynamic clinical time series and static demographic statistical data to generate final patient representations for prediction. The method extracts time sequence information of the clinical time series from a local and global perspective and effectively combines the time sequence information and the static data. Therefore, the method has high reliability, good accuracy and good practicability.

Description

Technical Field

[0001] This invention belongs to the field of big data processing, specifically relating to a method for predicting in-hospital mortality risk based on ICU medical record data. Background Technology

[0002] With economic and technological development and the improvement of people's living standards, people are paying more and more attention to medical resources. Therefore, how to plan and allocate medical resources has become a key research focus for researchers.

[0003] The number and management of intensive care unit (ICU) beds have always been important factors reflecting medical resources. In-hospital mortality probability data also influences the planning and allocation of medical resources to some extent. Therefore, predicting in-hospital mortality risk has become a new research hotspot.

[0004] Currently, methods for predicting in-hospital mortality risk are generally based on electronic medical record (EMR) data. Many studies now treat EMR data as multivariate time series data, using deep learning methods to extract relevant representations of patients' health status for mortality risk prediction. While these methods show promising applications, they still have some limitations in capturing the deep dependencies between clinical events.

[0005] Currently, commonly used technologies for predicting in-hospital mortality risk generally rely on recurrent neural network (RNN) architectures, parallel convolutional neural networks (CNNs) and attention mechanisms, or Transformer-based technologies. While these technologies can solve some technical problems, they all suffer from poor reliability, low accuracy, and limited practicality. Summary of the Invention

[0006] The purpose of this invention is to provide a reliable, accurate, and practical method for predicting in-hospital mortality risk based on ICU medical record data.

[0007] The in-hospital mortality risk prediction method based on ICU medical record data provided by this invention includes the following steps:

[0008] S1. Obtain the historical medical record dataset of ICU patients;

[0009] S2. Process the medical record dataset obtained in step S1 to obtain the training set sequence, validation set sequence, and test set sequence;

[0010] S3. Based on the temporal representation learning module and the multi-view representation fusion module based on the gating mechanism, a preliminary model for predicting in-hospital mortality risk is constructed.

[0011] S4. The preliminary model for predicting in-hospital mortality risk obtained in step S3 is trained using the training set sequence obtained in step S2, and validated and tested using the validation set sequence and the test set sequence to obtain the in-hospital mortality risk prediction model.

[0012] S5. Obtain the medical record data of the target to be predicted and input it into the in-hospital mortality risk prediction model obtained in step S4 to obtain the in-hospital mortality risk prediction result of the target to be predicted.

[0013] Step S1, which involves obtaining the historical medical record dataset of ICU patients, specifically includes the following steps:

[0014] Select medical record data of ICU patients whose ICU stay exceeded 48 hours;

[0015] ICU patients were divided into positive and negative samples, where positive samples represent ICU patients who eventually died and negative samples represent ICU patients who eventually survived.

[0016] Step S2 involves processing the medical record dataset obtained in step S1 to obtain the training set sequence, validation set sequence, and test set sequence. Specifically, this includes the following steps:

[0017] For each ICU patient, clinical variables are selected from the electronic medical record as a dynamic clinical time series, and demographic data are selected from the electronic medical record as a demographic data series.

[0018] Dynamic clinical time series are based on hourly units, and medical records within one hour are summarized as a clinical event;

[0019] The final data is divided into training set sequences, validation set sequences, and test set sequences. The training set sequences include training set clinical time-series sequences and training set demographic data sequences. The validation set sequences include validation set clinical time-series sequences and validation set demographic data sequences. The test set sequences include test set clinical time-series sequences and test set demographic data sequences.

[0020] The process of processing the medical record dataset obtained in step S1 to obtain the training set sequence, validation set sequence, and test set sequence specifically includes the following steps:

[0021] The dynamic clinical time series of ICU patients is denoted as P = [r1, r2, ..., r t ,...,r T ], where r t Let r be the number of clinical events at time t, and T be the number of clinical events; each clinical event r t For r t =[v1,v2,...,v f ,...,v F]∈R F It consists of various clinical variables of the patient, including continuous numerical variables and categorical discrete variables, v f Let f be the f-th clinical variable, and F be the number of clinical variables; during processing, the categorical discrete variables are treated as one-hot vectors and concatenated with the numerical continuous variables; the demographic data sequence of ICU patients is S = [s1, s2, ..., s...]. m ,...,s M ], s m Let M be the population data of the m-th individual, and M be the number of population data.

