A method, device, equipment and medium for patient clinical risk assessment

Abstract

This application relates to the field of data processing technology and provides a method, apparatus, device, and medium for patient clinical risk assessment. The assessment method includes: processing time-series electronic medical record data of a target patient to generate a primary patient representation vector; obtaining an absolute offset and its embedded representation; obtaining a relative offset and its embedded representation; extracting features from the absolute offset embedding representation to obtain absolute offset features; extracting features from the relative offset embedding representation to obtain relative offset features; fusing the absolute and relative offset features using the primary patient representation vector to obtain a final patient representation vector; and calculating the final patient representation vector using a clinical risk assessment module to obtain the patient's clinical risk result at the current time T. The assessment method of this application can improve the accuracy of risk assessment.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, device, equipment and medium for clinical risk assessment of patients. Background Technology

[0002] The Intensive Care Unit (ICU) is a specialized medical facility that applies critical care medicine theories and modern high-tech medical equipment. It is primarily designed to treat critically ill patients and employs a closed design to avoid external interference and contamination. The construction, operation, and human resource costs of ICU units equipped with high-end equipment are substantial. Patients in the ICU require close monitoring with various advanced medical equipment such as electrocardiogram monitors and ventilators, as well as 24-hour shifts of highly skilled critical care medicine staff for treatment and care.

[0003] Critically ill patients in the ICU often have complex and fluctuating conditions. Assessing their health status relies heavily on the subjective judgment of medical staff with extensive clinical knowledge and experience. Furthermore, the prolonged stay of critically ill patients in the ICU consumes significant medical costs and resources, posing a substantial challenge to ICU treatment decisions. Improving the timeliness and accuracy of treatment decisions and reducing the workload of clinical medical staff are key issues in improving the quality of ICU medical services and increasing the utilization rate of medical resources. Therefore, designing effective clinical decision support models that utilize rich ICU clinical data to make timely and accurate diagnoses of patients' conditions is crucial for achieving effective and rational treatment interventions, improving the prognosis of critically ill patients, enhancing the quality of further treatment, and raising the overall level of medical services. In current clinical practice prediction tasks, artificial intelligence models typically use data from a single patient as input, and after a series of inference operations, output a corresponding clinical feature vector representing the patient's data in the computer system. This vector is used to calculate and predict the specific patient's disease risk. However, existing medical AI predictions suffer from inaccurate patient clinical risk assessment. Summary of the Invention

[0004] This application provides a method, device, equipment, and medium for patient clinical risk assessment, which can solve the problem of inaccurate patient clinical risk assessment.

[0005] In a first aspect, embodiments of this application provide a method for assessing clinical risk in patients, the method comprising: The electronic medical record time-series data of the target patient is processed to generate a primary representation vector of the patient; the electronic medical record time-series data includes: multiple physiological indicators of the target patient in... T The value at time i, the i-th time TThe current moment is represented by the current moment. Obtain the absolute offset that reflects the degree of deviation between the target patient's physiological index value and the normal value, and obtain the absolute offset embedding representation of the absolute offset; Obtain the relative offsets that reflect the stability of the target patient's physiological indicators, and obtain the relative offset embedding representation of the relative offsets; Feature extraction is performed on the absolute offset embedding representation to obtain the absolute offset features; Feature extraction is performed on the relative offset embedding representation to obtain the relative offset features; The absolute offset feature and the relative offset feature are fused using the patient's primary representation vector to obtain the patient's final representation vector; The final patient representation vector is calculated using the clinical risk assessment module to obtain the patient's status at the current time. T Clinical risk outcomes.

[0006] Secondly, embodiments of this application provide a patient clinical risk assessment device, the assessment device comprising: The vector generation module processes the time-series data of the target patient's electronic medical record to generate a primary representation vector of the patient. The time-series data includes multiple physiological indicators of the target patient. T The value at time i, the i-th time The current moment is represented by the current moment. The first acquisition module acquires physiological indicators reflecting the target patient at the current moment. T The absolute offset of the deviation between the current value and the normal value is obtained, and the absolute offset embedding representation of the absolute offset is obtained. The second acquisition module acquires the relative offset used to reflect the stability of the physiological indicators of the target patient, and acquires the relative offset embedding representation of the relative offset. The first feature extraction module extracts features from the absolute offset embedding representation to obtain absolute offset features; The second feature extraction module extracts features from the relative offset embedding representation to obtain the relative offset features. The fusion module uses the patient's primary representation vector to fuse the absolute offset features and the relative offset features to obtain the patient's final representation vector; The assessment module uses the clinical risk assessment module to calculate the patient's final representation vector, obtaining the patient's status at the current moment. T Clinical risk outcomes.

[0007] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described patient clinical risk assessment method.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned patient clinical risk assessment method.

[0009] The above-mentioned solution in this application has the following beneficial effects: In the embodiments of this application, the electronic medical record time-series data of the target patient is processed to generate a primary patient representation vector. Then, an absolute offset reflecting the deviation of the target patient's physiological indicators from normal values ​​is obtained, along with an absolute offset embedding representation. Simultaneously, a relative offset reflecting the stability of the target patient's physiological indicators is obtained, along with a relative offset embedding representation. Feature extraction is then performed on the absolute offset embedding representation to obtain absolute offset features, and feature extraction is performed on the relative offset embedding representation to obtain relative offset features. The absolute and relative offset features are then fused using the primary patient representation vector to obtain the final patient representation vector. Finally, the clinical risk assessment module calculates the final patient representation vector to obtain the patient's current position at the current time. T The clinical risk outcome is determined by the absolute offset, which reflects the degree of deviation of the target patient's physiological indicators from normal values, and the relative offset, which reflects the stability of the target patient's physiological indicators. These features provide a direct and accurate representation of the patient. By fusing the absolute and relative offset features using the primary patient representation vector, a final patient representation vector is obtained. This feature fusion method can generate personalized patient representation vectors, improve the accuracy of the representation vectors, and enable sufficiently accurate clinical risk assessment of patients.

[0010] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart of a patient clinical risk assessment method provided in an embodiment of this application; Figure 2 A block diagram of a patient clinical risk assessment system provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a patient clinical risk assessment device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0014] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0015] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0016] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0017] The following is an illustrative description of the patient clinical risk assessment method provided in this application.

[0018] like Figure 1 As shown, the patient clinical risk assessment method provided in this application includes the following steps: Step 11: Process the time-series data of the target patient's electronic medical record to generate a primary representation vector of the patient.

[0019] The aforementioned electronic medical record time-series data includes: multiple physiological indicators of the target patient at... T The value at time i, the i-th time The current moment is defined as the moment in time.

[0020] Specifically, it can be done through formulas The time-series data of the target patient's electronic medical record are processed to obtain the patient's primary representation vector. .

