An acute kidney injury prediction system

Abstract

The application discloses an acute kidney injury prediction system, a data preprocessing module in the system is used for acquiring to-be-tested structured medical record data of a target object, determining at least two to-be-tested time structured data based on a time sliding window and the to-be-tested structured medical record data, and classifying the to-be-tested structured medical record data based on an index parameter dimension to obtain at least two to-be-tested space structured data, wherein there is at least one same to-be-tested time medical record data between adjacent to-be-tested time structured data; a model processing module is used for outputting a predicted acute kidney injury result based on each to-be-tested time structured data, each to-be-tested space structured data and a pre-trained acute kidney injury prediction model; and a prediction result output module is used for outputting the predicted acute kidney injury result. The embodiment of the application improves the prediction effect of the acute kidney injury prediction system.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to an acute kidney injury prediction system. Background Technology

[0002] Acute kidney injury (AKI) is characterized by its high incidence, high mortality, high rate of adverse clinical events, and poor prognosis. Clinical studies have shown that AKI is a common complication of sepsis, with sepsis-induced AKI (SI-AKI) accounting for more than 50% of AKI cases and being a significant cause of death and poor prognosis in intensive care unit (ICU) patients. However, the pathogenesis of SI-AKI is highly complex, with various theories circulating. The unclear pathogenesis and reliance on clinical experience make early diagnosis of SI-AKI a significant challenge.

[0003] In recent years, deep learning has been increasingly applied to electronic health record (EHR) systems to improve the prediction and diagnosis of patients' clinical manifestations, including early-stage disease detection, clinical risk assessment, and prediction of future hospitalization.

[0004] Deep learning has many different applicable algorithms depending on the research direction. Neural network models aimed at SI-AKI prediction and evaluation still have shortcomings, with relatively lagging prediction results and low accuracy. Summary of the Invention

[0005] This invention provides an acute kidney injury prediction system to address the problem of poor prediction performance of existing acute kidney injury prediction models and improve the timeliness and accuracy of acute kidney injury prediction results.

[0006] An acute kidney injury prediction system is provided according to an embodiment of the present invention. The system includes: a data preprocessing module, a model processing module, and a prediction result output module.

[0007] The data preprocessing module is used to acquire the structured medical record data to be tested of the target object, and to determine at least two structured time data to be tested based on the time sliding window and the structured medical record data to be tested, and to classify the structured medical record data to be tested based on the indicator parameter dimension to obtain at least two structured spatial data to be tested; wherein, there is at least one identical structured time data to be tested between adjacent structured time data to be tested.

[0008] The model processing module is used to output the predicted acute kidney injury result based on each of the time-structured data to be tested, each of the spatial-structured data to be tested, and the pre-trained acute kidney injury prediction model.

[0009] The prediction result output module is used to output the predicted acute kidney injury result.

[0010] The technical solution of this invention includes a data preprocessing module in an acute kidney injury (AKI) prediction system. This module acquires the target object's structured medical record data and determines at least two target temporal structured data based on a time sliding window and the target structured medical record data. It also classifies the target structured medical record data based on indicator parameters to obtain at least two target spatial structured data. Adjacent target temporal structured data contain at least one identical target temporal medical record data. A model processing module in the AKI prediction system outputs a predicted AKI result based on each target temporal structured data, each target spatial structured data, and a pre-trained AKI prediction model. A prediction result output module in the AKI prediction system outputs the predicted AKI result. This AKI prediction system considers the temporal continuity and spatial diversity of AKI, reduces the impact of data acquisition bias in structured medical record data on the AKI prediction result, solves the problem of poor prediction performance in existing AKI prediction models, and improves the timeliness and accuracy of the AKI prediction results output by the AKI prediction model.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the structure of an acute kidney injury prediction system provided in one embodiment of the present invention;

[0014] Figure 2 This is a time series diagram of structured medical record data to be tested, provided in one embodiment of the present invention.

[0015] Figure 3This is a schematic diagram of the structure of another acute kidney injury prediction system provided in one embodiment of the present invention;

[0016] Figure 4 This is a model architecture diagram of an acute kidney injury prediction model provided in one embodiment of the present invention;

[0017] Figure 5 A model architecture diagram of another acute kidney injury prediction model provided in one embodiment of the present invention;

[0018] Figure 6 This is a schematic diagram of the structure of another acute kidney injury prediction system provided in one embodiment of the present invention;

[0019] Figure 7 This is a time series diagram illustrating a training method for structured medical record data, provided in one embodiment of the present invention.

[0020] Figure 8 A flowchart illustrating the training process of an acute kidney injury prediction model provided in one embodiment of the present invention;

[0021] Figure 9 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] Figure 1This is a schematic diagram of an acute kidney injury prediction system provided in one embodiment of the present invention. The embodiment of the present invention is applicable to the early prediction of acute kidney injury in target subjects, especially to the early prediction of acute kidney injury in target subjects confirmed to have sepsis. The acute kidney injury prediction system can be implemented in hardware and / or software and can be configured in a terminal device.

[0025] like Figure 1 As shown, the acute kidney injury prediction system includes: a data preprocessing module 110, a model processing module 120, and a prediction result output module 130. The data preprocessing module 110 is used to acquire the target object's structured medical record data, and based on a time sliding window and the target structured medical record data, determine at least two target time-structured data sets, and classify the target structured medical record data sets based on indicator parameter dimensions to obtain at least two target spatial-structured data sets. The model processing module 120 is used to output the predicted acute kidney injury result based on each target time-structured data set, each target spatial-structured data set, and a pre-trained acute kidney injury prediction model. The prediction result output module 130 is used to output the predicted acute kidney injury result.