[0022] Step S3, which involves the temporal representation learning module and the multi-view representation fusion module based on gating mechanisms, constructs a preliminary model for predicting in-hospital mortality risk. This process includes the following steps:

[0023] A. Input the obtained training set sequence into the local temporal representation learning module to obtain the context representation;

[0024] B. Input the contextual representation obtained in step A into the global temporal representation learning module to obtain a comprehensive representation;

[0025] C. The multi-view representation fusion module based on gating mechanism is used to aggregate the comprehensive representation obtained in step B at time steps, and then fused with the result obtained in step A to obtain the target representation vector;

[0026] D. Input the target representation vector obtained in step C into the classifier for prediction to obtain the final mortality prediction risk result.

[0027] Step A, which involves inputting the obtained training set sequence into the local temporal representation learning module to obtain the contextual representation, specifically includes the following steps:

[0028] The local temporal representation learning module includes a clinical temporal sequence embedding layer, a demographic data embedding layer, and a Bi-GRU layer based on a local attention mechanism.

[0029] The aforementioned clinical time series embedding layer specifically involves passing the clinical time series through a one-layer feedforward network (FFN) with linear rectified units (ReLU) to obtain an embedding representation of the clinical time series. Specifically, the embedding representation of the clinical time series is calculated using the following formula:

[0030] x t =ReLU(W x r t +b x )

[0031] In the formula x tThe clinical time series embedding representation is at time t; ReLU() is a linear rectified unit; W x This is the weight matrix for clinical variables; r t b is the clinical time series sequence at time t input; x It is the bias vector;

[0032] The aforementioned demographic data embedding layer specifically encodes demographic data sequences into an embedding representation x that corresponds to the clinical book sequence through a linear mapping. t Within the same space, the embedding representation of the demographic data sequence is obtained; in specific implementation, the embedding representation of the demographic data sequence is calculated using the following formula:

[0033] d = W d S

[0034] In the formula, d represents the embedding representation of the demographic data sequence; W d S is the mapping matrix obtained through training; S is the input demographic data sequence.

[0035] The Bi-GRU layer based on the local attention mechanism specifically employs a Bi-GRU network to learn the temporal dependencies of the embedded representations of clinical time sequences, and uses the embedded representations d of demographic data as background information for ICU patients to initialize the hidden state units of the Bi-GRU network, thereby enhancing semantic information; in specific implementation, let The initial forward GRU hidden unit state is used; then, the embedding representation x of a given clinical time sequence is... t and the previous hidden unit state The hidden cell state of the forward GRU at time t is calculated. for Where GRU() is the gated loop unit function; then, let Given the initial backward GRU hidden cell state, obtain the backward GRU hidden cell state. for Then, the hidden cell state of the forward GRU is... and the hidden cell state of the backward GRU By splicing the data, we obtain the hidden unit state h of the Bi-GRU network at time t. t for The hidden state H obtained from the output of the Bi-GRU network is H = [h1, h2, ..., h...]. t ,...,h T ];

[0036] Finally, a local attention mechanism is employed to learn contextual representations of the importance of clinical events in ICU patients from the hidden state H. Specifically, a single-layer feedforward network is first used, followed by the use of a softmax function to learn the local attention weights a for the t-th clinical event. t For a t =σ((W) a ) T h t +b a ), where σ() is the softmax function, W a b is the learned weight vector a The learned offset value; based on the local attention weights a t Calculate the context representation vector c t For c t =a t ·h t Finally, the context representation C is obtained as C = [c1, c2, ..., c t ,...,c T ].

[0037] Step B, which involves inputting the contextual representation obtained in step A into the global temporal representation learning module to obtain a comprehensive representation, specifically includes the following steps:

[0038] The contextual representation obtained in step A is input into the global temporal representation learning module to learn an enhanced representation that includes global temporal dependencies, and a comprehensive representation is obtained by combining local and global temporal dependencies.

[0039] The global temporal representation learning module uses a stacked L-layer Transformer network to learn global temporal dependencies from contextual representations.

[0040] The Transformer network consists of positional encoding and L consecutive blocks, each of which includes a multi-head self-attention mechanism (MHSA) and a two-layer FFN module.