[0021] The patient's primary representation vector is the current time step in the formula. T Medical hidden representation on .

[0022] in, This represents a time series model that processes the electronic medical record time series data of the target patient. Indicates the first t Medical feature concealment representation at a given moment, This indicates that multiple physiological indicators of the target patient are in the first... t The value at each moment, t =1,2,..., T .

[0023] It should be noted that the electronic medical record time-series data for each moment in the aforementioned target patients' electronic medical records... It consists of the values ​​of multiple physiological indicators, namely , Indicates the first t The moment of the first i The values ​​of a physiological indicator i =1,2,..., C In some embodiments of this application, the time-series data of the target patient's electronic medical record can be acquired using common acquisition methods, such as detection by medical devices. For example, the aforementioned physiological indicators may include blood pressure, heart rate, body temperature, etc.

[0024] It is worth mentioning that the dynamic features in the time-series data of patients' electronic medical records are captured by time series modeling methods to generate a primary representation vector of patients, which facilitates the calculation in subsequent steps.

[0025] Step 12: Obtain the absolute offset that reflects the degree of deviation between the target patient's physiological index value and the normal value, and obtain the absolute offset embedding representation of the absolute offset.

[0026] It should be noted that normal values ​​are those that are theoretically considered to be normal physiological indicators.

[0027] In some embodiments of this application, the absolute deviation between the current value and the normal value is calculated by mimicking the thinking and behavior of a doctor when dealing with abnormal deviations in the current value of a patient's physiological indicators. For example, if a patient's current blood pressure value exceeds the normal value, it indicates that the patient may have hypertension.

[0028] It is worth mentioning that the absolute offset obtained through the above steps can intuitively and accurately reflect the degree of deviation between the current value and the normal value of the patient's physiological indicators. However, the mathematical form of the absolute offset is not conducive to subsequent calculations. Therefore, the absolute offset is calculated to obtain the absolute offset embedding representation to facilitate subsequent calculations.

[0029] Step 13: Obtain the relative offset that reflects the stability of the target patient's physiological indicators, and obtain the relative offset embedding representation of the relative offset.

[0030] In some embodiments of this application, by mimicking the thinking and behavior of doctors when dealing with abnormal deviations in the values ​​of patients' physiological indicators, the relative deviation is obtained by calculating the values ​​of physiological indicators. For example, if the current value of a patient's heart rate rises or falls sharply, it indicates that the patient may have coronary heart disease.

[0031] It is worth mentioning that the relative offset obtained through the above steps can intuitively and accurately reflect the stability of physiological indicators. However, the mathematical form of the relative offset is not conducive to subsequent calculations. Therefore, the relative offset is calculated to obtain the relative offset embedding representation to facilitate subsequent calculations.

[0032] Step 14: Extract features from the absolute offset embedding representation to obtain absolute offset features.

[0033] In some embodiments of this application, a Long Short Term Memory (LSTM) network model can be used to extract features from the absolute offset embedding representation to obtain absolute offset features.

[0034] It is worth mentioning that the absolute offset is a quantification of the degree of abnormality of the current physiological indicators. The longer the abnormal period, the worse the patient's condition. By using the LSTM model to extract features from the absolute offset embedding representation, we can capture the temporal changes of the patient's condition, highlight the abnormal time points in the patient's electronic medical record time series data, and obtain the absolute offset features with high accuracy.

[0035] Step 15: Extract features from the relative offset embedding representation to obtain the relative offset features.

[0036] In some embodiments of this application, a multi-head self-attention model can be used to extract features from the relative offset embedding representation.

[0037] It is worth noting that, due to the irregular sampling intervals of each physiological indicator, data preprocessing requires filling the unsampled blank time points with fixed values ​​(i.e., the latest value or average value) to maintain a fixed time window size. This may result in the relative offset containing a large number of consecutive zero values ​​between two non-zero values. Therefore, it is necessary to capture the correlation between two non-adjacent and non-zero offsets with a long time interval, rather than simply focusing on the relationship between offsets at adjacent time points. However, traditional recurrent neural networks cannot directly capture the correlation between non-adjacent time points, and the continuous accumulation of zero values ​​in the memory cells of recurrent neural networks may gradually dilute the key features in the time series data. To address this issue, some embodiments of this application use a multi-head self-attention model to extract features from the relative offset. This model can simultaneously integrate feature information from all time points to enhance the learning ability of long-dependent features. Furthermore, the relative offset utilizes hidden representations as positional embeddings and supplementary information to ensure that the multi-head self-attention model can further aggregate the temporal features in the original data, improving the accuracy of the relative offset features.

[0038] Step 16: Use the patient's primary representation vector to fuse the absolute offset features and the relative offset features to obtain the patient's final representation vector.

[0039] In some embodiments of this application, the absolute offset embedding representation is calculated using the patient's primary representation vector to obtain the absolute offset feature weights, and the relative offset embedding representation is calculated using the primary representation vector to obtain the relative offset feature weights. The absolute offset features and relative offset features are then fused using the absolute offset feature weights and relative offset feature weights to obtain the final patient representation vector.

[0040] It is worth mentioning that by using the patient's primary representation vector to calculate the absolute and relative offset features, the relationship between the absolute and relative offset features can be fully explored. By fusing the two based on the relationship between the absolute and relative offset features, a personalized and highly accurate final patient representation vector can be constructed.

[0041] Step 17: Calculate the patient's final representation vector using the clinical risk assessment module to obtain the patient's current position. T Clinical risk outcomes.

[0042] In some embodiments of this application, the above-mentioned clinical risk assessment module can be a predictive classifier: ; in, This indicates the predicted results of relevant clinical risks. =0 indicates that the patient does not have any relevant clinical risk. =1 indicates that the patient has a relevant clinical risk. express Activation function Indicates a linear layer. This represents the final representation vector of the patient.

[0043] Cross-entropy loss is used as the loss function for backpropagation parameter updates. The loss function is: ; in, This represents the loss value of patient data. i= 1,2,..., B , B This represents the total number of patient data. Indicates the first in this batch i The actual results for each patient =0 indicates that the patient does not have any relevant clinical risk. =1 indicates that the patient has relevant clinical risks.

[0044] It should be noted that the patient's final representation vector can be used for disease risk prediction or in-hospital mortality risk prediction, etc.

[0045] It is worth mentioning that the calculated absolute offset, which reflects the degree of deviation of physiological indicators from normal values, and the relative offset, which reflects the stability of physiological indicators, provide a high degree of directness and accuracy in representing patients. By weighted summing of the two vector features and their weights, the final patient representation vector is obtained. This feature fusion method can generate more personalized patient representation vectors, improve the accuracy of the representation vectors, and enable sufficiently accurate clinical risk assessment of patients.

