CN116364290B - Hemodialysis characterization identification and complications risk prediction system based on multi-view alignment - Google Patents
Hemodialysis characterization identification and complications risk prediction system based on multi-view alignment Download PDFInfo
- Publication number
- CN116364290B CN116364290B CN202310644753.5A CN202310644753A CN116364290B CN 116364290 B CN116364290 B CN 116364290B CN 202310644753 A CN202310644753 A CN 202310644753A CN 116364290 B CN116364290 B CN 116364290B
- Authority
- CN
- China
- Prior art keywords
- data
- patient
- view
- characterization
- risk prediction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G06N3/0455—Auto-encoder networks; Encoder-decoder networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
- G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biomedical Technology (AREA)
- Public Health (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Epidemiology (AREA)
- Primary Health Care (AREA)
- Databases & Information Systems (AREA)
- Pathology (AREA)
- External Artificial Organs (AREA)
Abstract
The application discloses a blood-penetration characterization identification and complications risk prediction system based on multi-view alignment, which comprises a data preparation module for acquiring and rectifying blood-penetration patient data and a blood-penetration characterization identification module for blood-penetration characterization identification and complications risk prediction. The application adopts a multi-view representation input method to acquire individual characteristic data, medication data, diagnosis data and inspection data of a patient, and constructs various patient views through a characteristic extraction unit and a multi-view mapping unit to provide comprehensive representation of the patient. According to the application, different characteristic extraction units are utilized to perform characteristic extraction on different types of patient data, semantic information of different data is reserved, and potential complementarity and consistency information among different views are mined by constructing consistency loss items and complementarity loss items of different views, so that more complete and non-redundant characteristic representation is obtained, and the performance of a learning task is improved. The application can provide accurate and effective decision support for clinical prediction.
Description
Technical Field
The application belongs to the technical field of medical health information, and particularly relates to a blood perspective characterization identification and complication risk prediction system based on multi-view alignment.
Background
Hemodialysis is an important treatment scheme for acute and chronic renal failure, and removes metabolic waste in vivo, maintains acid-base balance and reduces renal pressure of patients through diffusion, ultrafiltration, adsorption and convection principles. However, the risk of complications in the long-term hemodialysis process is high, the serious complications can endanger the life safety of patients, and if the complications of the hemodialysis patients can be predicted as early as possible and the patients are subjected to corresponding therapeutic intervention, the complications can be reduced or avoided. Therefore, the method improves the blood permeation complication prediction capability and has great significance for improving the life quality of blood permeation patients.
With the rise of artificial intelligence, more and more researchers use deep neural networks or machine learning methods to predict the risk of hemodialysis patient complications. At present, in the process of predicting the hemodialysis complications, the characterization input of patients is divided into two types:
the first is a single view representation input method. And taking the diagnosis view or the clinical examination view of the patient as the input of a prediction model to predict the risk of complications of the hemodialysis patient. Since a single view can only represent a patient from one angle of the patient, a comprehensive representation of the patient cannot be provided, and thus excellent predictive power cannot be achieved.
The second is a multi-view token fusion input method. And taking individual characteristics of the patient, clinical examination and other views as input of a prediction model after characteristic extraction and fusion, and predicting the risk of complications of the hemodialysis patient. Although the method solves the problem that a single view can not provide comprehensive representation of a patient, the existing multi-view characterization fusion input method can lose independent semantic information among features and does not fully utilize the relationship among the features, so that the performance of a learning task is affected.
Disclosure of Invention
Aiming at the defects of the prior art, the application provides a hemodialysis characterization identification and complications risk prediction system based on multi-view alignment for providing accurate and effective decision support for clinical prediction.
The application aims at realizing the following technical scheme: a multi-view alignment-based hemodialysis characterization identification and complications risk prediction system, which comprises a data preparation module and a hemodialysis characterization identification module;
the data preparation module is used for acquiring hemodialysis patient data, and classifying and integrating the data according to static data, one-dimensional time sequence data and two-dimensional time sequence data after cleaning the data;
the hemodialysis characterization and identification module comprises a feature extraction unit, a multi-view mapping unit and a risk prediction unit;
the feature extraction unit is used for respectively setting corresponding feature extraction units for the integrated static data, the one-dimensional time sequence data and the two-dimensional time sequence data to carry out different feature extraction so as to obtain different hemodialysis characterization;
the multi-view mapping unit is used for mapping different blood transmittance characteristics to different patient views;
the risk prediction unit is used for carrying out complication risk prediction according to the patient view, and the prediction loss function of the risk prediction unit comprises a target task loss item, a consistency loss item and a complementation loss item, wherein the consistency loss item is used for measuring consistency differences between different patient view outputs, and the complementation loss item is used for measuring mutual information between different patient view hemodialysis characterizations.