[0041] In practice, positional encoding is first added to the context vector representation using the following formula:

[0042] e t =c t +p t

[0043] In the formula e t c is the encoded context representation vector; t p is the context representation vector; t For position encoding, and position encoding p with an odd number of bits t,2k-1 for Even-numbered position code pt,2k for n is the context representation vector c t Dimension size;

[0044] Then, e t Packed into a matrix E = [e1, e2, ..., e t ,...,e T The matrix is ​​then fed into a linear mapping layer to generate the initial matrix. The subscript 0 indicates the first layer of the Transformer network;

[0045] Then, a multi-head self-attention mechanism (MHSA) with h attention heads is used to capture hidden dependencies between clinical events:

[0046]

[0047] In the formula This represents the hidden dependencies between clinical events; MHSA() is the multi-head self-attention mechanism function; Norm() is the batch normalization layer processing function; This represents the output of the (l-1)th layer of the Transformer network, where 1 ≤ l ≤ L, and L is the number of layers in the Transformer network.

[0048] Next, a two-layer FFN module is used to hide the dependencies between clinical events. Perform nonlinear transformation:

[0049]

[0050] In the formula is the output of the l-th layer Transformer network; FFN() is the processing function, and FFN(x) = W2(ReLU(W1x+b1)+b2), where W1 and W2 are the mapping matrices of the two layers of FFN modules, b1 and b2 are the bias vectors of the two layers of FFN modules, and ReLU() is the linear rectifier unit function;

[0051] Finally, take the result obtained from the last layer of the Transformer network. The global temporal dependency of the final clinical event is concatenated with the contextual representation to obtain the comprehensive representation U of the ICU patient.

[0052] Step C, which involves using a gated multi-view representation fusion module to aggregate the comprehensive representation obtained in step B over time and then fusing it with the result obtained in step A to obtain the target representation vector, specifically includes the following steps:

[0053] A multi-view representation fusion module based on a gating mechanism is used to aggregate the comprehensive representation obtained in step B over time to obtain a representation vector of dynamic features, and then it is fused with the embedded representation of the demographic data sequence obtained in step A to obtain the target representation vector.

[0054] In practice, the dynamic feature representation vector uses a Bi-GRU network to dynamically aggregate the comprehensive representations from several time steps from two directions, forming a robust unified dynamic feature representation vector g. T For g T =Bi-GRU(U), where Bi-GRU() is a bidirectional gated cyclic unit function and U is the comprehensive characterization of ICU patients;

[0055] Then, a concatenation operation is used to fuse the dynamic feature representation vector g. T The embedding representation d of the demographic data sequence yields the final target representation vector z as z = [g T ,d].

[0056] Step D, which involves inputting the target representation vector obtained in step C into a classifier for prediction to obtain the final mortality prediction risk result, specifically includes the following steps:

[0057] The target representation vector z obtained in step C is input into a classifier containing a sigmoid activation function for prediction, and the final mortality prediction risk result is obtained.

[0058] In practice, the classifier uses a fully connected layer containing a sigmoid activation function for prediction, and the calculation formula is as follows:

[0059]

[0060] In the formula For the final mortality prediction risk outcome; W y and b y All of these are parameters obtained through learning.

[0061] The training described in step S4 specifically includes the following steps:

[0062] During training, the cross-entropy loss is calculated by comparing the final predicted mortality risk with the actual mortality risk value, and the model parameters are optimized based on the cross-entropy result.

[0063] In practice, the cross-entropy loss function is:

[0064]

[0065] In the formula, θ represents all trainable parameters; N is the number of samples; y i The true labels for the samples; The predicted labels are obtained from the model.

[0066] The in-hospital mortality risk prediction method based on ICU medical record data provided by this invention generates a final patient representation for prediction by modeling dynamic clinical time series and static demographic statistics. The method of this invention extracts the temporal information of clinical time series from both local and global perspectives and effectively combines temporal information and static data. Therefore, the method of this invention has high reliability, good accuracy and practicality. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0068] Figure 2 This is a schematic diagram of the prediction model structure of the method of the present invention. Detailed Implementation

[0069] like Figure 1 The diagram shown illustrates the method flow of this invention: This method for predicting in-hospital mortality risk based on ICU medical record data, provided by this invention, includes the following steps:

[0070] S1. Obtain the historical medical record dataset of ICU patients; specifically including the following steps:

[0071] Select medical record data of ICU patients whose ICU stay exceeded 48 hours;

[0072] ICU patients were divided into positive and negative samples, where positive samples represented ICU patients who eventually died and negative samples represented ICU patients who eventually survived.