[0046] The specific steps of step 12 described above will be illustrated below with reference to specific embodiments.

[0047] In some embodiments of this application, the specific implementation process of step 12 above includes the following steps: Step 12.1: Obtain the absolute offset that reflects the degree of deviation between the target patient's physiological index value and the normal value.

[0048] Through formula Calculate the physiological indicators used to reflect the target patient at the current moment. T The absolute deviation of the value from the normal value.

[0049] in, Indicates the first [number] used to reflect the target patient i The physiological indicators at the first tThe absolute deviation of the value at any given time from the normal value. i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T , T Indicates the current time. Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i The normal values ​​corresponding to each physiological indicator This represents the inverse triangular nonlinear activation function.

[0050] Step 12.2: Obtain the absolute offset embedding representation of the absolute offset. .

[0051] In some embodiments of this application, the absolute offset can be processed by multi-scale one-dimensional convolution to obtain multiple physiological indicators at the 1st level. t Multiple absolute offsets corresponding to each moment The convolution result of the absolute offset embedding representation matrix The specific calculation formula is as follows: ; ; in, , Indicates the kernel size as k Multiple physiological indicators at the first t Multiple absolute offsets corresponding to each moment The absolute offset embeddings represent the convolution result of the matrix. express Dimensions The size of the vector dimension. t= 1,2,..., T , T Indicates the current moment. Indicates size is convolution kernel, Indicates along the time dimension Units of time The first in A segment, This represents a matrix consisting of multiple absolute offsets corresponding to multiple physiological indicators. q= 1,2,..., Q , Q Indicates the total number of segments. For bias parameters, For splicing operations, This indicates that when the kernel size is 3, multiple physiological indicators are in the first... t Multiple absolute offsets corresponding to each moment The absolute offset embeddings represent the convolution result of the matrix. This indicates that when the kernel size is 4, multiple physiological indicators are in the first stage. t Multiple absolute offsets corresponding to each moment The absolute offset embeddings represent the convolution result of the matrix. This indicates that when the kernel size is 5, multiple physiological indicators are in the first... t Multiple absolute offsets corresponding to each moment The absolute offset embedding represents the convolution result of the matrix.

[0052] The recorrection module calculates the... The weights are redistributed to obtain the absolute offset embedding representation.

[0053] Specifically, through the formula Calculate the first t Absolute offset embedding representation of the absolute offset at each moment .

[0054] in, and Both represent nonlinear activation functions. and All are parameter matrices.

[0055] It is worth mentioning that the absolute offset obtained through the above steps can intuitively and accurately reflect the degree of deviation between the patient's physiological indicators and normal values. However, the mathematical form of the absolute offset is not conducive to subsequent calculations. Therefore, multiple absolute offsets corresponding to multiple physiological indicators are calculated to obtain an embedded representation of the absolute offset, which facilitates subsequent calculations.

[0056] The specific steps of step 13 described above will be illustrated below with reference to specific embodiments.

[0057] In one embodiment of this application, the specific implementation process of step 13 above includes the following steps: 13.1 Obtain the relative offset used to reflect the stability of the target patient's physiological indicators.

[0058] Through the formula: ; Calculate the relative offset used to reflect the stability of physiological indicators of the target patient; in, Indicates the first [number] used to reflect the target patienti The physiological indicators at the first t The relative offset of stability at each moment i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T , T Indicates the current moment. Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i A physiological indicator in t- The value at time 1, This represents the inverse triangular nonlinear activation function.

[0059] 13.2 Obtaining the relative offset embedding representation of the relative offset .

[0060] In some embodiments of this application, the relative offset can be processed by multi-scale one-dimensional convolution to obtain multiple physiological indicators at the 1st... t Multiple relative offsets corresponding to each time point The convolution result of the relative offset embedding representation matrix The specific calculation formula is as follows: ; ; in, , Indicates the kernel size as k Multiple physiological indicators at the first t Multiple relative offsets corresponding to each time point The relative offset embeddings represent the convolution result of the matrix. express Dimensions The size of the vector dimension. t= 1,2,..., T , T Indicates the current moment. Indicates size is convolution kernel, Indicates along the time dimension Units of time The first in A segment, This represents a matrix consisting of multiple relative offsets corresponding to multiple physiological indicators. q= 1,2,..., Q , QIndicates the total number of segments. For bias parameters, For splicing operations, This indicates that when the kernel size is 3, multiple physiological indicators are in the first... t Multiple relative offsets corresponding to each time point The relative offset embeddings represent the convolution result of the matrix. This indicates that when the kernel size is 4, multiple physiological indicators are in the first stage. t Multiple relative offsets corresponding to each time point The relative offset embeddings represent the convolution result of the matrix. This indicates that when the kernel size is 5, multiple physiological indicators are in the first... t Multiple relative offsets corresponding to each time point The convolution result of the relative offset embedding representation matrix is ​​formed.

[0061] The recorrection module calculates the... The weights are redistributed to obtain the relative offset embedding representation.

[0062] Specifically, through the formula Calculate the first t The relative offset embedding representation of the relative offset at each time point .

[0063] in, and Both represent nonlinear activation functions. and All are parameter matrices.

[0064] It is worth mentioning that the relative offsets obtained through the above steps can intuitively and accurately reflect the stability of physiological indicators. However, the mathematical form of the relative offsets is not conducive to subsequent calculations. Therefore, multiple absolute offsets corresponding to multiple physiological indicators are calculated to obtain the relative offset embedding representation, which facilitates subsequent calculations.

[0065] The specific steps of step 14 described above will be illustrated below with reference to specific embodiments.

[0066] In one embodiment of this application, the specific implementation process of step 14 above includes the following steps: The first step involves using the recalibration module to calculate the time-series data of the patient's electronic medical record, obtaining the result at the [missing information - likely a specific timeframe or date]. t Embedded representation of electronic medical record time series data at each time point Specifically, the calculation formula is as follows: ;

[0067] in, and Both represent nonlinear activation functions. and All are parameter matrices. This indicates that multiple physiological indicators of the target patient are in the first... The value at each moment.

[0068] The second step is to use the LSTM model. Feature extraction is performed on the absolute offset embedding representation to obtain the absolute offset features. ,and .

[0069] It should be noted that the absolute offset feature The absolute offset embedding represents the current time. T The guidance indicates .

[0070] in, The absolute offset embedding is represented in the first position. t Guidance at each moment. The absolute offset embedding is represented in t- The guidance at moment 1, t =1,2,..., T , T Indicates the current moment. Indicates the first t The absolute offset embedding representation of the absolute offset at each time step, and LSTM() represents the LSTM model operation. This indicates a splicing operation.