Further, the hemodialysis patient data acquired by the data preparation module includes individual characteristic data, medication data, diagnostic data, examination data, and medical result data.
Further, the two-dimensional time sequence data characteristic extraction unit comprises two-way long-short-period memory networks, two attention layers and a characterization alignment layer; the first two-way long-short-period memory network is used for capturing the blood pressure dynamic change relation of a patient in one blood permeation process, the second two-way long-short-period memory network is used for capturing the blood pressure dynamic change relation of the patient between blood permeation at each time, an attention layer is connected behind each two-way long-short-period memory network, and the characteristic representation length of each characteristic extraction unit is guaranteed to be the same through the representation alignment layer.
Further, the one-dimensional time sequence data characteristic extraction unit comprises a two-way long-short-period memory network, an attention layer and a characterization alignment layer which are sequentially connected; the two-way long-short-term memory network is used for capturing the front-back change relation of one-dimensional time sequence data of a patient, and the characteristic length of each characteristic extraction unit is identical through the characteristic alignment layer.
The static data feature extraction unit is further constructed by an automatic encoder, the automatic encoder comprises an encoder part and a decoder part, the static data feature extraction unit is pre-trained, static data is subjected to single-heat encoding and then is input as the automatic encoder, mean square error is used as a loss function of the automatic encoder, a random gradient descent method is used for optimization, and after the pre-training is completed, the encoder part is used for feature extraction of the static data.
Further, the multi-view mapping unit is configured to map different hemodialysis characterizations to different patient views, including individual feature views, diagnostic views, examination views, and medication views, with the different view mapping units.
Further, the target task loss term of the risk prediction unit is used for measuring the difference between the target task true value and the output of the risk prediction unit, and cross entropy loss is used.
Further, the consistency loss term of the risk prediction unit uses KL divergence to measure the distribution differences between different patient view outputs.
Further, the complementarity loss term of the risk prediction unit measures the amount of complementary information contained in the blood vessel characterization of the different views using mutual information between the blood vessel characterization of the different views.
Further, the training of the system comprises two stages, wherein the first stage is a pre-training static data characteristic extraction unit and fixes parameters thereof; the second stage is to train other units except the static data feature extraction unit, calculate the loss value of the risk prediction unit by inputting patient data, and update the parameters of the risk prediction unit, the multi-view mapping unit, the two-dimensional time sequence data feature extraction unit and the one-dimensional time sequence data feature extraction unit by gradient back propagation.
The beneficial effects of the application are as follows:
1. aiming at the problem that the single-view representation input method cannot provide comprehensive representation of a patient, the multi-view representation input method is adopted, individual characteristic data, medication data, diagnosis data and inspection data of the patient are collected, various patient views are constructed through the characteristic extraction unit and the multi-view mapping unit, and comprehensive representation of the patient is provided.
2. Aiming at the problems that independent semantic information among features is lost and the relationship among the features is not fully utilized in the conventional multi-view characterization fusion input method, the application utilizes different feature extraction units to perform feature extraction on different types of patient data, reserves semantic information of different data, and extracts potential complementarity and consistency information among different views by constructing consistency loss items and complementarity loss items of different views to acquire more complete and non-redundant feature representation, thereby improving the performance of learning tasks.
3. According to the application, by constructing two-way long-short-term memory networks, semantic information of different time dimensions of two-dimensional time sequence data is extracted, and the obtained hemodialysis characterization information is more accurate.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions of the prior art, the drawings that are needed in the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a block diagram of a multi-view alignment based blood vessel characterization identification and complications risk prediction system in accordance with an exemplary embodiment;
FIG. 2 is a block diagram of a blood permeation characterization identification module, as shown in an exemplary embodiment;
FIG. 3 is a block diagram of a two-dimensional temporal data feature extraction unit shown in an exemplary embodiment;
FIG. 4 is a block diagram of a two-way long and short term memory network shown in an exemplary embodiment;
FIG. 5 is a block diagram of a one-dimensional time series data feature extraction unit shown in an exemplary embodiment;
FIG. 6 is a block diagram of an automatic encoder shown in an exemplary embodiment;
fig. 7 is a block diagram of a risk prediction unit according to an exemplary embodiment.
Detailed Description
For a better understanding of the technical solution of the present application, the following detailed description of the embodiments of the present application refers to the accompanying drawings.
It should be understood that the described embodiments are merely some, but not all, embodiments of the application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.
The terminology used in the embodiments of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in this application and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.
The embodiment of the application provides a system for identifying blood permeation characterization and predicting complications risk based on multi-view alignment, which comprises a data preparation module and a blood permeation characterization identification module as shown in fig. 1. The data preparation module is used for collecting and rectifying the hemodialysis patient data and comprises a data acquisition unit, a data cleaning unit and a data integration unit, and the hemodialysis characterization identification module is used for carrying out risk prediction on the hemodialysis patient data processed by the data preparation module and comprises a feature extraction unit, a multi-view mapping unit and a risk prediction unit.