[0073] S2. Process the medical record dataset obtained in step S1 to obtain the training set sequence, validation set sequence, and test set sequence; specifically including the following steps:

[0074] For each ICU patient, set clinical variables are selected from the electronic medical record as a dynamic clinical time series, and demographic data are selected from the electronic medical record as a demographic data series.

[0075] Dynamic clinical time series are based on hourly units, and medical records within one hour are summarized as a clinical event;

[0076] The final data is divided into training set sequences, validation set sequences, and test set sequences; the training set sequences include training set clinical time-series sequences and training set demographic data sequences; the validation set sequences include validation set clinical time-series sequences and validation set demographic data sequences; and the test set sequences include test set clinical time-series sequences and test set demographic data sequences.

[0077] The specific implementation includes the following steps:

[0078] The dynamic clinical time series of ICU patients is denoted as P = [r1, r2, ..., r t ,...,r T ], where r t Let r be the number of clinical events at time t, and T be the number of clinical events; each clinical event r t For r t =[v1,v2,...,v f ,...,v F ]∈R F It consists of various clinical variables of the patient, including continuous numerical variables and categorical discrete variables, v f Let f be the f-th clinical variable, and F be the number of clinical variables; during processing, the categorical discrete variables are treated as one-hot vectors and concatenated with the numerical continuous variables; the demographic data sequence of ICU patients is S = [s1, s2, ..., s...]. m ,...,s M ], s m Let M be the population data of the m-th individual, and M be the number of population data points.

[0079] S3. Based on the temporal representation learning module and the multi-view representation fusion module based on the gating mechanism, a preliminary model for predicting in-hospital mortality risk is constructed (model structure as follows). Figure 2 (As shown); specifically, it includes the following steps:

[0080] A. Input the obtained training set sequence into the local temporal representation learning module to obtain the contextual representation; specifically, this includes the following steps:

[0081] The local temporal representation learning module includes a clinical temporal sequence embedding layer, a demographic data embedding layer, and a Bi-GRU layer based on a local attention mechanism.

[0082] The aforementioned clinical time series embedding layer specifically involves passing the clinical time series through a one-layer feedforward network (FFN) with linear rectified units (ReLU) to obtain an embedding representation of the clinical time series. Specifically, the embedding representation of the clinical time series is calculated using the following formula:

[0083] x t =ReLU(W x r t +b x )

[0084] In the formula x t The embedded representation of the clinical time series at time t; ReLU() is the linear rectified unit function; W x This is the weight matrix for clinical variables; r tb is the clinical time series sequence at time t input; x It is the bias vector;

[0085] The aforementioned demographic data embedding layer specifically encodes demographic data sequences into an embedding representation x that corresponds to the clinical book sequence through a linear mapping. t Within the same space, the embedding representation of the demographic data sequence is obtained; in specific implementation, the embedding representation of the demographic data sequence is calculated using the following formula:

[0086] d = W d S

[0087] In the formula, d represents the embedding representation of the demographic data sequence; W d S is the mapping matrix obtained through training; S is the input demographic data sequence.

[0088] The Bi-GRU layer based on the local attention mechanism specifically employs a Bi-GRU network to learn the temporal dependencies of the embedded representations of clinical time sequences, and uses the embedded representations d of demographic data as background information for ICU patients to initialize the hidden state units of the Bi-GRU network, thereby enhancing semantic information; in specific implementation, let The initial forward GRU hidden unit state is used; then, the embedding representation x of a given clinical time sequence is... t and the previous hidden unit state The hidden cell state of the forward GRU at time t is calculated. for Where GRU() is the gated loop unit function; then, let Given the initial backward GRU hidden cell state, obtain the backward GRU hidden cell state. for Then, the hidden cell state of the forward GRU is... and the hidden cell state of the backward GRU By splicing the data, we obtain the hidden unit state h of the Bi-GRU network at time t. t for The hidden state H obtained from the output of the Bi-GRU network is H = [h1, h2, ..., h...]. t ,...,h T ];

[0089] Finally, a local attention mechanism is employed to learn contextual representations of the importance of clinical events in ICU patients from the hidden state H. Specifically, a single-layer feedforward network is first used, followed by the use of a softmax function to learn the local attention weights a for the t-th clinical event. t For a t =σ((W) a )T h t +b a ), where σ() is the softmax function, W a b is the learned weight vector a The learned offset value; based on the local attention weights a t Calculate the context representation vector c t For c t =a t ·h t Finally, the context representation C is obtained as C = [c1, c2, ..., c t ,...,c T ];

[0090] B. Input the contextual representation obtained in step A into the global temporal representation learning module to obtain a comprehensive representation; specifically, this includes the following steps:

[0091] The contextual representation obtained in step A is input into the global temporal representation learning module to learn an enhanced representation that includes global temporal dependencies, and a comprehensive representation is obtained by combining local and global temporal dependencies.