[0071] It is worth mentioning that the absolute offset is a quantification of the degree of abnormality of physiological indicators. The longer the abnormal period, the worse the patient's condition. By using the LSTM model to extract features from the absolute offset embedding representation, we can capture the temporal changes of the patient's condition and highlight the abnormal time points in the patient's electronic medical record time series data. The obtained absolute offset features have high accuracy.

[0072] The specific steps of step 15 described above will be illustrated below with reference to specific embodiments.

[0073] In one embodiment of this application, the specific implementation process of step 15 above includes the following steps: Step 15.1: Calculate the weights of multi-head self-attention using the formula. : ; in, Indicates the first The height in Weight at each moment, =1,2,...,G , G This indicates the total number of heads in the bulls' self-attention. ( ) represents the activation function. , Indicates the first The height in The standard self-attention query vector at time step 1. t =1,2,..., T , T Indicates the current moment. Represents the parameter matrix, Indicates the first The relative offset and electronic medical record time series data at each moment, , Represents the parameter matrix, Indicates the first t The relative offset embedding representation of the relative offset at each time point. Indicates the first Original patient data at any given time. This indicates a splicing operation. , Indicates the first The standard self-attention key vector of the head at time 1. Represents the parameter matrix, , This represents the relative offset at time 1 and the electronic medical record time series data. This represents the relative offset embedding representation at time 1. This represents the hidden representation of medical features at time 1. , Indicates the first Size at the current moment T The standard self-attention key vector, , Indicates the current moment T The relative offset and electronic medical record time series data, Indicates the current moment T The relative offset of the relative offset is embedded in the representation. This represents the patient's primary representation vector. S This represents the total number of dimensions of the vector.

[0074] Step 15.2, using the formula: ; Calculate the self-attention representation of relative offset ; in, Indicates the first The relative offset at each moment is represented by the self-attention method. This indicates that the first head is in the first position. Weight at each moment This indicates that the first head is in the first position. The standard self-attention value vector at each time step , Represents the parameter matrix, Indicates the first G The height in Weight at each moment, Indicates the first G The height in The standard self-attention value vector at each time step , Represents the parameter matrix, This indicates the transpose operation. This indicates a splicing operation.

[0075] Step 15.3, using the formula: ; Calculate relative offset features .

[0076] It is worth noting that, due to the irregular sampling intervals of each physiological indicator, data preprocessing requires filling unsampled blank time points with fixed values ​​(i.e., the latest value or average value) to maintain a fixed time window size. This can lead to a large number of consecutive zero values ​​in the relative offset between two non-zero values. Therefore, it is necessary to capture the correlation between two non-adjacent and non-zero offsets with a long time interval, rather than simply focusing on the relationship between offsets at adjacent time points. However, traditional recurrent neural networks cannot directly capture the correlation between non-adjacent time points, and the continuous accumulation of zero values ​​in the memory cells of recurrent neural networks may gradually dilute the key features in the time series data. Therefore, a multi-head self-attention model is used to extract features from the relative offset. This model can simultaneously integrate feature information from all time points to enhance the learning ability of long-dependent features. In addition, the relative offset utilizes hidden representations as positional embeddings and supplementary information to ensure that the multi-head self-attention model can further aggregate the temporal features in the original data, improving the accuracy of the relative offset features.

[0077] The specific steps of step 16 described above will be illustrated below with reference to specific embodiments.

[0078] In one embodiment of this application, the specific implementation process of step 16 above includes the following steps: Step 16.1: Calculate the absolute offset feature weights using the patient's primary representation vector.

[0079] Specifically, through the formula: ; Calculate the first The importance of the absolute offset at a given moment to the patient's final health status ; in, This represents the patient's primary representation vector. Indicates a linear layer. L Indicates the size of the vector dimension. ( ) represents the activation function. Indicates the first The absolute offset embedding representation at each time step. t =1,2,..., T , T Indicates the current moment; Through formula Contextual features for calculating absolute offset ; Through formula Calculate the absolute offset feature weights .

[0080] in, This indicates a linear layer.

[0081] Step 16.2: Calculate the relative offset feature weights using the patient's primary representation vector.

[0082] Specifically, through the formula: ; Calculate the first The importance of information about the relative offset at a given moment to the patient's final health status ; in, Indicates a linear layer. Indicates the first The relative offset embedding representation at each time step; Through formula Contextual features for calculating relative offsets ; Through formula Calculate the relative offset feature weights ; in, This indicates a linear layer.

[0083] Step 16.3: The absolute offset feature and the relative offset feature are fused by the absolute offset feature weight and the relative offset feature weight to obtain the final patient representation vector.

[0084] Specifically, through the formula: ; Calculate the final patient representation vector .

[0085] in, Indicates the absolute offset feature. This indicates the relative offset feature.

[0086] It is worth mentioning that by using the patient's primary representation vector to calculate the absolute and relative offset features, the relationship between the absolute and relative offset features can be fully explored. By fusing the two based on the relationship between the absolute and relative offset features, a personalized and highly accurate final patient representation vector can be constructed.

[0087] The above-mentioned clinical risk assessment method for patients will be illustrated by a specific example below.

[0088] like Figure 2 As shown in the figure, the workflow of the patient clinical risk assessment system provided in a specific example of this application is as follows: Time-series data (i.e., the time-series data of the patient's electronic medical record mentioned above) enters the time-series embedding module. Through the calculation of the time-series model, a primary patient representation vector is obtained, which then enters the adaptive module. Simultaneously, the time-series data enters the absolute offset representation module. After passing through the absolute offset extractor, the absolute offset is obtained. The absolute offset is then processed by a one-dimensional convolutional neural network and a re-correction module to obtain an absolute offset embedding representation. The absolute offset embedding representation is then processed by an LSTM to obtain absolute offset features. This absolute offset embedding representation enters the adaptive module and is calculated together with the primary patient representation vector to obtain absolute offset feature weights. The time-series data enters the relative offset representation module. After passing through the relative offset extractor, the relative offset is obtained. The relative offset is then processed by a one-dimensional convolutional neural network and a re-correction module to obtain a relative offset embedding representation. The relative offset embedding representation is then processed by multi-head self-attention to obtain relative offset features. This relative offset embedding representation enters the adaptive module and is calculated together with the primary patient representation vector to obtain relative offset feature weights. In the adaptive fusion module, the absolute offset feature weights, relative offset feature weights, absolute offset features, and relative offset features are calculated to obtain the patient's final representation vector. The patient's final representation vector is then passed to the prediction module, which passes through the multilayer perceptron in the prediction module to obtain the prediction result.

[0089] It is worth mentioning that the above system can effectively implement the patient clinical risk assessment method provided in this application and obtain assessment results.

[0090] The patient clinical risk assessment method provided in this application was experimentally validated based on the above system.