The following description further presents some embodiments of the implementation of the modules of the multi-view alignment-based blood vessel characterization identification and complications risk prediction system in accordance with the requirements of the present application.
1. A data preparation module comprising the following three units: the device comprises a data acquisition unit, a data cleaning unit and a data integration unit.
1.1 data acquisition Unit
Structured data of a hemodialysis patient is acquired using an electronic medical record system of a hospital, the data including: (1) individual characteristic data: height, sex, age, region, weight; (2) medication data: daily medication information including contrast agents, anticoagulants, antiplatelets, inotropic agents, vasodilators; (3) diagnostic data: cardiomyopathy, cerebral apoplexy, valvular disease, atrial fibrillation, and coronary heart disease; (4) checking data: blood pressure, urea, creatinine, potassium, hemoglobin; (5) medical results: complications occur.
1.2 data cleaning Unit
The data cleaning unit is used for cleaning the structured data of the hemodialysis patient acquired by the data acquisition unit, and comprises missing value processing, error value detection, elimination of repeated data and/or elimination operation of inconsistency.
1.3 data integration Unit
The data integration unit is used for classifying and processing the data cleaned by the data cleaning unit according to the static data, the one-dimensional time sequence data and the two-dimensional time sequence data.
The data with time characteristics are time sequence data, which comprise one-dimensional time sequence data and two-dimensional time sequence data, original complete information of the time sequence data is represented by (feature names, feature values and time stamps), for example, data such as contrast agents, anticoagulants, antiplatelets, positive inotropic drugs, vasodilators, cardiomyopathy, cerebral apoplexy, valvular disease, atrial fibrillation, coronary heart disease, urea, creatinine, potassium, hemoglobin and the like are one-dimensional time sequence data, and blood pressure data are two-dimensional time sequence data.
The data without time characteristics are static data, and the original complete information of the static data, such as height, gender, age, region, weight and the like, is represented by (feature names, feature values).
One-dimensional time sequence data integration is exemplified by creatinine, and the integrated creatinine data is oneOne-dimensional vectors, each value in the vector representing the results of a blood creatinine examination made by the patient each time before hemodialysis.
Two-dimensional time series data integration is exemplified by blood pressure, and the integrated blood pressure data is oneThe two-dimensional matrix, n, represents the number of patient's blood permeabilities and m represents the number of times a patient has been logged with blood pressure values during one blood permeance.
Static data is a discrete type of data that is integrated during processing using one-hot encoding.
The integrated original data is expressed as X, and the specific characteristics are expressed asFor example->。
2. A blood permeation characterization identification module, as shown in fig. 2, comprising the following three units: the device comprises a feature extraction unit F, a multi-view mapping unit S and a risk prediction unit G.
2.1 feature extraction Unit F
The feature extraction unit F is used for performing feature extraction on the integrated original data X to obtain a blood-penetration characterization, and using e to represent the blood-penetration characterization, and the different blood-penetration characterizations are represented asFor example->. The feature extraction unit F includes a two-dimensional time series data feature extraction unit, a one-dimensional time series data feature extraction unit, and a static data feature extraction unit.
(1) Two-dimensional time sequence data characteristic extraction unit
The two-dimensional time sequence data feature extraction unit is used for performing feature extraction on the two-dimensional time sequence data, and as shown in fig. 3, the two-dimensional time sequence data feature extraction unit consists of two-way long-short-term memory networks, two attention layers and a representation alignment layer.
Because of the uniqueness of the two-dimensional time sequence data, the blood pressure data comprises data of two time dimensions, for example, blood pressure data comprises dynamic changes of blood pressure of a patient in a blood permeation process, and also comprises dynamic changes of blood pressure of the patient between blood permeation processes. In order to capture the dynamic change of blood pressure in one blood permeation process of a patient and the dynamic change of blood pressure between each blood permeation, the application designs two-way long-short-period memory networks. As shown in fig. 4, the structure diagram of the bidirectional long-short-term memory network is shown, and compared with the unidirectional long-short-term memory network, the bidirectional long-short-term memory network comprises a forward long-short-term memory network and a reverse long-short-term memory network, so that the sequence can be traversed in both forward and reverse directions, and more characteristics can be extracted. In the blood pressure characteristic extraction process, the first two-way long-short-term memory network is used for capturing the blood pressure dynamic change relation in the process of one blood permeation of a patient, and the second two-way long-term memory network is used for capturing the blood pressure dynamic change relation between each blood permeation of the patient. Although the bidirectional long-short-term memory network can capture information in the forward direction and the reverse direction of blood pressure, it cannot determine which parts are more important and which parts can be ignored, for this purpose, an attention layer is added behind each bidirectional long-short-term memory network, and the attention layer calculates the importance of each time step in the sequence, and performs weighted average on the information of different time steps according to the importance, so as to obtain a more comprehensive blood-level representation, thereby improving the robustness and accuracy of the network. Meanwhile, in order to ensure that the characteristic characterization lengths extracted by each characteristic extraction unit are the same, the last layer of the two-dimensional time sequence data characteristic extraction unit is a characterization alignment layer, the two-layer full-connection network is used for realizing the two-layer full-connection network, the number of nodes of the two layers is 128 and 64 respectively, a Tanh activation function is used, and the number of nodes of the last output layer is 8.