[0092] The global temporal representation learning module uses a stacked L-layer Transformer network to learn global temporal dependencies from contextual representations.

[0093] The Transformer network consists of positional encoding and L consecutive blocks, each of which includes a multi-head self-attention mechanism (MHSA) and a two-layer FFN module.

[0094] In practice, positional encoding is first added to the context vector representation using the following formula:

[0095] e t =c t +p t

[0096] In the formula e t c is the encoded context representation vector; t p is the context representation vector; t For position encoding, and position encoding p with an odd number of bits t,2k-1 for Even-numbered position code p t,2k for n is the context representation vector c t Dimension size;

[0097] Then, e t Packed into a matrix E = [e1, e2, ..., e t ,...,eT The matrix is ​​then fed into a linear mapping layer to generate the initial matrix. The subscript 0 indicates the first layer of the Transformer network;

[0098] Then, a multi-head self-attention mechanism (MHSA) with h attention heads is used to capture hidden dependencies between clinical events:

[0099]

[0100] In the formula This represents the hidden dependencies between clinical events; MHSA() is the multi-head self-attention mechanism function; Norm() is the batch normalization layer processing function; This represents the output of the (l-1)th layer of the Transformer network, where 1 ≤ l ≤ L, and L is the number of layers in the Transformer network.

[0101] Next, a two-layer FFN module is used to hide the dependencies between clinical events. Perform nonlinear transformation:

[0102]

[0103] In the formula is the output of the l-th layer Transformer network; FFN() is the processing function, and FFN(x) = W2(ReLU(W1x+b1)+b2), where W1 and W2 are the mapping matrices of the two layers of FFN modules, b1 and b2 are the bias vectors of the two layers of FFN modules, and ReLU() is the linear rectifier unit function;

[0104] Finally, take the result obtained from the last layer of the Transformer network. The global temporal dependency of the final clinical event is concatenated with the contextual representation to obtain the comprehensive representation U of the ICU patient.

[0105] C. A multi-view representation fusion module based on a gating mechanism is used to aggregate the comprehensive representation obtained in step B over time steps, and then fused with the result obtained in step A to obtain the target representation vector; specifically, the following steps are included:

[0106] A multi-view representation fusion module based on a gating mechanism is used to aggregate the comprehensive representation obtained in step B over time to obtain a representation vector of dynamic features, and then it is fused with the embedded representation of the demographic data sequence obtained in step A to obtain the target representation vector.

[0107] In practice, the dynamic feature representation vector uses a Bi-GRU network to dynamically aggregate the comprehensive representations from several time steps from two directions, forming a robust unified dynamic feature representation vector g. T For g T =Bi-GRU(U), where Bi-GRU() is a bidirectional gated cyclic unit function and U is the comprehensive characterization of ICU patients;

[0108] Then, a concatenation operation is used to fuse the dynamic feature representation vector g. T The embedding representation d of the demographic data sequence yields the final target representation vector z as z = [g T ,d];

[0109] D. Input the target representation vector obtained in step C into the classifier for prediction to obtain the final mortality prediction risk result; specifically including the following steps:

[0110] The target representation vector z obtained in step C is input into a classifier containing a sigmoid activation function for prediction, and the final mortality prediction risk result is obtained.

[0111] In practice, the classifier uses a fully connected layer containing a sigmoid activation function for prediction, and the calculation formula is as follows:

[0112]

[0113] In the formula For the final mortality prediction risk outcome; W y and b y All of these are parameters obtained through learning;

[0114] S4. The preliminary model for predicting in-hospital mortality risk obtained in step S3 is trained using the training set sequence obtained in step S2, and validated and tested using the validation set sequence and the test set sequence to obtain the in-hospital mortality risk prediction model.