[0091] The MIMIC-III dataset is a large database recording health-related data of over 40,000 patients admitted to intensive care units (ICUs) in medical centers between 2001 and 2012. This experiment used standardized data from the MIMIC-III dataset to predict patients' in-hospital mortality risk. Using a 48-hour time window prior to patient admission to the ICU, 17 physiological indicators were collected from patients' real-time vital signs and laboratory test results, labeled as whether in-hospital mortality occurred. The dataset was randomly divided into training, validation, and test sets in a ratio of 60%:10%:30%.

[0092] The eICU-CRD dataset is a multicenter database containing health data from over 200,000 ICU inpatients across the United States between 2014 and 2015. This experiment uses the eICU-CRD data to predict mortality rates, employing a 24-hour time window prior to patient admission to the ICU. Fourteen physiological indicators were selected from the standardized data table of the eICU-CRD dataset, and samples with time series lengths less than 24 were filtered out. The label used was whether the patient experienced in-hospital mortality. The dataset was divided into training, validation, and test sets, with a ratio of 60%:10%:30%.

[0093] PARSE and each baseline model were evaluated using the harmonic mean of recall and precision (F1-Score), area under the characteristic curve (AUROC), area under the precision-recall curve (AUPRC), and the minimum value of sensitivity and precision, Min(Se, P+). For all four evaluation metrics, a higher value indicates better predictive performance.

[0094] F1-Score is the most direct evaluation metric for judging the performance of binary classification methods. AUROC and AUPRC are more informative evaluation metrics when dealing with imbalanced datasets in binary classification tasks. Min(Se, P+) is the maximum value of Min(sensitivity, precision) on the recall curve. The F1-Score is defined as follows: ; ; ; in , and These represent the number of true negatives, false negatives, and false positives in the confusion matrix with a threshold of 0.5.

[0095] In addition to the data acquisition and evaluation parameters mentioned above, this experiment also selected several advanced models as baseline models for comparison: Gated Recurrent Unit (GRU): learns long-term features from time-series data. RETAIN: mimics the actual diagnostic process of a physician by viewing electronic health record (EHR) data in reverse chronological order and uses two-level attention to generate patient representations. Dipole: uses a Bi-directional Long Short-Term Memory (Bi-LSTM) network to learn the temporal features of patient diagnostic data and introduces visit-level attention to extract time-dependent features. Transformer: introduces a multi-head self-attention mechanism to capture the relationships between time steps in time-series data. In this experiment, the encoder part and averaging operation of the Transformer were used to generate patient representations. ConCare: uses a GRU to model time-aware features for each clinical variable (including time-series and static data), and then applies a multi-head self-attention mechanism to aggregate all features. For fair comparison, this experiment used a version with only time-series data as input. AdaCare: Uses dilated convolutions with multi-scale receptive fields to extract local temporal clinical features and applies GRU to capture temporal dependencies between time steps. Greedy Randomized Adaptive Search Procedures (GRASP): Simultaneously considers the current patient's clinical information and auxiliary information extracted from similar patients using the K-Means algorithm to improve the accuracy of patient representation. In this experiment, based on the comparative experimental results in the original paper, ConCare is used as the backbone model for the temporal feature extraction module in GRASP.

[0096] To better verify the performance of this application, Figure 2 This study implemented three different versions of a patient clinical risk assessment system. Each version used a different time series model in the time series embedding module to obtain primary time series features and was named PARSE (GRU), PARSE (Trans.), and PARSE (AdaCare), respectively.

[0097] The results of this experiment are shown in the table below.

[0098]

[0099] It can be seen that the patient clinical risk assessment method provided in this application can make the patient clinical risk assessment sufficiently accurate.

[0100] The patient clinical risk assessment device provided in this application is described below as an example.

[0101] like Figure 3 As shown in the figure, this application embodiment provides a patient clinical risk assessment device 300, which includes: a vector generation module 301, a first acquisition module 302, a second acquisition module 303, a first feature extraction module 304, a second feature extraction module 305, a fusion module 306, and an assessment module 307.

[0102] The vector generation module 301 processes the electronic medical record time-series data of the target patient to generate a primary representation vector of the patient. The electronic medical record time-series data includes multiple physiological indicators of the target patient. T The value at each moment, T The time is the current time; The first acquisition module 302 is used to process the electronic medical record time-series data of the target patient to generate a primary representation vector of the patient; the electronic medical record time-series data includes: multiple physiological indicators of the target patient in... T The value at each moment, T The time is the current time; The second acquisition module 303 is used to acquire a relative offset that reflects the stability of the physiological indicators of the target patient, and to acquire a relative offset embedding representation of the relative offset. The first feature extraction module 304 is used to extract features from the absolute offset embedding representation to obtain absolute offset features. The second feature extraction module 305 is used to extract features from the relative offset embedding representation to obtain the relative offset features. The fusion module 306 is used to fuse the absolute offset features and the relative offset features using the patient's primary representation vector to obtain the patient's final representation vector; Assessment module 307 is used to calculate the patient's final representation vector using the clinical risk assessment module, to obtain the patient's current position. T Clinical risk outcomes.

[0103] Optionally, the vector generation module 301 includes a vector generation unit, which is used to generate vectors using formulas. The time-series data of the target patient's electronic medical record are processed to obtain the patient's primary representation vector. ; in, This represents a time series model that processes the electronic medical record time series data of the target patient. Indicates the first Medical feature concealment representation at a given moment, This indicates that multiple physiological indicators of the target patient are in the first... The value at each moment, t =1,2,..., T .

[0104] Optionally, the first acquisition module 302 includes a first calculation unit and a second calculation unit; The first calculation unit is used to calculate using formulas Calculate the physiological indicators used to reflect the target patient at the current moment. T The absolute deviation of the value from the normal value; in, Indicates the first [number] used to reflect the target patient i The physiological indicators at the first t The absolute deviation of the value at any given time from the normal value. i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T , T Indicates the current moment. Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i The normal values ​​corresponding to each physiological indicator This represents the inverse triangular nonlinear activation function; The second calculation unit is used to calculate using formulas Calculate the first t Absolute offset embedding representation of the absolute offset at each moment ; in, and Both represent nonlinear activation functions. and All are parameter matrices. Indicates multiple physiological indicators at the first t Multiple absolute offsets corresponding to each moment The absolute offset embedding represents the convolution result of the matrix.

[0105] Optionally, the second acquisition module 303 includes a first calculation unit and a second calculation unit. The first calculation unit is used to calculate using the formula: ; Calculate the relative offset used to reflect the stability of physiological indicators of the target patient; in, Indicates the first [number] used to reflect the target patient i The physiological indicators at the first t The relative offset of stability at each moment i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T , T Indicates the current moment. Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i A physiological indicator in t- The value at time 1, This represents the inverse triangular nonlinear activation function; The second calculation unit is used to calculate using formulas Calculate the first t The relative offset embedding representation of the relative offset at each time point ; in, and Both represent nonlinear activation functions. and All are parameter matrices. Indicates multiple physiological indicators at the first t Multiple relative offsets corresponding to each time point The convolution result of the relative offset embedding representation matrix is ​​formed.

[0106] Optionally, the first feature extraction module 304 includes a first feature extraction unit, which is used to utilize an LSTM model. Feature extraction is performed on the absolute offset embedding representation to obtain the absolute offset features. ,and ; in, The absolute offset embedding is represented in the first position. t Guidance at each moment. The absolute offset embedding is represented in t- The guidance at moment 1, The absolute offset is embedded at the current time. T The guidance indicates that t =1,2,..., T , Indicates the first t Absolute offset embedding representation of absolute offset at each time point Indicates the first t The embedded representation of the time-series electronic medical record data at each time point, LSTM() represents the LSTM model operation. This indicates a splicing operation.

[0107] Optionally, the second feature extraction module 305 includes a first computing unit, a second computing unit, and a third computing unit; The first calculation unit is used to calculate the weights of multi-head self-attention using a formula. : ; in, Indicates the first The height in Weight at each moment, =1,2,..., G , G This indicates the total number of heads in the bulls' self-attention. ( ) represents the activation function. , Indicates the first The height in The standard self-attention query vector at time step 1. t =1,2,..., T , T Indicates the current moment. Represents the parameter matrix, Indicates the first The relative offset and electronic medical record time series data at each moment, , Represents the parameter matrix, Indicates the first t The relative offset embedding representation of the relative offset at each time point. Indicates the first Original patient data at any given time. This indicates a splicing operation. , Indicates the first The standard self-attention key vector of the head at time 1. Represents the parameter matrix, , This represents the relative offset at time 1 and the electronic medical record time series data. Indicates the first Relative offset embedding representation of relative offset at one time point This represents the hidden representation of medical features at time 1. , Indicates the first Size at the current moment T The standard self-attention key vector, , Indicates the current moment T The relative offset and electronic medical record time series data, Indicates the current moment T The relative offset of the relative offset is embedded in the representation. This represents the patient's primary representation vector. S This represents the total number of dimensions of the vector; The second calculation unit is used to calculate using the formula: ; Calculate the self-attention representation of relative offset ; in, Indicates the first The relative offset at each moment is represented by the self-attention method. This indicates that the first head is in the first position. Weight at each moment, This indicates that the first head is in the first position. The standard self-attention value vector at each time step , Represents the parameter matrix, Indicates the first G The height in Weight at each moment, Indicates the first G The height in The standard self-attention value vector at each time step , Represents the parameter matrix, This indicates the transpose operation. Indicates a splicing operation; The third calculation unit is used to calculate using the formula: ; Calculate relative offset features .

[0108] Optionally, the fusion module 306 includes a fusion unit, wherein the fusion unit, The first calculation unit is used to calculate using the formula: ; Calculate the first The importance of the absolute offset at a given moment to the patient's final health status ; in, This represents the patient's primary representation vector. Indicates a linear layer. L Indicates the size of the vector dimension. ( ) represents the activation function. Indicates the first The absolute offset embedding representation of the absolute offset at each time point t =1,2,..., T , T Indicates the current moment; Through formula Contextual features for calculating absolute offset ; Through formula Calculate the absolute offset feature weights ; in, Indicates a linear layer; The second calculation unit is used to calculate using the formula: ; Calculate the first The importance of information about the relative offset at a given moment to the patient's final health status ; in, Indicates a linear layer. Indicates the first The relative offset embedding representation of the relative offset at each moment; Through formula Contextual features for calculating relative offsets ; Through formula Calculate the relative offset feature weights ; in, Indicates a linear layer; The third calculation unit is used to calculate using the formula: ; Calculate the final patient representation vector ; in, Indicates the absolute offset feature. This indicates the relative offset feature.

[0109] Optionally, the assessment module 307 includes an assessment unit, which is used to calculate the patient's final representation vector using the clinical risk assessment module to obtain the patient's current position. T Clinical risk outcomes.

[0110] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0111] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0112] like Figure 4 As shown, an embodiment of this application provides a terminal device, wherein the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 4 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0113] Specifically, when the processor D100 executes the computer program D102, it processes the time-series data of the target patient's electronic medical record to generate a primary representation vector of the patient; obtains an absolute offset reflecting the deviation of the target patient's physiological indicators from normal values, and obtains an absolute offset embedding representation of the absolute offset; obtains a relative offset reflecting the stability of the target patient's physiological indicators, and obtains a relative offset embedding representation of the relative offset; extracts features from the absolute offset embedding representation to obtain absolute offset features; extracts features from the relative offset embedding representation to obtain relative offset features; fuses the absolute offset features and relative offset features using the primary representation vector of the patient to obtain the final representation vector of the patient; and calculates the final representation vector of the patient using the clinical risk assessment module to obtain the patient's clinical risk result.

[0114] The electronic medical record time series data includes: multiple physiological indicators of the target patient at...T The value at each moment, T The timeframe refers to the current moment.

[0115] Optionally, when the processor D100 executes the computer program D102, it is further configured to implement the formula The time-series data of the target patient's electronic medical record are processed to obtain the patient's primary representation vector. .

[0116] in, This represents a time series model that processes the electronic medical record time series data of the target patient. Indicates the first Medical feature concealment representation at a given moment, This indicates that multiple physiological indicators of the target patient are in the first... The value at each moment, t =1,2,..., T .

[0117] Optionally, when the processor D100 executes the computer program D102, it is further configured to implement the formula Calculate the physiological indicators used to reflect the target patient at the current moment. T The absolute deviation of the value from the normal value; in, Indicates the first [number] used to reflect the target patient i The physiological indicators at the first t The absolute deviation of the value at any given time from the normal value. i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T , T Indicates the current moment. Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i The normal values ​​corresponding to each physiological indicator This represents the inverse triangular nonlinear activation function; Through formula Calculate the first Absolute offset embedding representation of the absolute offset at each moment ; in, and Both represent nonlinear activation functions. and All are parameter matrices. Indicates multiple physiological indicators at the first t Multiple absolute offsets corresponding to each moment The absolute offset embedding represents the convolution result of the matrix.

[0118] Optionally, when the processor D100 executes the computer program D102, it is also used to implement the formula: ; Calculate the relative offset used to reflect the stability of physiological indicators of the target patient; in, Indicates the first [number] used to reflect the target patient i The physiological indicators at the first t The relative offset of stability at each moment i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T , T Indicates the current moment. Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i A physiological indicator in t- The value at time 1, This represents the inverse triangular nonlinear activation function; Through formula Calculate the first t The relative offset embedding representation of the relative offset at each time point ; in, and Both represent nonlinear activation functions. and All are parameter matrices. Indicates multiple physiological indicators at the first t Multiple relative offsets corresponding to each time point The convolution result of the relative offset embedding representation matrix is ​​formed.

[0119] Optionally, when the processor D100 executes the computer program D102, it is also used to implement the use of the LSTM model. Feature extraction is performed on the absolute offset embedding representation to obtain the absolute offset features. ,and ; in, The absolute offset embedding is represented in the first position. t Guidance at each moment. The absolute offset embedding is represented in t- The guidance at moment 1, The absolute offset is embedded at the current time. T The guidance indicates that t =1,2,..., T , Indicates the first t Absolute offset embedding representation of absolute offset at each time point Indicates the first t The embedded representation of the time-series electronic medical record data at each time point, LSTM() represents the LSTM model operation. This indicates a splicing operation.

[0120] Optionally, when the processor D100 executes the computer program D102, it is also configured to calculate the weights of multi-head self-attention using a formula. : ; in, Indicates the first The height in Weight at each moment, =1,2,..., G , G This indicates the total number of heads in the bulls' self-attention. ( ) represents the activation function. , Indicates the first The height in The standard self-attention query vector at time step 1. t =1,2,..., T , T Indicates the current moment. Represents the parameter matrix, Indicates the first The relative offset and electronic medical record time series data at each moment, , Represents the parameter matrix, Indicates the first t The relative offset embedding representation of the relative offset at each time point. Indicates the first Original patient data at any given time. This indicates a splicing operation. , Indicates the first The standard self-attention key vector of the head at time 1. Represents the parameter matrix, , This represents the relative offset at time 1 and the electronic medical record time series data. This represents the relative offset embedding representation at time 1. This represents the hidden representation of medical features at time 1. , Indicates the first Size at the current moment T The standard self-attention key vector, , Indicates the current moment T The relative offset and electronic medical record time series data, Indicates the current moment T The relative offset of the relative offset is embedded in the representation. This represents the patient's primary representation vector. S This represents the total number of dimensions of the vector; Through the formula: ; Calculate the self-attention representation of relative offset ; in, Indicates the first The relative offset at each moment is represented by the self-attention method. This indicates that the first head is in the first position. Weight at each moment, This indicates that the first head is in the first position. The standard self-attention value vector at each time step , Represents the parameter matrix, Indicates the first G The height in Weight at each moment, Indicates the first G The height in The standard self-attention value vector at each time step , Represents the parameter matrix, This indicates the transpose operation. Indicates a splicing operation; Through the formula: ; Calculate relative offset features .

[0121] Optionally, when the processor D100 executes the computer program D102, it is also used to implement the formula: ; Calculate the first The importance of the absolute offset at a given moment to the patient's final health status ; in, This represents the patient's primary representation vector. Indicates a linear layer. L Indicates the size of the vector dimension. ( ) represents the activation function. Indicates the first The absolute offset embedding representation of the absolute offset at each time point. t =1,2,..., T , T Indicates the current moment; Through formula Contextual features for calculating absolute offset ; Through formula Calculate the absolute offset feature weights ; in, Indicates a linear layer; Through the formula: ; Calculate the first The importance of information about the relative offset at a given moment to the patient's final health status ; in, Indicates a linear layer. Indicates the first The relative offset embedding representation of the relative offset at each moment; Through formula Contextual features for calculating relative offsets ; Through formula Calculate the relative offset feature weights ; in, Indicates a linear layer; Through the formula: ; Calculate the final patient representation vector ; in, Indicates the absolute offset feature. This indicates the relative offset feature.

[0122] It is worth mentioning that, since the absolute offset, which reflects the degree of deviation of the target patient's physiological indicators from the normal value, and the relative offset, which reflects the stability of the target patient's physiological indicators, have high directness and accuracy in representing the patient, the final patient representation vector is obtained by fusing the absolute offset feature and the relative offset feature using the primary patient representation vector. This feature fusion method can generate personalized patient representation vectors, improve the accuracy of the representation vectors, and enable sufficiently accurate clinical risk assessment of patients.

[0123] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0124] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0125] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0126] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0127] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a patient clinical risk assessment method device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0128] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0129] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for assessing clinical risk in patients, characterized in that, include: The electronic medical record time series data of the target patient is processed to generate a primary representation vector of the patient; The electronic medical record time-series data includes: multiple physiological indicators of the target patient at... T The value at time i, the i-th time T The current moment is represented by the current moment. Obtain the absolute offset that reflects the degree of deviation between the physiological index value of the target patient and the normal value, and obtain the absolute offset embedding representation of the absolute offset; Obtain a relative offset that reflects the stability of the physiological indicators of the target patient, and obtain a relative offset embedding representation of the relative offset; Feature extraction is performed on the absolute offset embedding representation to obtain absolute offset features; Feature extraction is performed on the relative offset embedding representation to obtain the relative offset features; The absolute offset feature and the relative offset feature are fused using the patient's primary representation vector to obtain the patient's final representation vector; The final representation vector of the patient is calculated using the clinical risk assessment module to obtain the patient's status at the current time. T Clinical risk outcomes; The step of obtaining an absolute offset that reflects the degree of deviation between the physiological index value of the target patient and the normal value, and obtaining an absolute offset embedding representation of the absolute offset, includes: Through formula Calculate the absolute deviation of the physiological indicator value used to reflect the degree of deviation between the target patient's value and the normal value; in, Indicates the first [number] used to reflect the target patient i The physiological indicators at the first t The absolute deviation of the value at any given time from the normal value. i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T ， Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i The normal values ​​corresponding to each physiological indicator This represents the inverse triangular nonlinear activation function; Through formula Calculate in the first t The absolute offset embedding representation of the absolute offset at each moment ; in, and Both represent nonlinear activation functions. and All are parameter matrices. This indicates that the multiple physiological indicators are in the first... t Multiple absolute offsets corresponding to each moment The convolution result of the absolute offset embedding representation matrix; The step of obtaining a relative offset reflecting the stability of the target patient's physiological indicators and obtaining a relative offset embedding representation of the relative offset includes: Through the formula: Calculate the relative offset of the physiological indicators used to reflect the stability of the target patient; in, Indicates the first [number] used to reflect the target patient i The physiological indicators at the first t The relative offset of stability at each moment i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T , Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i The physiological indicators at the first t- The value at one moment, This represents the inverse triangular nonlinear activation function; Through formula Calculate in the first t The relative offset embedding representation of the relative offset at each moment ; in, and Both represent nonlinear activation functions. and All are parameter matrices. This indicates that the multiple physiological indicators are in the first... t Multiple relative offsets corresponding to each time point The convolution result of the relative offset embedding representation matrix is ​​formed.

2. The evaluation method according to claim 1, characterized in that, The process of processing the time-series data of the target patient's electronic medical record to generate a primary representation vector of the patient includes: Through formula The time-series data of the target patient's electronic medical record are processed to obtain the patient's primary representation vector. ; in, This represents a time series model for processing the electronic medical record time series data of the target patient. Indicates the first The medical feature hiding representation at a given moment, This indicates that multiple physiological indicators of the target patient are in the first... The value at each moment, t =1,2,..., T .

3. The evaluation method according to claim 1, characterized in that, The feature extraction from the absolute offset embedding representation to obtain the absolute offset features includes: Using LSTM model Feature extraction is performed on the absolute offset embedding representation to obtain the absolute offset features. ,and ; in, The absolute offset is embedded in t The guidance of time. The absolute offset embedding is represented in t- The guidance at moment 1, The absolute offset is embedded at the current time. T The guidance indicates that t =1,2,..., T , Indicates the first t The absolute offset of the absolute offset at each moment is embedded in the representation. Indicates the first t The embedded representation of the electronic medical record time series data at each time point, LSTM() represents the LSTM model operation. This indicates a splicing operation.

4. The evaluation method according to claim 1, characterized in that, The step of extracting features from the relative offset embedding representation to obtain relative offset features includes: Feature extraction is performed on the relative offset embedding representation using a multi-head self-attention model; The weight of multi-head self-attention is calculated using a formula. : in, Indicates the first Height Weight of time, =1,2,..., G , G This indicates the total number of heads in the bulls' self-attention. ( ) represents the activation function. , Indicates the first Height The standard self-attention query vector at time step. t =1,2,..., T , T Indicates the current moment. Represents the parameter matrix, Indicates the first The relative offset at each moment and the electronic medical record time series data, , Represents the parameter matrix, Indicates the first t The relative offset embedding representation of the relative offset at each time point. Indicates the first Original patient data at any given time. This indicates a splicing operation. , Indicates the first The height in The standard self-attention key vector at one time step. Represents the parameter matrix, , Indicates the first The relative offset at one moment and the electronic medical record time series data, This represents the relative offset embedding representation of the relative offset at time 1. This represents the hidden representation of medical features at time 1. , Indicates the first Size at the current moment T The standard self-attention key vector, , Indicates the current moment T The relative offset and the electronic medical record time series data, Indicates the current moment T The relative offset of the relative offset is embedded in the representation. This represents the primary representation vector of the patient. S This represents the total number of dimensions of the vector; Through the formula: Calculate the self-attention representation of relative offset ; in, Indicates the first The relative offset at each moment is represented by the self-attention method. This indicates that the first head is in the first position. Weight at each moment, This indicates that the first head is in the first position. The standard self-attention value vector at each time step , Represents the parameter matrix, Indicates the first G The height in Weight at each moment, Indicates the first G The height in The standard self-attention value vector at each time step , Represents the parameter matrix, This indicates the transpose operation. Indicates a splicing operation; Through the formula: Calculate relative offset features .

5. The evaluation method according to claim 1, characterized in that, The process of fusing the absolute offset feature and the relative offset feature using the patient's primary representation vector to obtain the patient's final representation vector includes: Through the formula: Calculate in the first The importance of the absolute offset information at each moment to the patient's final health status ; in, This represents the primary representation vector of the patient. Indicates a linear layer. L Indicates the size of the vector dimension. ( ) represents the activation function. Indicates the first t The absolute offset embedding representation of the absolute offset at each time point. t =1,2,..., T , T Indicates the current moment; Through formula Contextual features for calculating the absolute offset ; Through formula Calculate the absolute offset feature weights ; in, Indicates a linear layer; Through the formula: Calculate in the first The importance of the relative offset information at each moment to the patient's final health status ; in, Indicates a linear layer. Indicates the first t The relative offset of the relative offset at each moment is embedded in the relative offset representation; Through formula Contextual features for calculating the relative offset ; Through formula Calculate the relative offset feature weights ; in, Indicates a linear layer; Through the formula: Calculate the final representation vector of the patient ; in, This represents the absolute offset feature. This represents the relative offset feature.

6. A device for assessing patient clinical risk based on time-series data, characterized in that, include: The vector generation module is used to process the time-series data of the target patient's electronic medical record and generate a primary representation vector of the patient; Electronic medical record time-series data includes: multiple physiological indicators of the target patient at... T The value at time i, the i-th time T The current moment is represented by the current moment. The first acquisition module is used to acquire an absolute offset that reflects the degree of deviation between the physiological index value of the target patient and the normal value, and to acquire an absolute offset embedding representation of the absolute offset. The second acquisition module is used to acquire a relative offset that reflects the stability of the physiological indicators of the target patient, and to acquire a relative offset embedding representation of the relative offset. The first feature extraction module is used to extract features from the absolute offset embedding representation to obtain absolute offset features; The second feature extraction module is used to extract features from the relative offset embedding representation to obtain relative offset features; The fusion module uses the patient's primary representation vector to fuse the absolute offset feature and the relative offset feature to obtain the patient's final representation vector; The assessment module uses the clinical risk assessment module to calculate the patient's final representation vector, obtaining the patient's status at the current moment. T Clinical risk outcomes; The first acquisition module is specifically used for: obtaining formulas Calculate the absolute deviation of the physiological indicator value used to reflect the degree of deviation from the normal value of the target patient; using the formula Calculate in the first t The absolute offset embedding representation of the absolute offset at each moment ;in, Indicates the first [number] used to reflect the target patient i The physiological indicators at the first t The absolute deviation of the value at any given time from the normal value. i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i The normal values ​​corresponding to each physiological indicator This represents the inverse triangular nonlinear activation function; and Both represent nonlinear activation functions. and All are parameter matrices. This indicates that the multiple physiological indicators are in the first... t Multiple absolute offsets corresponding to each moment The convolution result of the absolute offset embedding representation matrix; The second acquisition module is specifically used for: Through the formula: Calculate the relative offset of the physiological indicators used to reflect the stability of the target patient; in, Indicates the first [number] used to reflect the target patient i The physiological indicators at the first t The relative offset of stability at each moment i =1,2,..., C , C This represents the total number of physiological indicators. t= 1,2,..., T , Indicates the first i The physiological indicators at the first t The value at each moment, Indicates the first i The physiological indicators at the first t- The value at one moment, This represents the inverse triangular nonlinear activation function; Through formula Calculate in the first t The relative offset embedding representation of the relative offset at each moment ; in, and Both represent nonlinear activation functions. and All are parameter matrices. This indicates that the multiple physiological indicators are in the first... t Multiple relative offsets corresponding to each time point The convolution result of the relative offset embedding representation matrix is ​​formed.

7. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the patient clinical risk assessment method as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the patient clinical risk assessment method as described in any one of claims 1 to 5.

Images