The two-dimensional time series data feature extraction unit calculation process will be described below by taking the extraction of a blood pressure characterization of a patient as an example.
UsingRepresenting patient blood pressure data, wherein
Represents the nth hemodialysis blood pressure data of the patient, and m times of blood pressure values are recorded in the blood permeation process, wherein +.>The blood pressure value at the mth time in the nth blood permeation process is shown. The first two-way long and short term memory network is calculated as follows:
wherein the method comprises the steps ofThe blood pressure value at the t time in the nth blood permeation process of the patient is represented; />Representing the forward computational unit function in the first two-way long and short term memory network,/for>Representing a reverse computational unit function in a first two-way long and short term memory network; [,]representing a splicing function; />The forward hidden output, representing the time t, is a length 4 vector,the reverse hidden output representing the time t is a vector of length 4,/for example>The hidden output at the t-th time is represented by the forward hidden output at the t-th time +>And reverse hidden output +.>Splicing to obtain the final product; then
The first two-way long-short-period memory network hiding output of the nth hemodialysis blood pressure data of the patient is indicated.
The first attention layer calculation procedure is as follows:
wherein the method comprises the steps ofAttention calculating function representing the first attention layer,/->Representing an activation function->Respectively, query, key, value in the attention model, by matrix +.>Obtaining linear change; />Representation matrix->Is a dimension of (2); t represents a transpose; />Are all dimension +.>A trainable weight matrix; />Representing a matrix multiplication; />The first attentive layer of the nth blood permeation blood pressure data of the patient is represented to be hidden and output.
The second two-way long-short term memory network is calculated as follows:
wherein the method comprises the steps ofA first attention layer hidden output of the kth blood-permeable blood pressure data of the patient is represented; />Representing a forward computing unit function in a second two-way long-short term memory network; />Representing a reverse computing unit function in a second two-way long-short term memory network; [,]representing a splicing function; />The forward hidden output representing the kth time is a vector of length 4, +.>The reverse concealment output representing the kth time is a length-4 vector, ++>Represents the hidden output of the kth time, is outputted by the positive hidden output of the kth time +.>And reverse hidden output +.>Splicing to obtain the final product; thenThe second two-way long-short-term memory network hidden output of the blood pressure data of the patient is represented.
The second attention layer calculation process is as follows:
wherein the method comprises the steps ofAttention calculating function representing the second attention layer,/->Representing an activation function->Respectively, query, key, value in the attention model, by matrix +.>Obtaining linear change; />Representation matrix->Is a dimension of (2); t represents a transpose; />Are all dimension +.>A trainable weight matrix; />Representing a matrix multiplication; />A second attentive layer of the patient's blood pressure data is shown as hidden output.
At the same time, in order to ensure the feature characterization length extracted by each feature extraction unitThe last layer of the two-dimensional time sequence data characteristic extraction unit is a representation alignment layer. The characterization alignment layer of each feature has the same structure and different parameters, the structure is realized by two layers of fully connected networks, the number of nodes of the two layers is 128 and 64 respectively, a Tanh activation function is used, and the number of nodes of the final output layer is 8. Final patient blood pressure characterizationWherein->Alignment of the function of the layer representation for the characterization of blood pressure,/->A second attentive layer hidden output of blood pressure data representing patient, < >>Is a flattening function.
(2) One-dimensional time sequence data characteristic extraction unit
As shown in FIG. 5, the one-dimensional time series data characteristic unit is composed of a two-way long-short-period memory network, an attention layer and a characterization alignment layer. The method comprises the steps of capturing a front-back change relation of one-dimensional time sequence data of a patient by using a two-way long-short-term memory network, adding an attention layer behind the two-way long-short-term memory network by introducing an attention mechanism, calculating a result of the two-way long-short-term memory network, and carrying out weighted average on information of different time steps according to importance, so as to obtain a more comprehensive blood penetration representation. The characterization alignment layer is used for ensuring that the feature characterization extracted by the one-dimensional time sequence data feature extraction unit is identical to the feature characterization extracted by other feature extraction units in length, and is realized by two layers of full-connection layers, wherein the node number of each layer is 256 and 128 respectively, a Tanh activation function is used, and the node number of the final output layer is 8.