[0115] During training, the cross-entropy loss is calculated by comparing the final predicted mortality risk with the actual mortality risk value, and the model parameters are optimized based on the cross-entropy result.

[0116] In practice, the cross-entropy loss function is:

[0117]

[0118] In the formula, θ represents all trainable parameters; N is the number of samples; y i The true labels for the samples; The predicted labels obtained from the model;

[0119] S5. Obtain the medical record data of the target to be predicted and input it into the in-hospital mortality risk prediction model obtained in step S4 to obtain the in-hospital mortality risk prediction result of the target to be predicted.

[0120] The following examples compare the method of the present invention with the prior art:

[0121] Bi-GRU: This method is a standard bidirectional gated recurrent unit network, where the hidden state vector of the last time step is used to predict the risk of death.

[0122] Transformer e It is the encoder of the Transformer network. Here, the output of the last encoder is flattened and a fully connected layer is used for death risk prediction;

[0123] Retain: This method is a two-level neural attention model that predicts by detecting influential historical clinical events and important clinical variables;

[0124] SAnd: This method learns patient representations from clinical time-series sequences based on a masked self-attention mechanism and a dense interpolation strategy, and applies them to downstream clinical prediction tasks.

[0125] ConCare: This method embeds clinical feature sequences individually and uses a multi-head self-attention mechanism to capture the interdependencies between dynamic features and static baseline information for mortality risk prediction.

[0126] AdaCare: This method uses dilated convolutions with multi-scale receptive fields and gated recurrent units to capture long-term and short-term temporal information from EMR data to predict mortality risk.

[0127] Since most comparisons do not consider demographic data, this application incorporates the embedding vector d of demographic data into the final patient representation of these methods for predicting mortality risk in order to fairly compare predictive performance.

[0128] The in-hospital mortality risk prediction method of this invention is defined as a binary classification task, and the real-world dataset is imbalanced. In experiments, this invention uses AUROC, AUPRC, and min(Se,P+) evaluation metrics to assess the method's performance. AUROC and AUPRC are the most valuable evaluation metrics for classifying imbalanced data, while min(Se,P+) is a custom metric proposed in the 2012 Physionet / CinC Challenge. Higher values ​​indicate a stronger ability of the model to distinguish between positive and negative samples, i.e., higher prediction accuracy. This invention reports the mean and standard deviation of each performance metric on the test set for MIMIC-III and e-ICU.

[0129] The method of the present invention and the comparative method were evaluated on the same test set. The experimental results are shown in Table 1, where the numbers in parentheses are the standard deviations.

[0130] Table 1. Schematic diagram of the prediction performance comparison between the method of the present invention and the comparative method.

[0131]

[0132] As shown in Table 1, the method proposed in this invention achieves the best performance among all comparative methods on the MIMIC-III and e-ICU datasets. Specifically, the average AUROC, AUPRC, and min(Se,P+) values ​​obtained from the model evaluation on the MIMIC-III test set are 0.8680 (0.001), 0.5254 (0.002), and 0.5138 (0.005) for the five experiments, respectively. On the e-ICU test set, the average AUROC, AUPRC, and min(Se,P+) values ​​obtained from the model evaluation on the five experiments are 0.8733 (0.005), 0.5801 (0.014), and 0.5588 (0.017) for the five experiments, respectively.

[0133] Therefore, experimental results on the MIMIC-III and e-ICU datasets show that the proposed method has good EMR time-series dependency extraction ability and patient representation learning ability, and can accurately predict the risk of in-hospital mortality.

Claims

1. A method for predicting in-hospital mortality risk based on ICU medical record data, comprising the following steps: S1. Obtain the historical medical record dataset of ICU patients; S2. Process the medical record dataset obtained in step S1 to obtain the training set sequence, validation set sequence, and test set sequence; S3. Based on the temporal representation learning module and the multi-view representation fusion module based on the gating mechanism, a preliminary model for predicting in-hospital mortality risk is constructed; specifically, the following steps are included: A. Input the obtained training set sequence into the local temporal representation learning module to obtain the context representation; B. Input the contextual representation obtained in step A into the global temporal representation learning module to obtain a comprehensive representation; C. A multi-view representation fusion module based on a gating mechanism is used to aggregate the comprehensive representation obtained in step B over time steps, and then fused with the result obtained in step A to obtain the target representation vector; D. Input the target representation vector obtained in step C into the classifier for prediction to obtain the final mortality prediction risk result; S4. The preliminary model for predicting in-hospital mortality risk obtained in step S3 is trained using the training set sequence obtained in step S2, and validated and tested using the validation set sequence and the test set sequence to obtain the in-hospital mortality risk prediction model. S5. Obtain the medical record data of the target to be predicted and input it into the in-hospital mortality risk prediction model obtained in step S4 to obtain the in-hospital mortality risk prediction result of the target to be predicted.