[0026] In this embodiment, the input data of the acute kidney injury prediction system includes the structured medical record data to be tested. Specifically, the structured medical record data to be tested can be used to characterize the medical records data that have been structurally analyzed from the perspective of medical informatics, using natural language to represent medical documents, according to the requirements of medical terminology, and whose semantic structures are stored in a database in a relational (object-oriented) structure.

[0027] In this embodiment, the time sliding window represents a preset number of time-related medical record data contained in the time-related structured data to be tested. For example, the preset number can be 3 or 7, and the sliding step size of the time sliding window can be 1 time node or 2 time nodes. Here, the preset number and sliding step size are not limited.

[0028] In this embodiment, there is at least one identical medical record data point among adjacent structured data points of the time to be tested. Specifically, the number of identical medical record data points among adjacent structured data points of the time to be tested is greater than or equal to the sliding step size.

[0029] In one optional embodiment, the structured medical record data to be tested includes medical record data corresponding to at least two time nodes, and the time length corresponding to each time node satisfies a preset time step. Accordingly, the data preprocessing module 110 is specifically used to: for each time node, take the current time node as the window termination node, and obtain reference time structured data from the structured medical record data to be tested based on the time sliding window and the window termination node; if the number of reference time medical record data in the reference time structured data is less than the preset number corresponding to the time sliding window, perform interpolation processing on the reference time structured data to obtain the test time structured data corresponding to the current time node.

[0030] For example, the preset time step is 1 hour. For instance, the structured medical record data to be tested includes medical record data 1 for the period from 8:00 to 9:00, medical record data 2 for the period from 9:00 to 10:00, and medical record data for the period from 13:00 to 14:00, etc.

[0031] For example, assuming the preset number of time sliding windows is N, if the sliding step of the time sliding window is 1, then the number of reference time structured data corresponding to the first N-1 time nodes is less than the preset number of time sliding windows.

[0032] In an optional embodiment, performing interpolation on the reference time structured data can involve copying the reference time medical record data within the reference time structured data, so that the number of medical record data to be tested in the current test time structured data meets the preset number corresponding to the time sliding window. For example, the reference time medical record data to be copied can be randomly selected, or it can be one of the previous or next few; there is no limitation on the reference time medical record data to be copied.

[0033] In one alternative embodiment, the start time point in the structured medical record data to be tested is the first time point after the target subject is diagnosed with sepsis.

[0034] The advantage of this setup is that it allows the acute kidney injury prediction system to be specifically used for early prediction of acute kidney injury in target individuals confirmed with sepsis.

[0035] Figure 2 This is a time series diagram illustrating a structured medical record data to be tested, provided as an embodiment of the present invention. Specifically, Figure 2 The dots within the square boxes represent the time-specific medical records in the structured medical record data to be tested. Figure 2The structured medical record data to be tested shown includes medical record data corresponding to 12 time nodes. Among them, time node 1 is the first time node after the target subject is diagnosed with sepsis. Figure 2 The arrows in the timeline indicate the direction of the timeline, and the position of the dots indicates the time position of each time node on the timeline. For example, time node 1 is located at 8:00-9:00, time node 2 is located at 9:00-10:00, time node 3 is located at 10:00-11:00, time node 4 is located at 12:00-13:00, and so on.

[0036] by Figure 2 For example, if the preset number of time sliding windows is 6 and the sliding step size is 1 time node, then the structured data of the time to be tested corresponding to time node 1 is [1 1 1 1 1 1], the structured data of the time to be tested corresponding to time node 2 is [1 2 1 2 1 2], the structured data of the time to be tested corresponding to time node 3 is [1 2 3 1 2 3], the structured data of the time to be tested corresponding to time node 4 is [1 2 3 4 1 2], the structured data of the time to be tested corresponding to time node 5 is [1 2 3 4 5 1], the structured data of the time to be tested corresponding to time node 6 is [1 2 3 4 5 6], the structured data of the time to be tested corresponding to time node 7 is [2 3 4 5 6 7], and so on.

[0037] Based on the above embodiments, optionally, the data preprocessing module 110 is specifically used for: obtaining at least two original time-based medical record data corresponding to the original structured medical record data of the target object based on a preset time step; performing a first preprocessing operation on each original time-based medical record data to obtain the time-based medical record data to be tested; wherein, the first preprocessing operation includes at least one of aggregation operation, outlier handling operation, standardization operation, and missing value handling operation; and generating the structured medical record data to be tested corresponding to the target object based on each time-based medical record data to be tested.

[0038] Specifically, the raw structured medical record data can be used to characterize the unprocessed structured medical record data corresponding to the target object. The raw structured medical record data contains structured medical record data corresponding to each time point.

[0039] For example, assuming a preset time step of 1 hour, the structured medical record data within the time ranges of 8:00-9:00, 9:00-10:00, 10:00-11:00, and 11:00-12:00 from the original structured medical record data will be used as the original time-based medical record data. For instance, if the original structured medical record data does not contain structured medical record data within the 10:00-11:00 time range, then the time nodes in the structured medical record data to be tested will not include 10:00-11:00.

[0040] In one embodiment, performing a first preprocessing operation on the original time-based medical record data to obtain the time-based medical record data to be tested includes: performing an aggregation operation on the original time-based medical record data based on the indicator parameter dimension to obtain the time-based medical record data to be tested.

[0041] For example, the aggregation operation can be mean aggregation, maximum aggregation, minimum aggregation, or median aggregation, etc. For instance, if the original time-based medical record data includes two sets of time-based medical record data, where the red blood cell count in the first set is 450 / ml and the red blood cell count in the second set is 550 / ml, then the red blood cell count in the time-based medical record data to be tested can be 450 / ml, 450 / ml, or 500 / ml.