Before each time the patient is transfused, the patient is subjected to serum creatinine examination and usedRepresenting patient's blood creatinine data, wherein>The results of the creatinine examination performed prior to the nth hemodialysis of the patient are shown. Next, taking creatinine as an example, a calculation process of the one-dimensional time sequence data feature extraction unit is described:
wherein the method comprises the steps ofRepresenting the results of a creatinine examination of the patient prior to the t-th hemodialysis, the +.>Representing a forward long and short term memory calculation unit function, < ->Representing a reverse long-short-term memory calculation unit function; [,]representing a splicing function;the forward hidden output representing the t-th time is a vector of length 4, +.>The reverse concealment output representing the t-th time is a vector of length 4,/h>Represents the hidden output of the t th time, and is outputted by the forward hidden output of the t th time +.>And reverse hidden output +.>Splicing to obtain the final product; />Representing hidden output of the two-way long-short-term memory network; />Representing an attention calculating function +.>Representing an activation function->Respectively, query, key, value in the attention model, by matrix +.>Obtaining linear change; />Representation matrix->
Is a dimension of (2); t represents a transpose;are all dimension +.>A trainable weight matrix; />Representing a matrix multiplication; />The patient's blood-permeable creatinine data attention layer is shown with hidden output.
Meanwhile, in order to ensure that the characteristic characterization lengths extracted by each characteristic extraction unit are the same, the last layer of the one-dimensional time sequence data characteristic extraction unit is a characterization alignment layer. The characterization alignment layer of each feature has the same structure and different parameters, the structure is realized by two layers of fully connected networks, the number of nodes of the two layers is 128 and 64 respectively, a Tanh activation function is used, and the number of nodes of the final output layer is 8. Characterization of the final patient's blood creatinineWherein->Alignment layer representation function for characterization of creatinine>Is a flattening function.
(3) Static data feature extraction unit
Because the static data are all discrete data, the feature lengths obtained by different features are different after the single-heat encoding. In order to ensure that the characteristic length of the static data after the single-heat encoding is the same as the characteristic length of other hemodialysis, the static data characteristic extraction unit is constructed by using an automatic encoder. As shown in fig. 6, which is a structural diagram of an automatic encoder, the automatic encoder comprises an encoder and a decoder, the encoder and the decoder are respectively realized by a fully-connected network consisting of 128 nodes, and the activation function is ReLU; hemodialysis is characterized by a vector of length 8.
In the pre-training process, static data of a patient are subjected to single-heat coding and then are used as input of an automatic encoder, mean square error is used as a loss function of the automatic encoder, the automatic encoder is optimized by using a random gradient descent method, and after the pre-training is finished, an encoder part in the automatic encoder is used for extracting characteristics of the static data.
2.2 Multi-view mapping Unit S
The multi-view mapping unit S is used for mapping different blood vessel characterizations to different patient views using different view mapping units, usingRepresenting a patient view, the different patient views being denoted +.>For example->. The multi-view mapping unit S includes an individual feature view mapping unit, a diagnostic view mapping unit, an inspection view mapping unit, and a medication view mapping unit.
Each view mapping unit is realized by a three-layer full-connection network, the three-layer node numbers are 64, 16 and 8 respectively, the activation functions are ReLU, the network is optimized by using a random gradient descent method, and finally the node number of the output layer is 10. But the blood perspective representation input by each view mapping unit is different, wherein the representation input by the individual feature view mapping unitThe method comprises the steps of carrying out a first treatment on the surface of the Characterization of medication view mapping unit inputThe method comprises the steps of carrying out a first treatment on the surface of the Characterization of the diagnostic View mapping Unit input +.>The method comprises the steps of carrying out a first treatment on the surface of the Checking the representation of the view mapping unit input +.>。
Taking an individual feature view mapping unit as an example, the view mapping unit calculation process is described:
wherein the method comprises the steps ofRepresenting individual feature view mapping unit network functions, +.>For individual characterization, ->For flattened function, ++>Representing the individual feature view mapping unit output, i.e. the individual feature view.
2.3 Risk prediction Unit G
The risk prediction unit G is used for performing risk prediction on the input patient view, and judging the risk of complications of the patient.
As shown in fig. 7, the risk prediction unit is implemented by a three-layer fully-connected network, and the three-layer nodes are 128, 64 and 10 respectively. The activation function of the first two layers is ReLU, and the last output layer activation function isThe entire network is optimized using a random gradient descent method.