2. The method for predicting in-hospital mortality risk based on ICU medical record data according to claim 1, characterized in that... Step S1, which involves obtaining the historical medical record dataset of ICU patients, specifically includes the following steps: Select medical record data of ICU patients whose ICU stay exceeded 48 hours; ICU patients were divided into positive and negative samples, where positive samples represent ICU patients who eventually died and negative samples represent ICU patients who eventually survived.

3. The method for predicting in-hospital mortality risk based on ICU medical record data according to claim 2, characterized in that... Step S2 involves processing the medical record dataset obtained in step S1 to obtain the training set sequence, validation set sequence, and test set sequence. Specifically, this includes the following steps: For each ICU patient, set clinical variables are selected from the electronic medical record as a dynamic clinical time series, and demographic data are selected from the electronic medical record as a demographic data series. Dynamic clinical time series are based on hourly units, and medical records within one hour are summarized as a clinical event; The final data is divided into training set sequences, validation set sequences, and test set sequences. The training set sequences include training set clinical time-series sequences and training set demographic data sequences. The validation set sequences include validation set clinical time-series sequences and validation set demographic data sequences. The test set sequences include test set clinical time-series sequences and test set demographic data sequences.

4. The method for predicting in-hospital mortality risk based on ICU medical record data according to claim 3, characterized in that... The process of processing the medical record dataset obtained in step S1 to obtain the training set sequence, validation set sequence, and test set sequence specifically includes the following steps: The dynamic clinical time series of ICU patients is denoted as... ,in for Clinical events at specific moments The number of clinical events; each clinical event for It consists of various clinical variables of the patient, including continuous numerical variables and categorical discrete variables. For the first One clinical variable, The number of clinical variables; during processing, categorical discrete variables are treated as... Vectors, concatenated with continuous numerical variables; demographic data sequence of ICU patients. , For the m-th person's demographic data, The quantity of demographic data.

5. The method for predicting in-hospital mortality risk based on ICU medical record data according to claim 4, characterized in that... Step A, which involves inputting the obtained training set sequence into the local temporal representation learning module to obtain the contextual representation, specifically includes the following steps: The local temporal representation learning module includes a clinical temporal sequence embedding layer, a demographic data embedding layer, and a Bi-GRU layer based on a local attention mechanism. The aforementioned clinical time series embedding layer specifically involves passing the clinical time series through a one-layer feedforward network (FFN) with linear rectified units (ReLU) to obtain an embedding representation of the clinical time series. Specifically, the embedding representation of the clinical time series is calculated using the following formula: In the formula The embedded representation of the clinical time sequence at time t; For linear rectifier unit functions; This is the weight matrix for clinical variables; The input is the clinical time series sequence at time t; It is the bias vector; The aforementioned demographic data embedding layer specifically encodes demographic data sequences into an embedding representation that corresponds to clinical book sequence sequences through a linear mapping. Within the same space, the embedding representation of the demographic data sequence is obtained; in specific implementation, the embedding representation of the demographic data sequence is calculated using the following formula: In the formula An embedded representation of demographic data sequences; The mapping matrix obtained during training; The input is a sequence of demographic data; The Bi-GRU layer based on the local attention mechanism specifically employs a Bi-GRU network to learn the temporal dependencies of the embedding representations of clinical time sequences, and uses embedding representations of demographic data. The hidden state units of the Bi-GRU network are initialized with background information of ICU patients to enhance semantic information; in specific implementation, let... , The initial forward GRU hidden unit states are used; then, the embedding representation of a given clinical time sequence is performed. and the previous hidden unit state The hidden cell state of the forward GRU at time t is calculated. for ,in This is a gated loop unit function; then, let , Given the initial backward GRU hidden cell state, obtain the backward GRU hidden cell state. for Then, the hidden cell state of the forward GRU is... and the hidden cell state of the backward GRU By splicing the data, we obtain the hidden unit states of the Bi-GRU network at time t. for Hidden states obtained from the output of the Bi-GRU network for ; Finally, a local attention mechanism is used to start from the hidden state. The learning process yields contextual representations of the importance of clinical events in ICU patients. Specifically, a single-layer feedforward network is first used, followed by the softmax function to learn the local attention weights for the t-th clinical event. for ,in For the softmax function, The learned weight vector, The learned offset value; based on the local attention weights Calculate the context representation vector for ; Finally, the context representation is obtained. for .