[0042] The advantage of this setup is that it reduces the amount of data input into the model processing module 120, thereby improving the prediction efficiency of the acute kidney injury prediction system.

[0043] In another embodiment, a first preprocessing operation is performed on the original time-based medical record data to obtain the time-based medical record data to be tested, including: determining the box plots corresponding to each indicator parameter based on the indicator parameter dimensions and the original time-based medical record data, and determining the parameter values ​​corresponding to each indicator parameter in the time-based medical record data to be tested based on each box plot.

[0044] A box plot is a statistical graph used to display the dispersion characteristics of a set of data, showing the maximum, minimum, median, and upper and lower quartiles. Specifically, for each box plot, parameter values ​​in the original time-series medical records that are greater than the maximum value in the box plot are designated as first outliers, and parameter values ​​in the original time-series medical records that are less than the minimum value in the box plot are designated as second outliers. Each first outlier in the original time-series medical records is replaced with the maximum value in the corresponding box plot, and each second outlier in the original time-series medical records is replaced with the minimum value in the corresponding box plot, thus obtaining the time-series medical records to be tested.

[0045] The advantage of this setting is that it can reduce the impact of outliers in the original time-based medical record data on the acute kidney injury prediction system, thereby improving the accuracy of the prediction results of the acute kidney injury prediction system.

[0046] In another embodiment, performing a first preprocessing operation on the raw time-based medical record data to obtain the time-based medical record data to be tested includes: performing a standardization operation on the raw time-based medical record data to obtain the time-based medical record data to be tested. For example, the standardization function used in the standardization operation can be a deviation standardization function or a Z-score standardization function.

[0047] The advantage of this setting is that it can unify the dimensions of various indicator parameters in the medical record data of the time to be tested, reduce the impact of different dimensions on the accuracy of the prediction results of the acute kidney injury prediction system, and improve the accuracy of the prediction results of the acute kidney injury prediction system.

[0048] In another embodiment, a first preprocessing operation is performed on the original time medical record data to obtain the time medical record data to be tested, including: if there is missing data in the original time medical record data, using the original time medical record data as the current original time medical record data, and adding the parameter values ​​of the indicator parameters corresponding to the missing data in the previous or next original time medical record data to the current original time medical record data.

[0049] Based on the above embodiments, if neither the previous original time medical record data nor the next parameter time medical record data has a parameter value corresponding to the missing data, then the standard value of the indicator parameter corresponding to the missing data is added to the current original time medical record data.

[0050] The advantage of this setting is that it reduces the impact of missing data on the accuracy of the acute kidney injury prediction system's prediction results, thereby improving the overall accuracy of the acute kidney injury prediction system's prediction results.

[0051] For example, the network architecture of an acute kidney injury prediction model includes, but is not limited to, CNN (Convolutional Neural Networks), FCN (Fully Convolutional Networks), Residual Networks (ResNet), DNN (Deep Neural Networks), RNN (Recurrent Neural Networks), or Transformer networks, etc. No specific limitation is made to the network architecture of the acute kidney injury prediction model here.

[0052] In this embodiment, the model processing module 120 is specifically used to: input each test time-structured data and each test spatial-structured data into a pre-trained acute kidney injury prediction model to obtain the output predicted acute kidney injury result. For example, the predicted acute kidney injury result can be an occurrence probability, a non-occurrence probability, an occurrence grade, a non-occurrence grade, or whether it has occurred, etc. The occurrence probability and occurrence grade characterize the likelihood that the target subject will subsequently be diagnosed with acute kidney injury; a higher occurrence probability or occurrence grade indicates a higher likelihood of the target subject being diagnosed with acute kidney injury, such as an occurrence probability of 0.9 and an occurrence grade of I. The non-occurrence probability and non-occurrence grade characterize the likelihood that the target subject will not subsequently be diagnosed with acute kidney injury; a higher non-occurrence probability or non-occurrence grade indicates a lower likelihood of the target subject being diagnosed with acute kidney injury, such as a non-occurrence probability of 0.1 and an occurrence grade of III. The output content of the predicted acute kidney injury result is not limited here.

[0053] For example, the output formats of the prediction result output module 130 include, but are not limited to, text display, sound playback, and indicator light output. For instance, the text display could be "Occurrence probability is 0.9," and the sound playback could be voice playback or a prompt tone. If the occurrence probability is low, the volume and / or pitch of the prompt tone output by the prediction result output module 130 would be low; if the occurrence probability is high, the volume and / or pitch of the prompt tone output by the prediction result output module 130 would be high. The indicator light output would have different colors and / or frequencies. If the occurrence probability is low, the indicator light color output by the prediction result output module 130 would be green and / or have a low frequency; if the occurrence probability is high, the indicator light color output by the prediction result output module 130 would be red and / or have a high frequency. The output format of the prediction result output module 130 is not limited here.

[0054] The technical solution of this embodiment includes a data preprocessing module in the acute kidney injury prediction system. This module acquires the target object's structured medical record data and determines at least two target temporal structured data based on a time sliding window and the target structured medical record data. It also classifies the target structured medical record data based on indicator parameters to obtain at least two target spatial structured data. Adjacent target temporal structured data contain at least one identical target temporal medical record data. The model processing module in the acute kidney injury prediction system outputs a predicted acute kidney injury result based on each target temporal structured data, each target spatial structured data, and a pre-trained acute kidney injury prediction model. The prediction result output module in the acute kidney injury prediction system outputs the predicted acute kidney injury result from the model processing module. This embodiment of the acute kidney injury prediction system considers the temporal continuity and spatial diversity of acute kidney injury, reduces the impact of data acquisition bias in structured medical record data on the acute kidney injury prediction result, solves the problem of poor prediction performance in existing acute kidney injury prediction models, and improves the timeliness and accuracy of the acute kidney injury prediction results output by the acute kidney injury prediction model.

[0055] Figure 3 This is a schematic diagram of another acute kidney injury prediction system provided in one embodiment of the present invention. This embodiment further refines the acute kidney injury prediction model in the model processing module 120 of the above embodiment.

[0056] like Figure 3 As shown, the acute kidney injury prediction model in the model processing module 120 of the acute kidney injury prediction system includes a location encoding unit 121, a structured time calculation unit 122, a structured spatial calculation unit 123, and a spatiotemporal joint calculation unit 124. Correspondingly, the model processing module 120 is specifically used for: inputting each target time structured data and each target spatial structured data into the location encoding unit 121 to obtain the output target time structured features and target spatial structured features; inputting each target time structured feature into the structured time calculation unit 122 in the acute kidney injury prediction model to obtain the output structured time vector; inputting each target spatial structured feature into the structured spatial calculation unit 123 in the acute kidney injury prediction model to obtain the output structured spatial vector; and inputting the structured time vector and structured spatial vector into the spatiotemporal joint calculation unit 124 in the acute kidney injury prediction model to obtain the output predicted acute kidney injury result.

[0057] Specifically, the position encoding unit 121 is used to store the relative or absolute positions of each index parameter in the time-structured data and the spatial-structured data to be tested.

[0058] For example, the position encoding unit 121 satisfies the formula:

[0059] PE (pos，2i) =sin(pos / 10000) 2i / d )

[0060] PE (pos，2i+1) =cos(pos / 10000) 2i / d )

[0061] Where pos represents the position of the indicator parameter in the time-structured data or spatial-structured data to be tested, pos≤l, l represents the length of the time-structured data or spatial-structured data to be tested, d is the dimension of word embedding, 2i represents an even dimension, 2i+1 represents an odd dimension, where 2i≤d, 2i+1≤d.

[0062] In an optional embodiment, the structured time computation unit 122 performs feature extraction operations on each test time structured feature using a self-attention mechanism to obtain a structured time vector. Specifically, the structured time computation unit 122 includes a first masking self-attention layer and a first residual and layer normalization layer. The first masking self-attention layer is used to output each time attention feature based on the input test time structured features, and the first residual and layer normalization layer is used to output a structured time vector based on the input test time structured features and the time attention features output by the first masking self-attention layer.

[0063] For example, the mask self-attention layer satisfies the formula:

[0064]

[0065] Where Q, K, and V are calculated from each time-structured data to be measured using three different linear transformation matrices, d k Represents the vector dimension, divided by The purpose is to prevent the internal volume from becoming too large.

[0066] Furthermore, the multi-head attention mechanism can also be defined as:

[0067] MultiHead(Q,K,V)=Concat(head1,...,head h W O

[0068] head i =Attention(QW i Q KW i K VW iV )

[0069] Among them, W O W i Q W i K and W i V The multi-head attention mechanism, which represents the learnable parameter weight matrix, provides multiple representation subspaces for the attention layer, expanding the ability of the acute kidney injury prediction model to focus on different locations, thereby greatly improving the performance of the attention layer.

[0070] Specifically, the structured spatial computation unit 123 performs feature extraction operations on each test spatial structured feature to obtain a structured spatial vector. In an optional embodiment, the structured spatial computation unit 123 includes a one-dimensional convolutional layer, a fully connected layer, and a layer normalization layer. The one-dimensional convolutional layer outputs spatial convolutional features based on the input test spatial structured features; the fully connected layer outputs structured spatial features based on the spatial convolutional features output by the one-dimensional convolutional layer; and the layer normalization layer outputs a structured spatial vector based on the structured spatial features output by the fully connected layer.

[0071] In an optional embodiment, the structured time vector and structured space vector are fused by the spatiotemporal joint computing unit 124, and the predicted acute kidney injury result is determined based on the obtained structured fused vector. Specifically, the spatiotemporal joint computing unit 124 includes a fusion layer, a second masked multi-head self-attention layer, a second residual and layer normalization layer, a first feedforward neural network, a third residual and layer normalization layer, and an output layer.

[0072] The system comprises the following layers: a fusion layer, which fuses the structured temporal vector and the structured spatial vector to output a structured fusion vector; a second masked multi-head self-attention layer, which outputs fused attention features based on the structured fusion vector from the fusion layer; a second residual and layer normalization layer, which outputs a first fusion feature based on the structured fusion vector from the fusion layer and the fused attention features from the second masked multi-head self-attention layer; a first feedforward neural network, which outputs a first feedforward feature based on the first fusion feature from the second residual and layer normalization layer; a third residual and layer normalization layer, which outputs a second fusion feature based on the first fusion feature from the second residual and layer normalization layer and the first feedforward feature from the first feedforward neural network; and an output layer, which outputs a predicted acute kidney injury result based on the second fusion feature from the third residual and layer normalization layer. In an optional embodiment, the output layer includes a fully connected layer and an activation function.

[0073] For example, the first feedforward neural network can be defined as:

[0074] FFN(x) = max(0, xW1+b1)W2+b2

[0075] Where W1 and W2 are learnable parameter weight matrices, and b1 and b2 are pre-set bias values.

[0076] Figure 4 This is a model architecture diagram of an acute kidney injury prediction model provided in one embodiment of the present invention. Figure 4 The leftmost 3x3 grid represents the original structured medical record data. The original structured data is input into the data preprocessing module 110, which outputs the structured data of each time period and each spatial period to be tested. Figure 4 The "circle + curve" following the data preprocessing module 110 represents the location encoding unit 121 in the acute kidney injury prediction model. The location encoding unit 121 outputs N-dimensional time structured features and T-dimensional spatial structured features. The N-dimensional time structured features are input into the structured time calculation unit 122 in the acute kidney injury prediction model to obtain the output structured time vector. The T-dimensional spatial structured features are input into the structured space calculation unit 123 in the acute kidney injury prediction model to obtain the output structured space vector.

[0077] Figure 4 The "circle + plus sign" in the figure represents the fusion layer in the spatiotemporal joint computing unit 124. Specifically, the structured time vector output by the structured time computing unit 122 and the structured space vector output by the structured space computing unit 123 are input into the spatiotemporal joint computing unit 124 in the acute kidney injury prediction model to obtain the output predicted acute kidney injury result.

[0078] Based on the above embodiments, optionally, the data preprocessing module 110 is further configured to: perform a second preprocessing operation on the acquired raw unstructured medical record data to obtain unstructured medical record data to be tested; wherein, the second preprocessing operation includes field filtering operation and vector transformation operation; correspondingly, the acute kidney injury prediction model also includes an unstructured computing unit, and the model processing module 120 is further configured to: input the unstructured medical record data to be tested into the position encoding unit 121 in the pre-trained acute kidney injury prediction model to obtain the output unstructured features; input the unstructured features into the unstructured computing unit in the pre-trained acute kidney injury prediction model to obtain the output unstructured vector, and input the unstructured vector into the spatiotemporal joint computing unit 124 in the acute kidney injury prediction model.

[0079] Specifically, the unstructured medical record data to be tested can be used to characterize the medical record data stored in the database in text form. For example, the unstructured medical record data to be tested includes, but is not limited to, hospitalization records, progress notes, ICU consultations, medication records, and related health instructions, etc., but there is no limitation on the unstructured medical record data to be tested here.

[0080] Specifically, the data preprocessing module 110 is used to: remove fields and / or characters unrelated to acute kidney injury from the original unstructured medical record data; input the removed original unstructured medical record data into a pre-trained word embedding model to obtain the output unstructured medical record data to be tested. Exemplary word embedding models include, but are not limited to, the GloVe model, CBOW model, or skip-Gram model, etc., but the word embedding model is not limited here.

[0081] Specifically, the location encoding unit 121 is used to store the relative or absolute position of the word in the unstructured medical record data to be tested.

[0082] In one optional embodiment, the unstructured computation unit includes a third masked multi-head self-attention layer, a fourth residual and layer normalization layer, a second feedforward neural network, and a fifth residual and layer normalization layer. Specifically, the third masked multi-head self-attention layer outputs a position encoding vector matrix based on the input unstructured features. The fourth residual and layer normalization layer outputs a normalized matrix based on the input unstructured features and the position encoding vector matrix output by the third masked multi-head self-attention layer. The second feedforward neural network outputs a second feedforward feature based on the normalized matrix output by the fourth residual and layer normalization layer. The fifth residual and layer normalization layer outputs an unstructured vector based on the normalized matrix output by the fourth residual and layer normalization layer and the second feedforward feature output by the second feedforward neural network.

[0083] In this embodiment, the input data of the spatiotemporal joint computing unit 124 includes an unstructured vector output by the unstructured computing unit, a structured time vector output by the structured time computing unit 122, and a structured spatial vector output by the structured spatial computing unit 123. Specifically, the second masking multi-head self-attention layer in the spatiotemporal joint computing unit 124 is used to output fused attention features based on the structured fusion vector output by the fusion layer and the unstructured vector output by the unstructured computing unit; the second residual and layer normalization layer is used to output a first fusion feature based on the structured fusion vector output by the fusion layer and the fused attention feature output by the second masking multi-head self-attention layer.

[0084] Figure 5 This is a model architecture diagram of another acute kidney injury prediction model provided in one embodiment of the present invention. Specifically, Figure 5 The “xxx” and “yyy” in the leftmost box represent the original unstructured medical record data. This original unstructured medical record data is input into the data preprocessing module 110, which outputs the unstructured medical record data to be tested. The location encoding unit 121 outputs unstructured features based on the input unstructured medical record data. These unstructured features are then input into the unstructured computation unit to obtain the output unstructured vector. This unstructured vector is then input into the spatiotemporal joint computation unit 124 in the acute kidney injury prediction model.

[0085] The advantage of this setup is that it increases the diversity of input data for the acute kidney injury prediction model, enabling the model to simultaneously possess the temporal continuity and spatial diversity of structured medical record data as well as the learning ability of unstructured medical record data. This further improves the timeliness and accuracy of the acute kidney injury prediction results output by the model.

[0086] The technical solution of this embodiment, by setting an acute kidney injury prediction model including a location encoding unit, a structured time calculation unit, a structured spatial calculation unit, and a spatiotemporal joint calculation unit, and correspondingly, the model processing module is specifically used for: inputting each target time structured data and each target spatial structured data into the location encoding unit to obtain the output target time structured features and target spatial structured features; inputting each target time structured feature into the structured time calculation unit in the acute kidney injury prediction model to obtain the output structured time vector; and inputting each target spatial structured feature into the structured time calculation unit in the acute kidney injury prediction model. In the spatial computing unit, the structured spatial vector is output. The structured time vector and the structured spatial vector are input into the spatiotemporal joint computing unit in the acute kidney injury prediction model to obtain the output predicted acute kidney injury result. This solves the shortcomings of the existing acute kidney injury prediction model architecture. The improved Transformer model is more suitable for the acute kidney injury prediction task provided by the embodiments of this invention, fully explores the spatiotemporal joint feature information of structured medical record data and the text information of unstructured medical record data, and further improves the timeliness and accuracy of the acute kidney injury prediction result output by the acute kidney injury prediction model.

[0087] Figure 6 This is a schematic diagram of another acute kidney injury prediction system provided in one embodiment of the present invention. The present invention further refines the training process of the acute kidney injury prediction system in the above embodiment.

[0088] like Figure 6As shown, the acute kidney injury prediction system also includes a model training module 310, which is used to: acquire training structured medical record data of the training subjects; determine at least two training time structured data based on the time sliding window and the training structured medical record data, and classify the training structured medical record data based on the indicator parameter dimensions to obtain at least two training spatial structured data; wherein, there is at least one identical training time medical record data among adjacent training time structured data; output the training acute kidney injury result based on each training time structured data, each training spatial structured data, and the untrained acute kidney injury prediction model; and adjust the model parameters of the untrained acute kidney injury prediction model based on the training acute kidney injury result and the standard acute kidney injury result until the trained acute kidney injury prediction model is obtained.

[0089] In this embodiment, the model training module 310 is specifically used to: input the structured data of each training time and the structured data of each training space into the untrained acute kidney injury prediction model to obtain the output training acute kidney injury results.

[0090] In an optional embodiment, the model training module 310 is further configured to: acquire training unstructured medical record data of the training object, and input the training unstructured medical record data into the untrained acute kidney injury prediction model.

[0091] Specifically, the training structured medical record data and / or training unstructured medical record data can be obtained from public databases. For example, a public database could be the MIMIC-III database, which is a database for intensive care patients jointly built by MIT's Physionet Lab, BIDMC (Beth Israel Deaconess Medical Center), and Philips. Currently, the database contains medical data of 53,423 adult patients.

[0092] Specifically, based on the third international consensus definition of sepsis (Sepsis-3), the sepsis patient population applicable to the embodiments of this invention was selected. The MIMIC-III database was then screened to obtain structured and / or unstructured medical record data for the training subjects. Furthermore, in accordance with the Kidney Disease Improving Global Outcomes (KDIGO) guidelines, the standard acute kidney injury outcome for the training subjects was calculated. The standard acute kidney injury outcome includes whether the training subject was diagnosed with acute kidney injury or not.

[0093] In one optional embodiment, the training structured medical record data includes training time medical record data corresponding to at least two time nodes, and the time length corresponding to each time node meets the preset time step. The model training module 310 is specifically used to: take the time node when the training object is diagnosed with sepsis as the start time node, and take the start time node corresponding to the preset look-ahead window before the training object is in a preset state as the end time node; wherein, the preset state includes being diagnosed with acute kidney injury or the end of the medical visit; and acquire at least two training time structured data in the training structured medical record data based on the time sliding window, the start time node, and the end time node.

[0094] For example, the window duration corresponding to the preset look-ahead window is 24 hours or 48 hours. Here, the window duration corresponding to the preset look-ahead window is not limited.

[0095] Figure 7 This is a time series diagram illustrating a training method for structured medical record data, provided in one embodiment of the present invention. Specifically, Figure 7 The last two white dots represent structured medical record data in the preset state of confirmed acute kidney injury. When the preset state of the training object ends in the consultation state, there are no white dots in the time series of the training structured medical record data after the preset state.

[0096] by Figure 7 For example, if the preset start time node corresponding to the prospective window is the 7th time node in the training structured medical record data, then in this embodiment, the training structured medical record data includes training time structured data corresponding to the 7 time nodes respectively.

[0097] Based on the above embodiments, optionally, the training medical record data corresponding to the model training module 310 can be used to form a training set, a validation set, and a test set. The training medical record data includes training structured medical record data, or it can include training structured medical record data and training unstructured medical record data. Specifically, based on the data proportions corresponding to the training set, validation set, and test set, each training medical record data is randomly divided into a training set, a validation set, and a test set. For example, the data proportions corresponding to the training set, validation set, and test set are 80%, 10%, and 10%, respectively; the data proportions are not limited here.

[0098] In this embodiment, the model training module 310 is specifically used for: training an incomplete acute kidney injury prediction model based on the training medical record data of the training objects in the training set to obtain a completed acute kidney injury prediction model; performing a verification operation on the completed acute kidney injury prediction model based on the training medical record data in the verification set; adjusting the hyperparameters in the completed acute kidney injury prediction model if the verification result does not meet the preset verification conditions, and using the adjusted acute kidney injury prediction model as the incomplete acute kidney injury prediction model; repeating the steps of training the incomplete acute kidney injury prediction model based on the training medical record data of the training objects in the training set to obtain a completed acute kidney injury prediction model, until the verification result meets the preset verification conditions; and evaluating the completed acute kidney injury prediction model whose verification result meets the preset verification conditions based on the training medical record data in the test set to obtain a model evaluation result.

[0099] For example, adjusting the hyperparameters in the trained acute kidney injury prediction model includes: performing a grid search operation on the hyperparameters in the trained acute kidney injury prediction model to obtain the optimal set of hyperparameters as the hyperparameters in the adjusted acute kidney injury prediction model.

[0100] For example, the evaluation parameters in the model evaluation results include, but are not limited to, accuracy, F1 score, AUROC (the area under the receiver operating characteristic, ROC curve area) and AUPRC (the area under the precision-recall curve, PRC curve area), etc.

[0101] The ROC curve, or Receiver Operating Characteristic curve, represents the probability of acute kidney injury (AKI) occurring under specific stimulus conditions. The horizontal axis represents the False Positive Rate (FPR), and the vertical axis represents the True Positive Rate (TPR). The points on the curve are determined by the TPR and FPR output at different classification thresholds based on the probability of AKI occurrence. For example, if the probability of AKI occurrence is greater than or equal to a set threshold, the training case data is predicted as positive; otherwise, it is predicted as negative. AUROC represents the area under the ROC curve and is used to evaluate classifier performance. A positive and a negative training case data are randomly selected and input into the AKI prediction model. The model outputs the probabilities of AKI occurrence for the two training cases, arranged from largest to smallest. The probability that the positive training case data appears before the negative training case data is the AUC value, which represents the likelihood that the output probability of the positive training case data is greater than that of the negative training case data.