The prediction loss function of the risk prediction unit consists of a target task loss term, a consistency loss term and a complementarity loss term, wherein the target task loss term measures the difference between a target task true value and the output of the risk prediction unit, and the smaller the loss is, the more successful the risk prediction fitting is. The consistency loss term measures the consistency difference between different view outputs, and the smaller the loss, the more consistent the different view output distribution. The term of complementarity loss measures mutual information between the hemodialysis representations of different views, and the smaller the loss, the more complementary information the hemodialysis representations of different views contain. The calculation process of each loss term is as follows:
(1) Target task loss item
Target task loss item in this embodimentUsing the cross entropy loss, the calculation formula is as follows:
wherein the method comprises the steps ofRepresenting a full view of the patient; />Patient-indicating genuine label->Indicating that the patient is happeningComplications(s) (I)>Indicating that the patient is not experiencing complications; />Representing the output value of the risk prediction unit, for predicting the probability of a complication occurring in the patient,/for the patient>The full view of the patient is represented as the input risk prediction unit output value.
(2) Consistency loss term
Consistency loss term in this embodimentThe calculation formula of (2) is as follows:
wherein the method comprises the steps ofRepresenting a view name set, wherein d is the total number of views; />Representing a distribution difference metric function between different view outputs, in this embodiment using KL divergence metrics, the calculation formula is as follows:
wherein the method comprises the steps ofRespectively representing the ith view of the patient, the jth view as input the risk prediction unit output value,
。
(3) Complementarity loss term
Complementarity loss term in this embodimentThe calculation formula of (2) is as follows:
wherein the method comprises the steps ofRepresenting a view name set, wherein d is the total number of views; />The mutual information measurement function between the blood transmission characterization of different views is represented, and the calculation formula in the embodiment is as follows:
wherein the method comprises the steps ofRespectively representing the ith view, the jth view of the patient as input the risk prediction unit output value,/for the patient>Representing the output value of the risk prediction unit when the ith view and the jth view of the patient are spliced to be used as input.
(4) Prediction loss function of risk prediction unitThe calculation formula of (2) is as follows:
wherein the method comprises the steps ofIs a super parameter, is used for balancing the force of two constraints of consistency and complementarity, and is in the embodiment，/>。
The training of the whole multi-view alignment-based hemodialysis characterization recognition and complications risk prediction system is divided into two stages:
the first stage is to pre-train the static data feature extraction unit, train the automatic encoder by using the integrated static data of the patient, and after the automatic encoder is trained, the encoder part in the automatic encoder is used for feature extraction of the static data, and when the other units are trained in the second stage, the parameters of the encoder part of the static data feature extraction unit are not updated any more.
The second stage is to train other units except the static data feature extraction unit, calculate the loss value of the risk prediction unit by inputting patient data, and update the parameters of the risk prediction unit, the multi-view mapping unit, the two-dimensional time sequence data feature extraction unit and the one-dimensional time sequence data feature extraction unit by gradient back propagation.
After the whole system is trained, a doctor inputs the data of the hemodialysis patient to be predicted into the system, and the system gives the risk probability of the complications of the patient after calculation.
According to the hemodialysis characterization recognition and complications risk prediction system based on multi-view alignment, different types of hemodialysis features are extracted by using different feature extraction units, semantic information of different data is reserved, and the problem that fusion features lose independent semantic information among the features is solved. According to the application, the consistency loss items and the complementarity loss items of different views are constructed, the potential complementarity and consistency information between different views is mined, and the more complete and non-redundant characteristic representation is obtained, so that the performance of a learning task is improved. The application can provide accurate and effective decision support for clinical prediction.
It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article or apparatus that comprises the element.
The foregoing describes specific embodiments of the present disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
The terminology used in the one or more embodiments of the specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of the specification. As used in this specification, one or more embodiments and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any or all possible combinations of one or more of the associated listed items.
The foregoing description of the preferred embodiment(s) is (are) merely intended to illustrate the embodiment(s) of the present application, and it is not intended to limit the embodiment(s) of the present application to the particular embodiment(s) described.
Claims (6)
1. The system for identifying the hemodialysis characterization and predicting the complications risk based on multi-view alignment is characterized by comprising a data preparation module and a hemodialysis characterization identification module;
the data preparation module is used for acquiring hemodialysis patient data, and classifying and integrating the data according to static data, one-dimensional time sequence data and two-dimensional time sequence data after cleaning the data;
the hemodialysis characterization and identification module comprises a feature extraction unit, a multi-view mapping unit and a risk prediction unit;
the feature extraction unit is used for respectively setting corresponding feature extraction units for the integrated static data, the one-dimensional time sequence data and the two-dimensional time sequence data to carry out different feature extraction so as to obtain different hemodialysis characterization;
the multi-view mapping unit is used for mapping different blood transmittance characteristics to different patient views by using different view mapping units, wherein the patient views comprise an individual characteristic view, a diagnosis view, an examination view and a medication view;
the risk prediction unit is used for carrying out complication risk prediction according to a patient view, and a prediction loss function of the risk prediction unit comprises a target task loss item, a consistency loss item and a complementarity loss item;
the consistency loss term is used for measuring consistency differences among different patient view outputs, in particular to measuring distribution differences among different patient view outputs by using KL divergence;
the complementarity loss item is used for measuring mutual information between the blood transmission representations of different patient views, in particular measuring the complementary information quantity contained in the blood transmission representations of different views by using the mutual information between the blood transmission representations of different views;
the training of the system comprises two stages, wherein the first stage is a pre-training static data characteristic extraction unit and fixes parameters thereof; the second stage is to train other units except the static data feature extraction unit, calculate the loss value of the risk prediction unit by inputting patient data, and update the parameters of the risk prediction unit, the multi-view mapping unit, the two-dimensional time sequence data feature extraction unit and the one-dimensional time sequence data feature extraction unit by gradient back propagation.