6. The method for predicting in-hospital mortality risk based on ICU medical record data according to claim 5, characterized in that... Step B, which involves inputting the contextual representation obtained in step A into the global temporal representation learning module to obtain a comprehensive representation, specifically includes the following steps: The contextual representation obtained in step A is input into the global temporal representation learning module to learn an enhanced representation that includes global temporal dependencies, and a comprehensive representation is obtained by combining local and global temporal dependencies. The global temporal representation learning module uses a stacked L-layer Transformer network to learn global temporal dependencies from contextual representations. The Transformer network consists of positional encoding and L consecutive blocks, each of which includes a multi-head self-attention mechanism (MHSA) and a two-layer FFN module. In practice, positional encoding is first added to the context vector representation using the following formula: In the formula This is the encoded context representation vector; For context representation vectors; Position encoding, and position encoding for odd-numbered positions. for Even-numbered position encoding for , , Context representation vector Dimension size; Then, Packed into a matrix Then it is fed into the linear mapping layer to generate the initial matrix. The subscript 0 indicates the first layer of the Transformer network; Then, a multi-head self-attention mechanism (MHSA) with h attention heads is used to capture hidden dependencies between clinical events: In the formula Hidden dependencies between clinical events; This is a function for the multi-head self-attention mechanism; This is the batch normalization layer processing function; For the first The output of the layered Transformer network, , The number of layers in the Transformer network; Next, a two-layer FFN module is used to hide the dependencies between clinical events. Perform nonlinear transformation: In the formula For the first The output of the layered Transformer network; For processing functions, and ,in and These are the mapping matrices for the two layers of FFN modules. and These are the bias vectors for the two FFN modules, respectively. For linear rectifier unit functions; Finally, take the result obtained from the last layer of the Transformer network. The global temporal dependency of the final clinical event is concatenated with the contextual representation to obtain the comprehensive representation of ICU patients. for .

7. The method for predicting in-hospital mortality risk based on ICU medical record data according to claim 6, characterized in that... Step C, which involves using a gated multi-view representation fusion module to aggregate the comprehensive representation obtained in step B over time and then fusing it with the result obtained in step A to obtain the target representation vector, specifically includes the following steps: A multi-view representation fusion module based on a gating mechanism is used to aggregate the comprehensive representation obtained in step B over time to obtain a representation vector of dynamic features, and then it is fused with the embedded representation of the demographic data sequence obtained in step A to obtain the target representation vector. In practice, the representation vector of dynamic features adopts a Bi-GRU network, which dynamically aggregates the comprehensive representations of several time steps from two directions to form a robust unified dynamic feature representation vector. for ,in This is a bidirectional gated loop unit function. A comprehensive characteristic of ICU patients; Then, a concatenation operation is used to fuse the dynamic feature representation vectors. Embedded representation of demographic data sequences The final target representation vector is obtained. for .

8. The method for predicting in-hospital mortality risk based on ICU medical record data according to claim 7, characterized in that... Step D, which involves inputting the target representation vector obtained in step C into a classifier for prediction to obtain the final mortality prediction risk result, specifically includes the following steps: The target representation vector obtained in step C The input is fed into a classifier containing a sigmoid activation function for prediction, resulting in the final mortality prediction risk. In practice, the classifier uses a fully connected layer containing a sigmoid activation function for prediction, and the calculation formula is as follows: In the formula For the final mortality prediction risk outcome; and All of these are parameters obtained through learning.

9. The method for predicting in-hospital mortality risk based on ICU medical record data according to claim 8, characterized in that... The training described in step S4 specifically includes the following steps: During training, the cross-entropy loss is calculated by comparing the final predicted mortality risk with the actual mortality risk value, and the model parameters are optimized based on the cross-entropy result. In practice, the cross-entropy loss function is: In the formula Represents all trainable parameters; The number of samples; The true labels for the samples; The predicted labels are obtained from the model.

Images