[0102] Similarly, the horizontal axis of the PRC curve represents recall, and the vertical axis represents precision. As the classification threshold gradually decreases, more and more training case data are predicted as positive. However, these positive training case data also contain real negative training case data, meaning that both TPR and FPR will increase simultaneously. When the threshold is at its maximum, the corresponding ROC curve coordinate point is (0, 0), and when the threshold is at its minimum, the corresponding coordinate point is (1, 1). Our ideal goal is TPR = 1 and FPR = 0, corresponding to the ROC curve coordinate point (0, 1). Therefore, in the training of the acute kidney injury prediction model, we choose the probability value represented by the point on the ROC curve closest to the coordinate point (0, 1) as the classification threshold. Based on the optimal classification threshold, the accuracy, sensitivity, specificity, recall, and precision of the acute kidney injury prediction results output by the acute kidney injury prediction model can be calculated. Specifically, the formulas are: Accuracy = (TP+TN) / (TP+TN+FP+FN); Sensitivity = TP / (TP+FN); Specificity = TN / (TN+FP); Recall = TP / (TP+FN); F1 score = 2*Recall*Precision / (Recall+Precision); Precision = TP / (TP+FP). Here, TP (TruePositive) represents training case data that was actually positive but predicted as positive, FP (False Positive) represents training case data that was actually negative but predicted as positive, TN (True Negative) represents training case data that was actually negative but predicted as negative, and FN (False Negative) represents training case data that was actually positive but predicted as negative.

[0103] Figure 8 This is a flowchart illustrating the training process of an acute kidney injury prediction model according to an embodiment of the present invention. Specifically, the training process of the acute kidney injury prediction model mainly includes data collection and population screening, modeling and analysis, and result evaluation.

[0104] In the data collection and population screening process, raw structured medical record data, raw unstructured medical record data, and tags are extracted from the database. Based on the third international consensus definition of sepsis (Sepsis-3), raw structured medical record data, raw unstructured medical record data, and tags applicable to the sepsis patient population in this embodiment of the invention are selected from the extracted results. A pre-trained word embedding model is used based on a medical clinical text corpus. The raw unstructured medical record data corresponding to the sepsis patient population is pruned and deleted. The pruned and deleted raw unstructured medical record data is input into the pre-trained word embedding model to obtain the output training unstructured medical record data. Based on a preset time step, the raw structured medical record data corresponding to the sepsis patient population is time-aggregated and sorted to obtain the structured time series corresponding to the raw structured medical record data. Outlier processing, indicator standardization, and missing value processing are then performed on the structured time series to obtain the training structured medical record data. Based on a preset prospective window, the tags corresponding to the sepsis patient population, and the structured time series, the standard acute kidney injury outcome is determined. Specifically, the training medical record data consists of unstructured medical record data, structured medical record data, and standard acute kidney injury results.

[0105] In the modeling and analysis process, 80% of the training medical record data obtained from data collection and population screening was used as the training set, 10% as the validation set, and 10% as the test set. Based on the training medical record data of the training subjects in the training set, the incompletely trained acute kidney injury (AKI) prediction model was trained to obtain a fully trained AKI prediction model. The fully trained AKI prediction model was then validated based on the training medical record data in the validation set. If the validation results did not meet the preset validation conditions, a hyperparameter grid search algorithm was used to adjust the hyperparameters in the fully trained AKI prediction model. The adjusted AKI prediction model was then used as the incompletely trained AKI prediction model. This process was repeated until the validation results met the preset validation conditions.

[0106] During the results evaluation process, the hyperparameters in the trained acute kidney injury prediction model that meet the preset validation conditions obtained from the modeling and analysis process are considered the optimal hyperparameters. The trained acute kidney injury prediction model that meets the preset validation conditions obtained from the modeling and analysis process is evaluated based on the training medical record data in the test set to obtain the model evaluation results.

[0107] The technical solution of this embodiment solves the acute kidney injury prediction system by setting a model training module. This module acquires training structured medical record data of the training subjects; based on a time sliding window and the training structured medical record data, it determines at least two training time-based structured data sets and classifies the training structured medical record data sets based on indicator parameter dimensions to obtain at least two training space-based structured data sets; wherein, there is at least one identical training time-based medical record data set among adjacent training time-based structured data sets; based on each training time-based structured data set, each training space-based structured data set, and the incompletely trained acute kidney injury prediction model, it outputs the training results for acute kidney injury; based on the training results for acute kidney injury and the standard results for acute kidney injury, it adjusts the model parameters of the incompletely trained acute kidney injury prediction model until a fully trained acute kidney injury prediction model is obtained. This solves the training problem of the acute kidney injury prediction model and ensures the integrity of the acute kidney injury prediction system.

[0108] It should be noted that the acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0109] Figure 9 This is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device 10 can be configured with functional modules of the acute kidney injury prediction system in this embodiment of the present invention.

[0110] like Figure 9 As shown, electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 10 may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components, their connections and relationships, and their functions shown in the embodiments of the invention are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0111] like Figure 9As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor 11. The processor 11 can perform various appropriate actions and processes based on the computer programs stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 may also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0112] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0113] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as acute kidney injury prediction methods.

[0114] In some embodiments, the acute kidney injury prediction method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or mounted on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the acute kidney injury prediction method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the acute kidney injury prediction method by any other suitable means (e.g., by means of firmware).