2. The multi-view alignment based hemodialysis characterization identification and complications risk prediction system of claim 1, wherein the hemodialysis patient data acquired by the data preparation module includes individual characteristic data, medication data, diagnostic data, examination data, and medical result data.
3. The multi-view alignment-based blood vessel characterization identification and complications risk prediction system according to claim 1, wherein the two-dimensional time series data feature extraction unit comprises two-way long-short-term memory networks, two attention layers and a characterization alignment layer; the first two-way long-short-period memory network is used for capturing the blood pressure dynamic change relation of a patient in one blood permeation process, the second two-way long-short-period memory network is used for capturing the blood pressure dynamic change relation of the patient between blood permeation at each time, an attention layer is connected behind each two-way long-short-period memory network, and the characteristic representation length of each characteristic extraction unit is guaranteed to be the same through the representation alignment layer.
4. The multi-view alignment-based hemodialysis characterization recognition and complications risk prediction system according to claim 1, wherein the one-dimensional time sequence data feature extraction unit comprises a two-way long-short-term memory network, an attention layer and a characterization alignment layer which are sequentially connected; the two-way long-short-term memory network is used for capturing the front-back change relation of one-dimensional time sequence data of a patient, and the characteristic length of each characteristic extraction unit is identical through the characteristic alignment layer.
5. The multi-view alignment-based blood vessel characterization recognition and complications risk prediction system according to claim 1, wherein the static data feature extraction unit is constructed by an automatic encoder, the automatic encoder comprises an encoder part and a decoder part, the static data feature extraction unit is pre-trained, the static data is input as an automatic encoder after being subjected to single-heat encoding, the mean square error is used as a loss function of the automatic encoder, the random gradient descent method is used for optimizing, and after the pre-training is completed, the encoder part is used for feature extraction of the static data.
6. The multiple view alignment based blood vessel characterization identification and complications risk prediction system according to claim 1, wherein the target task loss term of the risk prediction unit is used to measure the difference between the target task truth and the risk prediction unit output, using cross entropy loss.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310644753.5A CN116364290B (en) | 2023-06-02 | 2023-06-02 | Hemodialysis characterization identification and complications risk prediction system based on multi-view alignment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310644753.5A CN116364290B (en) | 2023-06-02 | 2023-06-02 | Hemodialysis characterization identification and complications risk prediction system based on multi-view alignment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116364290A CN116364290A (en) | 2023-06-30 |
| CN116364290B true CN116364290B (en) | 2023-09-08 |
Family
ID=86928579
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310644753.5A Active CN116364290B (en) | 2023-06-02 | 2023-06-02 | Hemodialysis characterization identification and complications risk prediction system based on multi-view alignment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116364290B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117034142B (en) * | 2023-10-07 | 2024-02-09 | 之江实验室 | Unbalanced medical data missing value filling method and system |
| CN117574244B (en) * | 2024-01-15 | 2024-04-02 | 成都秦川物联网科技股份有限公司 | Ultrasonic water meter fault prediction method, device and equipment based on Internet of things |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110392914A (en) * | 2017-02-03 | 2019-10-29 | 费奥普斯有限公司 | The system and method for determining the risk of hemodynamics insufficiency after heart intervention treating |
| CN111340067A (en) * | 2020-02-10 | 2020-06-26 | 天津大学 | Redistribution method for multi-view classification |
| CN113658721A (en) * | 2021-07-19 | 2021-11-16 | 南京邮电大学 | Alzheimer disease process prediction method |
| CN114883003A (en) * | 2022-06-08 | 2022-08-09 | 中南大学 | ICU (intensive care unit) hospitalization duration and death risk prediction method based on convolutional neural network |
| CN114913982A (en) * | 2022-07-18 | 2022-08-16 | 之江实验室 | End-stage renal disease complication risk prediction system based on contrast learning |
| CN115223679A (en) * | 2022-08-05 | 2022-10-21 | 华中科技大学同济医学院附属同济医院 | Perioperative period risk early warning method based on machine learning |