[0115] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0116] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the modules described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0117] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. An acute kidney injury prediction system, characterized in that, include: Data preprocessing module, model processing module, and prediction result output module; The data preprocessing module is used to acquire the structured medical record data to be tested of the target object, and to determine at least two structured time data to be tested based on the time sliding window and the structured medical record data to be tested, and to classify the structured medical record data to be tested based on the indicator parameter dimension to obtain at least two structured spatial data to be tested; wherein, there is at least one identical structured time data to be tested between adjacent structured time data to be tested. The model processing module is used to output the predicted acute kidney injury result based on each of the time-structured data to be tested, each of the spatial-structured data to be tested, and the pre-trained acute kidney injury prediction model. The prediction result output module is used to output the predicted acute kidney injury result; The acute kidney injury prediction model includes a location encoding unit, a structured time calculation unit, a structured spatial calculation unit, and a spatiotemporal joint calculation unit. The model processing module is specifically used for: Each of the time-structured data to be tested and each of the spatial-structured data to be tested are input into the position encoding unit to obtain the output time-structured features and spatial-structured features to be tested. Each of the structured time features to be measured is input into the structured time calculation unit in the acute kidney injury prediction model to obtain the output structured time vector; Each of the aforementioned spatial structured features to be tested is input into the structured spatial calculation unit in the acute kidney injury prediction model to obtain the output structured spatial vector; The structured time vector and the structured space vector are input into the spatiotemporal joint computation unit in the acute kidney injury prediction model to obtain the output predicted acute kidney injury result.

2. The acute kidney injury prediction system according to claim 1, characterized in that, The structured medical record data to be tested includes medical record data corresponding to at least two time nodes, and the time length corresponding to each time node meets the preset time step. Accordingly, the data preprocessing module is specifically used for: For each time node, the current time node is used as the window termination node, and reference time structured data is obtained from the structured medical record data to be tested based on the time sliding window and the window termination node. If the number of reference time medical records in the reference time structured data is less than the preset number corresponding to the time sliding window, interpolation processing is performed on the reference time structured data to obtain the test time structured data corresponding to the current time node.

3. The acute kidney injury prediction system according to claim 2, characterized in that, The start time node in the structured medical record data to be tested is the first time node after the target subject is diagnosed with sepsis.

4. The acute kidney injury prediction system according to claim 2, characterized in that, The data preprocessing module is specifically used for: Based on a preset time step, at least two original time-based medical record data corresponding to the original structured medical record data of the target object are obtained; For each original time-based medical record data, a first preprocessing operation is performed on the original time-based medical record data to obtain the time-based medical record data to be tested; wherein, the first preprocessing operation includes at least one of aggregation operation, outlier handling operation, standardization operation, and missing value handling operation; Based on the medical record data of each time period to be tested, the structured medical record data to be tested corresponding to the target object is generated.

5. The acute kidney injury prediction system according to claim 1, characterized in that, The structured temporal computation unit includes a first masking self-attention layer and a first residual and layer normalization layer; the structured spatial computation unit includes a one-dimensional convolutional layer, a fully connected layer, and a layer normalization layer; and the spatiotemporal joint computation unit includes a fusion layer, a second masking multi-head self-attention layer, a second residual and layer normalization layer, a first feedforward neural network, a third residual and layer normalization layer, and an output layer.

6. The acute kidney injury prediction system according to claim 1, characterized in that, The acute kidney injury prediction system further includes a model training module, which is used for: Obtain the training structured medical record data of the training subjects; Based on the time sliding window and the training structured medical record data, at least two training time structured data are determined, and the training structured medical record data are classified based on the indicator parameter dimension to obtain at least two training spatial structured data; wherein, there is at least one identical training time medical record data among adjacent training time structured data. Based on the structured training time data, the structured training space data, and the incomplete training of the acute kidney injury prediction model, the training results for acute kidney injury are output. Based on the training results of acute kidney injury and the standard acute kidney injury results, the model parameters of the untrained acute kidney injury prediction model are adjusted until a trained acute kidney injury prediction model is obtained.

7. The acute kidney injury prediction system according to claim 6, characterized in that, The training structured medical record data includes training time medical record data corresponding to at least two time nodes, and the time length corresponding to each time node meets a preset time step. The model training module is specifically used for: The time point at which the training subject is diagnosed with sepsis is taken as the start time point, and the start time point corresponding to the preset prospective window before the training subject is in the preset state is taken as the end time point; wherein, the preset state includes diagnosis of acute kidney injury or the end of the medical treatment state. Based on a time sliding window, a start time node, and an end time node, at least two training time structured data points are obtained from the training structured medical record data.

8. The acute kidney injury prediction system according to claim 1, characterized in that, The data preprocessing module is further configured to: perform a second preprocessing operation on the acquired raw unstructured medical record data to obtain the unstructured medical record data to be tested; wherein, the second preprocessing operation includes field filtering operation and vector transformation operation; Accordingly, the acute kidney injury prediction model also includes an unstructured computation unit, and the model processing module is further used to: input the unstructured medical record data to be tested into the position encoding unit in the pre-trained acute kidney injury prediction model to obtain the output unstructured features. The unstructured features are input into the unstructured computation unit of the pre-trained acute kidney injury prediction model to obtain the output unstructured vector, and the unstructured vector is input into the spatiotemporal joint computation unit of the acute kidney injury prediction model.

9. The acute kidney injury prediction system according to claim 8, characterized in that, The unstructured computing unit includes a third masked multi-head self-attention layer, a fourth residual and layer normalization layer, a second feedforward neural network, and a fifth residual and layer normalization layer.

Images