| CN115547502A (en) * | 2022-11-23 | 2022-12-30 | 浙江大学 | Hemodialysis patient risk prediction device based on time sequence data |
| CN115831377A (en) * | 2022-07-01 | 2023-03-21 | 中南大学 | Intra-hospital death risk prediction method based on ICU (intensive care unit) medical record data |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11406269B2 (en) * | 2008-10-29 | 2022-08-09 | Flashback Technologies, Inc. | Rapid detection of bleeding following injury |
| EP3866176A1 (en) * | 2020-02-17 | 2021-08-18 | Siemens Healthcare GmbH | Machine-based risk prediction for peri-procedural myocardial infarction or complication from medical data |
-
2023
- 2023-06-02 CN CN202310644753.5A patent/CN116364290B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110392914A (en) * | 2017-02-03 | 2019-10-29 | 费奥普斯有限公司 | The system and method for determining the risk of hemodynamics insufficiency after heart intervention treating |
| CN111340067A (en) * | 2020-02-10 | 2020-06-26 | 天津大学 | Redistribution method for multi-view classification |
| CN113658721A (en) * | 2021-07-19 | 2021-11-16 | 南京邮电大学 | Alzheimer disease process prediction method |
| CN114883003A (en) * | 2022-06-08 | 2022-08-09 | 中南大学 | ICU (intensive care unit) hospitalization duration and death risk prediction method based on convolutional neural network |
| CN115831377A (en) * | 2022-07-01 | 2023-03-21 | 中南大学 | Intra-hospital death risk prediction method based on ICU (intensive care unit) medical record data |
| CN114913982A (en) * | 2022-07-18 | 2022-08-16 | 之江实验室 | End-stage renal disease complication risk prediction system based on contrast learning |
| CN115223679A (en) * | 2022-08-05 | 2022-10-21 | 华中科技大学同济医学院附属同济医院 | Perioperative period risk early warning method based on machine learning |
| CN115547502A (en) * | 2022-11-23 | 2022-12-30 | 浙江大学 | Hemodialysis patient risk prediction device based on time sequence data |
Non-Patent Citations (1)
| Title |
|---|
| Cardiac Complication Risk Profiling for Cancer Survivors via Multi-View Multi-Task Learning;Thai-Hoang Pham et al;《2021 IEEE International Conference on Data Mining (ICDM)》;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116364290A (en) | 2023-06-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116364290B (en) | Hemodialysis characterization identification and complications risk prediction system based on multi-view alignment | |
| Qin et al. | A machine learning methodology for diagnosing chronic kidney disease | |
| CN109785976A (en) | A kind of goat based on Soft-Voting forecasting system by stages | |
| Wu et al. | A novel method for classification of ECG arrhythmias using deep belief networks | |
| WO2021143780A1 (en) | Multi-label learning-based post-physical examination chronic disease prognosis system | |
| CN110503635B (en) | Hand bone X-ray film bone age assessment method based on heterogeneous data fusion network | |
| CN111798954A (en) | Drug combination recommendation method based on time attention mechanism and graph convolution network | |
| CN111248883B (en) | Blood pressure prediction method and device | |
| CN109192305A (en) | A kind of heart function automatic analysis method based on deep-cycle neural network | |
| Chen et al. | A machine learning method correlating pulse pressure wave data with pregnancy | |
| CN113270192A (en) | Hemodialysis risk prediction system based on combination of dynamic and static data and depth self-encoder | |
| CN115579141A (en) | Interpretable disease risk prediction model construction method and disease risk prediction device | |
| CN113096818A (en) | ODE and GRUD-based method for evaluating incidence of acute diseases | |
| Sheikhalishahi et al. | Benchmarking machine learning models on eICU critical care dataset | |
| CN117034142B (en) | Unbalanced medical data missing value filling method and system | |
| Barakat et al. | An HPC-driven data science platform to speed-up time series data analysis of patients with the acute respiratory distress syndrome | |
| CN112201348B (en) | Knowledge-aware-based multi-center clinical data set adaptation device | |
| CN109119155A (en) | ICU mortality prediction assessment system based on deep learning | |
| CN115547502B (en) | Hemodialysis patient risk prediction device based on time sequence data | |
| CN114504298B (en) | Physiological characteristic discriminating method and system based on multisource health perception data fusion | |
| CN115376638A (en) | Physiological characteristic data analysis method based on multi-source health perception data fusion | |
| Fernández et al. | Artificial neural networks applications in dialysis | |
| CN113571190A (en) | Lung function decline prediction device and prediction method thereof | |
| CN113822420A (en) | Dead space fraction model building method and system based on volumetric carbon dioxide map | |
| CN115512833B (en) | Establishment of long-term cost effectiveness prediction system for lung cancer patient based on deep learning Markov framework |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |