Multi-modal hemodynamic analysis method and device applied to sepsis

Abstract

The invention relates to the technical field of data processing, and discloses a multi-modal hemodynamic analysis method and device applied to sepsis, and the method comprises the steps: monitoring multi-modal data of a plurality of sepsis patients; according to sepsis data in each view included in the multi-modal data of each sepsis patient, constructing an adjacent matrix of each view; based on a dynamic fusion algorithm of a preset multi-view map clustering model, performing embedding processing on the adjacent matrixes of all views to obtain a multi-view embedding matrix; based on a preset clustering algorithm, according to all the multi-view map embedding matrixes, clustering analysis is carried out on all the sepsis patients to obtain a clustering result, and all the sepsis patients in each sepsis patient group included in the clustering result correspond to one hemodynamic feature. Therefore, by implementing the method, the classification accuracy of the sepsis patients can be improved based on diversified monitoring data, and the analysis accuracy of the hemodynamic characteristics of each type of patients can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a multimodal hemodynamic analysis method and device applied to sepsis. Background Art

[0002] Sepsis can be caused by infection in any part of the body, such as the lungs, abdomen, urinary system, etc. Symptoms of sepsis may include fever, chills, rapid breathing, increased heart rate, decreased blood pressure, changes in consciousness, etc., that is, the symptoms of sepsis are complex and diverse.

[0003] However, in the existing technology, in the management of sepsis, the hemodynamic analysis is usually performed simply by monitoring the function and state of the patient's left ventricle through critical care echocardiography (CCE). The monitored data is relatively single and incomplete, which easily leads to low classification accuracy of sepsis patients and low analysis accuracy of their hemodynamic characteristics, thereby affecting the subsequent treatment effect. It can be seen that it is particularly important to propose a new technical solution for hemodynamic data analysis of sepsis. Summary of the invention

[0004] The present invention provides a multimodal hemodynamic analysis method and device for sepsis, which can improve the classification accuracy of sepsis patients based on diversified monitoring data, and is conducive to improving the analysis accuracy of the hemodynamic characteristics of each type of patient.

[0005] In order to solve the above technical problems, the first aspect of the present invention discloses a multimodal hemodynamic analysis method for sepsis, the method comprising:

[0006] Monitoring multimodal data of each of the plurality of sepsis patients, wherein the multimodal data includes two or more of medical indicator data, diagnostic text data, and imaging data data;

[0007] constructing an adjacency matrix of each of the views according to the sepsis data in each of the multiple views included in the multimodal data of each of the sepsis patients;

[0008] Based on a dynamic fusion algorithm of a preset multi-view spectral clustering model, the adjacency matrices of all the views are embedded to obtain a multi-view embedding matrix;

[0009] Based on a preset clustering algorithm, cluster analysis is performed on all the sepsis patients according to all the multi-view spectrum embedding matrices to obtain a clustering result, wherein the clustering result includes multiple sepsis patient groups, and all sepsis patients in each sepsis patient group correspond to a hemodynamic feature; the hemodynamic feature is used as a basis for generating a treatment strategy for all sepsis patients in the corresponding sepsis patient group.

[0010] As an optional implementation, in the first aspect of the present invention, constructing an adjacency matrix of each view according to the sepsis data in each of the multiple views included in the multimodal data of each sepsis patient comprises:

[0011] According to a preset deep neural network model, feature extraction is performed on the sepsis data in each of the multiple views contained in the multimodal data of each sepsis patient to obtain feature data of each sepsis patient in each of the views;

[0012] Based on a preset nearest neighbor analysis algorithm, and according to the feature data of all the sepsis patients in each of the views, a sample set corresponding to each of the sepsis patients is constructed, wherein each sample in the sample set corresponds to a sample type, and the sample type includes a positive sample type or a negative sample type;

[0013] According to the characteristic data of the sample set corresponding to each of the sepsis patients in each of the views, optimizing the mapping function included in the pre-constructed target network, wherein the target network is a twin network with the same network branches;

[0014] The adjacency matrix of each of the views is calculated according to the optimized mapping function and the feature data of the sample set corresponding to each of the sepsis patients in each of the views.

[0015] As an optional implementation, in the first aspect of the present invention, the method based on a preset nearest neighbor analysis algorithm, according to the feature data of all the sepsis patients in each of the views, constructs a sample set corresponding to each of the sepsis patients, including:

[0016] Based on a preset distance measurement algorithm, calculating the similarity between each of the sepsis patients according to the feature data of all the sepsis patients in each of the views;

[0017] Based on a preset nearest neighbor analysis algorithm, and according to the similarities between all the sepsis patients, a neighbor set corresponding to each of the sepsis patients is constructed;

[0018] According to the determined sample type of each sepsis patient, a sample set corresponding to each sepsis patient is determined.

[0019] As an optional implementation, in the first aspect of the present invention, determining the sample set corresponding to each of the sepsis patients according to the determined sample type of each of the sepsis patients comprises:

[0020] According to the determined sample type of each of the sepsis patients, samples having the same sample type as that of the sepsis patient are determined from a neighbor set corresponding to each of the sepsis patients as a positive sample set corresponding to the sepsis patient;

[0021] According to the sample type of each sepsis patient, determine, from the neighbor set corresponding to each sepsis patient, samples that are different from the sample type of the sepsis patient as a negative sample set corresponding to the sepsis patient;

[0022] A positive sample set corresponding to each of the sepsis patients and a negative sample set corresponding to the sepsis patient are determined as the sample set corresponding to the sepsis patient.

[0023] As an optional implementation, in the first aspect of the present invention, optimizing the mapping function included in the pre-constructed target network according to the feature data of the sample set corresponding to each of the sepsis patients in each of the views comprises:

[0024] For each of the sepsis patients, the sample set is divided according to the sample types of all samples contained in the sample set corresponding to the sepsis patient to obtain multiple sample groups, each of the sample groups has a corresponding discriminant label, and the discriminant label includes a label that the corresponding sample groups are positive samples of each other or a label that the corresponding sample groups are negative samples of each other;

[0025] Inputting feature data of each sample group in each view into a pre-constructed target network, and outputting a feature vector of the sample group through all network branches included in the target network;

[0026] Calculating the contrast loss of the target network according to the feature vectors of all the sample groups, and optimizing the contrast loss of the target network to obtain an optimized contrast loss;

[0027] According to the optimized contrast loss, the mapping function included in the target network is optimized.

[0028] As an optional implementation, in the first aspect of the present invention, the multi-view spectral clustering model includes a plurality of deep neural network layers, and each of the deep neural network layers includes one of a local learning layer, a global learning layer and an orthogonal constraint layer;

[0029] And, the dynamic fusion algorithm based on the preset multi-view spectral clustering model embeds the adjacency matrices of all the views to obtain a multi-view embedding matrix, including:

[0030] Based on the local learning layer, the sepsis data of each sepsis patient in each view and the adjacency matrix of each view are processed to obtain single view feature information of each view;

[0031] Based on the global learning layer, single view feature information of all the views is processed to obtain a spectral embedding matrix of each of the views;

[0032] Based on the orthogonal constraint layer, the spectrum embedding matrices of all the views are orthogonalized to obtain a multi-view spectrum embedding matrix.

[0033] As an optional embodiment, in the first aspect of the present invention, the method further comprises:

[0034] In the process of learning using the local learning layer and the global learning layer, a local learning loss of the local learning layer is calculated according to the acquired learning data of the local learning layer; and a global learning loss of the global learning layer is calculated according to the acquired learning data of the global learning layer;

[0035] Calculating the overall loss of the multi-view spectral clustering model according to the local learning loss and the global learning loss;

[0036] According to the overall loss, evaluating the performance index of the multi-view spectrum clustering model; and determining whether the performance index reaches a preset standard performance index;

[0037] When it is determined that the performance index does not reach the standard performance index, the overall loss is optimized to obtain an optimized overall loss, until the performance index of the multi-view spectrum clustering model reaches a preset standard performance index;

[0038] When it is determined that the performance indicator reaches the standard performance indicator, it is determined that the learning of the local learning layer and the global learning layer is completed.

[0039] A second aspect of the present invention discloses a multimodal hemodynamic analysis device for sepsis, the device comprising:

[0040] A monitoring module, used to monitor multimodal data of each of the multiple sepsis patients, wherein the multimodal data includes two or more of medical indicator data, diagnostic text data and imaging data;

[0041] A construction module, configured to construct an adjacency matrix of each of the views according to the sepsis data in each of the multiple views included in the multimodal data of each of the sepsis patients;

[0042] A processing module, configured to embed the adjacency matrices of all the views based on a dynamic fusion algorithm of a preset multi-view spectral clustering model to obtain a multi-view embedding matrix;

[0043] A clustering module is used to perform cluster analysis on all the sepsis patients based on a preset clustering algorithm and according to all the multi-view spectrum embedding matrices to obtain a clustering result, wherein the clustering result includes multiple sepsis patient groups, and all sepsis patients in each sepsis patient group correspond to a hemodynamic feature; the hemodynamic feature is used as a basis for generating a treatment strategy for all sepsis patients in the corresponding sepsis patient group.

[0044] As an optional implementation, in the second aspect of the present invention, the construction module constructs the adjacency matrix of each view according to the sepsis data in each of the multiple views included in the multimodal data of each sepsis patient, specifically including:

[0045] According to a preset deep neural network model, feature extraction is performed on the sepsis data in each of the multiple views contained in the multimodal data of each sepsis patient to obtain feature data of each sepsis patient in each of the views;

[0046] Based on a preset nearest neighbor analysis algorithm, and according to the feature data of all the sepsis patients in each of the views, a sample set corresponding to each of the sepsis patients is constructed, wherein each sample in the sample set corresponds to a sample type, and the sample type includes a positive sample type or a negative sample type;

[0047] According to the characteristic data of the sample set corresponding to each of the sepsis patients in each of the views, optimizing the mapping function included in the pre-constructed target network, wherein the target network is a twin network with the same network branches;

[0048] The adjacency matrix of each of the views is calculated according to the optimized mapping function and the feature data of the sample set corresponding to each of the sepsis patients in each of the views.

[0049] As an optional implementation, in the second aspect of the present invention, the construction module constructs a sample set corresponding to each of the sepsis patients based on a preset nearest neighbor analysis algorithm according to the feature data of all the sepsis patients in each of the views, specifically comprising:

[0050] Based on a preset distance measurement algorithm, calculating the similarity between each of the sepsis patients according to the feature data of all the sepsis patients in each of the views;

[0051] Based on a preset nearest neighbor analysis algorithm, and according to the similarities between all the sepsis patients, a neighbor set corresponding to each of the sepsis patients is constructed;

[0052] According to the determined sample type of each sepsis patient, a sample set corresponding to each sepsis patient is determined.

[0053] As an optional implementation, in the second aspect of the present invention, the construction module determines the sample set corresponding to each of the sepsis patients according to the determined sample type of each of the sepsis patients, specifically including:

[0054] According to the determined sample type of each of the sepsis patients, samples having the same sample type as that of the sepsis patient are determined from a neighbor set corresponding to each of the sepsis patients as a positive sample set corresponding to the sepsis patient;

[0055] According to the sample type of each sepsis patient, determine, from the neighbor set corresponding to each sepsis patient, samples that are different from the sample type of the sepsis patient as a negative sample set corresponding to the sepsis patient;

[0056] A positive sample set corresponding to each of the sepsis patients and a negative sample set corresponding to the sepsis patient are determined as the sample set corresponding to the sepsis patient.

[0057] As an optional implementation, in the second aspect of the present invention, the construction module optimizes the mapping function included in the pre-constructed target network according to the feature data of the sample set corresponding to each of the sepsis patients in each of the views, and specifically includes:

[0058] For each of the sepsis patients, the sample set is divided according to the sample types of all samples contained in the sample set corresponding to the sepsis patient to obtain multiple sample groups, each of the sample groups has a corresponding discriminant label, and the discriminant label includes a label that the corresponding sample groups are positive samples of each other or a label that the corresponding sample groups are negative samples of each other;

[0059] Inputting feature data of each sample group in each view into a pre-constructed target network, and outputting a feature vector of the sample group through all network branches included in the target network;

[0060] Calculating the contrast loss of the target network according to the feature vectors of all the sample groups, and optimizing the contrast loss of the target network to obtain an optimized contrast loss;

[0061] According to the optimized contrast loss, the mapping function included in the target network is optimized.

[0062] As an optional implementation, in the second aspect of the present invention, the multi-view spectral clustering model includes a plurality of deep neural network layers, and each of the deep neural network layers includes one of a local learning layer, a global learning layer and an orthogonal constraint layer;

[0063] Furthermore, the processing module embeds the adjacency matrices of all the views based on the dynamic fusion algorithm of the preset multi-view spectral clustering model, and the method of obtaining the multi-view embedding matrix specifically includes:

[0064] Based on the local learning layer, the sepsis data of each sepsis patient in each view and the adjacency matrix of each view are processed to obtain single view feature information of each view;

[0065] Based on the global learning layer, single view feature information of all the views is processed to obtain a spectral embedding matrix of each of the views;

[0066] Based on the orthogonal constraint layer, the spectrum embedding matrices of all the views are orthogonalized to obtain a multi-view spectrum embedding matrix.

[0067] As an optional implementation, in the second aspect of the present invention, the device further includes:

[0068] A calculation module, used for calculating the local learning loss of the local learning layer according to the acquired learning data of the local learning layer in the process of learning using the local learning layer and the global learning layer; and calculating the global learning loss of the global learning layer according to the acquired learning data of the global learning layer;

[0069] The calculation module is further used to calculate the overall loss of the multi-view spectral clustering model according to the local learning loss and the global learning loss;

[0070] An evaluation module, used to evaluate the performance index of the multi-view spectral clustering model according to the overall loss;

[0071] A judgment module, used to judge whether the performance index reaches a preset standard performance index;

[0072] An optimization module, configured to optimize the overall loss to obtain an optimized overall loss when the judgment module determines that the performance index does not reach the standard performance index, until the performance index of the multi-view spectrum clustering model reaches a preset standard performance index;

[0073] The determination module is used to determine that the learning of the local learning layer and the global learning layer is completed when the judgment module determines that the performance indicator reaches the standard performance indicator.

[0074] The third aspect of the present invention discloses another multimodal hemodynamic analysis device for sepsis, the device comprising:

[0075] A memory storing executable program code;

[0076] a processor coupled to the memory;

[0077] The processor calls the executable program code stored in the memory to execute the multimodal hemodynamic analysis method for sepsis disclosed in the first aspect of the present invention.

[0078] A fourth aspect of the present invention discloses a computer storage medium, wherein the computer storage medium stores computer instructions, and when the computer instructions are called, they are used to execute the multimodal hemodynamic analysis method for sepsis disclosed in the first aspect of the present invention.

[0079] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0080] In an embodiment of the present invention, multimodal data of each sepsis patient among multiple sepsis patients is monitored, and the multimodal data includes two or more of medical indicator data, diagnostic text data and imaging data data; an adjacency matrix of each view is constructed according to the sepsis data in each view in multiple views contained in the multimodal data of each sepsis patient; based on a dynamic fusion algorithm of a preset multi-view spectrum clustering model, the adjacency matrices of all views are embedded to obtain a multi-view embedding matrix; based on a preset clustering algorithm, all sepsis patients are clustered according to all multi-view spectrum embedding matrices to obtain clustering results, and the clustering results include multiple sepsis patient groups, and all sepsis patients in each sepsis patient group correspond to a hemodynamic feature; the hemodynamic feature is used as a basis for generating a treatment strategy for all sepsis patients in the corresponding sepsis patient group. It can be seen that the implementation of the present invention can monitor the multimodal data of each sepsis patient among multiple sepsis patients, such as two or more of medical indicator data, diagnostic text data and imaging data, and construct an adjacency matrix of each view according to the sepsis data in each view in the multiple views contained in the multimodal data of each sepsis patient, which can improve the accuracy and efficiency of constructing the adjacency matrix of each view. Subsequently, based on the dynamic fusion algorithm of the preset multi-view spectrum clustering model, the adjacency matrices of all views are embedded to obtain a multi-view embedding matrix, which can improve the accuracy and reliability of the multi-view embedding matrix obtained by processing; based on the preset clustering algorithm, all sepsis patients are clustered and analyzed according to all multi-view spectrum embedding matrices to obtain clustering results, so that all sepsis patients in each sepsis patient group included in the clustering results correspond to a hemodynamic feature, which can improve the classification accuracy of sepsis patients based on multimodal data, which is conducive to improving the analysis accuracy of the hemodynamic features of each type of sepsis patients, thereby facilitating the generation accuracy of the treatment strategy for each type of patients based on the accurately analyzed hemodynamic features, and further facilitating achieving a good treatment effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0082] Figure 1 It is a flow chart of a multimodal hemodynamic analysis method for sepsis disclosed in an embodiment of the present invention;

[0083] Figure 2is a flow chart of another multimodal hemodynamic analysis method for sepsis disclosed in an embodiment of the present invention;

[0084] Figure 3 is a schematic structural diagram of a multimodal hemodynamic analysis device for sepsis disclosed in an embodiment of the present invention;

[0085] Figure 4 is a schematic structural diagram of another multimodal hemodynamic analysis device for sepsis disclosed in an embodiment of the present invention;

[0086] Figure 5 It is a schematic structural diagram of another multimodal hemodynamic analysis device for sepsis disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0087] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0088] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product or end including a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or ends.

[0089] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0090] The present invention discloses a multimodal hemodynamic analysis method and device for sepsis, which can monitor the multimodal data of each sepsis patient among multiple sepsis patients, such as two or more of medical indicator data, diagnostic text data and imaging data data, and construct an adjacency matrix of each view according to the sepsis data in each view in multiple views contained in the multimodal data of each sepsis patient, which can improve the construction accuracy and efficiency of the adjacency matrix of each view, and then embed the adjacency matrices of all views based on the dynamic fusion algorithm of the preset multi-view spectrum clustering model to obtain a multi-view embedding matrix, which can Improve the accuracy and reliability of the multi-view embedding matrix obtained by processing; then based on the preset clustering algorithm, perform cluster analysis on all sepsis patients according to all multi-view spectrum embedding matrices to obtain clustering results, so that all sepsis patients in each sepsis patient group included in the clustering results correspond to a hemodynamic feature, which can improve the classification accuracy of sepsis patients based on multimodal data, and is conducive to improving the analysis accuracy of the hemodynamic features of each type of sepsis patient, thereby facilitating the generation accuracy of the treatment strategy for each type of patient based on the accurately analyzed hemodynamic features, and thus facilitating achieving good treatment results. The following are detailed descriptions.

[0091] Embodiment 1

[0092] See also Figure 1 , Figure 1 : is a flow chart of a multimodal hemodynamic analysis method for sepsis disclosed in an embodiment of the present invention. Figure 1 The multimodal hemodynamic analysis method for sepsis described above can be applied to a multimodal hemodynamic analysis device for sepsis, wherein the device may include an analysis device or an analysis server, wherein the analysis server may include a cloud server or a local server, which is not limited in the embodiments of the present invention. Figure 1 As shown, the multimodal hemodynamic analysis method applied to sepsis may include the following operations:

[0093] 101.Monitoring multimodal data of each septic patient in multiple septic patients.

[0094] In the embodiment of the present invention, optionally, the multimodal data may include two or more of medical indicator data, diagnostic text data and imaging data.

[0095] Among them, medical indicator data refers to the numerical data obtained by physiological, biochemical, and immune tests on sepsis patients, such as heart rate, blood pressure, white blood cell count, blood lactate level, etc. These indicators can reflect the hemodynamic status and organ function of sepsis patients, and are an important basis for evaluating the severity of the disease and the treatment effect. Diagnostic text data refers to the text data recorded by doctors for clinical diagnosis and treatment plan formulation of sepsis patients, such as medical records, prescriptions, and examination reports. These texts can reflect the cause, symptoms, complications, drug reactions, and other information of sepsis patients, and are an important basis for understanding individual differences in patients and formulating personalized treatment plans. Imaging data data refers to image or video data obtained by imaging examinations of sepsis patients, such as X-rays, ultrasounds, CT scans, etc. These data can reflect the anatomical structure and functional changes of sepsis patients, and are an important basis for observing the degree of organ damage and treatment effects of patients. In this way, multimodal data can provide more comprehensive and integrated information, reflecting the condition and treatment effect of sepsis patients from different angles and levels, which helps to improve the accuracy and robustness of clustering; and can make up for the deficiencies and omissions of single-modal data, for example, medical indicators may have measurement errors or lack of sensitivity, diagnostic texts may have subjective biases or unclear expressions, and imaging data may have noise or low resolution. Through the fusion of multimodal data, the complementary and synergistic effects between different modalities can be utilized to improve the quality and credibility of data.

[0096] 102. Construct an adjacency matrix of each view according to the sepsis data in each view of the multiple views included in the multimodal data of each sepsis patient.

[0097] 103. Based on the dynamic fusion algorithm of the preset multi-view spectral clustering model, the adjacency matrices of all views are embedded to obtain a multi-view embedding matrix.

[0098] In an embodiment of the present invention, the multi-view spectral clustering model includes multiple deep neural network layers, and each deep neural network layer includes one of a local learning layer, a global learning layer and an orthogonal constraint layer.

[0099] Specifically, first, the data of each view is processed by the local learning layer to learn the unique embedding of each view and retain the local invariance of each view; secondly, the unique embedding of all views is processed by the global learning layer to learn the shared embedding between multiple views and explore the global consistency and complementarity between multiple views; finally, the output of the global learning layer is processed by the orthogonal constraint layer to ensure the orthogonality of the multi-view spectral embedding matrix and obtain the consistent spectral embedding. The mathematical expression can be Among them, Y N×c =[Y 1 ,Y 2 ,...,YV ];in, is the data of the vth view in the Nth sample, is the mapping function of the multi-view spectral clustering model, Y N×c It is the consistent spectrum embedding matrix obtained after mapping by the multi-view spectrum clustering model. Each row represents a data sample (e.g., the Nth row represents the Nth data sample), and each column represents a cluster (e.g., the cth column represents the cth cluster, c=1,2,...,v). This matrix can be used to cluster data, that is, each sample will be assigned to a cluster, which is the column where its maximum value is located in the matrix. The generalization ability and adaptability of the multi-view spectrum clustering model can be enhanced by the adjacency matrix based on multimodal data. Through the fusion of multimodal data, the multi-view spectrum clustering model can be more easily adapted to different scenarios and environments, and the portability and universality of the multi-view spectrum clustering model can be improved.

[0100] 104. Based on the preset clustering algorithm, all sepsis patients are clustered and analyzed according to the embedding matrix of all multi-view spectra to obtain the clustering results.

[0101] In an embodiment of the present invention, the clustering result may include multiple sepsis patient groups, and all sepsis patients in each sepsis patient group correspond to a hemodynamic feature; the hemodynamic feature is used as a basis for generating a treatment strategy for all sepsis patients in the corresponding sepsis patient group.

[0102] It can be seen that the implementation Figure 1 The multimodal hemodynamic analysis method for sepsis described herein can monitor the multimodal data of each sepsis patient among multiple sepsis patients, such as two or more of medical indicator data, diagnostic text data, and imaging data data, and construct an adjacency matrix of each view according to the sepsis data in each view in multiple views contained in the multimodal data of each sepsis patient, thereby improving the accuracy and efficiency of constructing the adjacency matrix of each view, and then embedding the adjacency matrices of all views based on the dynamic fusion algorithm of the preset multi-view spectral clustering model to obtain a multi-view embedding matrix, thereby improving the processing accuracy and efficiency. The accuracy and reliability of the multi-view embedding matrix are obtained; then based on the preset clustering algorithm, all sepsis patients are clustered and analyzed according to all multi-view spectrum embedding matrices to obtain clustering results, so that all sepsis patients in each sepsis patient group included in the clustering results correspond to a hemodynamic feature, which can improve the classification accuracy of sepsis patients based on multimodal data, and is beneficial to improving the analysis accuracy of the hemodynamic features of each type of sepsis patient, thereby facilitating the generation accuracy of the treatment strategy for each type of patient based on the accurately analyzed hemodynamic features, and further facilitating achieving good treatment effects.

[0103] In an optional embodiment, the dynamic fusion algorithm based on the preset multi-view spectral clustering model in step 103 embeds the adjacency matrices of all views to obtain a multi-view embedding matrix, which may include:

[0104] Based on the local learning layer, the sepsis data of each sepsis patient in each view and the adjacency matrix of each view are processed to obtain the single view feature information of each view;

[0105] Based on the global learning layer, the single-view feature information of all views is processed to obtain the spectral embedding matrix of each view;

[0106] Based on the orthogonal constraint layer, the spectral embedding matrices of all views are orthogonalized to obtain the multi-view spectral embedding matrix.

[0107] In an embodiment of the present invention, the local learning layer can adapt to different forms of data and learn the unique embedding of each view to extract the unique features of a single view. Specifically, the local learning layer uses a fully connected layer and a ReLU function to perform a nonlinear transformation on the original data of each view, thereby obtaining a new single view feature representation.

[0108] In the global learning layer, parameter sharing and feature transfer methods can be used to learn the consistent features and complementary features of multiple views, thereby obtaining a consistent spectral embedding that integrates multi-view information.

[0109] Specifically, the parameter sharing method includes cascading the local learning layer outputs of all views, and then using the fully connected layer for learning to embed features for multiple views. The feature migration method includes migrating information from different views through nonlinear mapping and fusing it with the local information of the current view to achieve feature migration between different views and form a new view.

[0110] In the orthogonal constraint layer, the spectral embedding of each view can be orthogonalized using Cholesky decomposition, so that the product of the spectral embedding matrix of each view and its transpose is equal to the identity matrix. This ensures that the feature embedding output by the model meets the orthogonality condition, thereby avoiding invalid solutions.

[0111] It can be seen that this optional embodiment can process the sepsis data of each sepsis patient in each view and the adjacency matrix of each view based on the local learning layer to obtain the single view feature information of each view, which can improve the extraction accuracy and efficiency of the single view features of each view, and based on the global learning layer, the single view feature information of all views is processed to obtain the spectral embedding matrix of each view, which can improve the accuracy and reliability of the processed spectral embedding matrix of each view, and then based on the orthogonal constraint layer, the spectral embedding matrices of all views are orthogonalized to obtain a multi-view spectral embedding matrix, which can improve the accuracy and reliability of the processed multi-view spectral embedding matrix.

[0112] In this optional embodiment, as an optional implementation, the method may further include:

[0113] In the process of learning using the local learning layer and the global learning layer, the local learning loss of the local learning layer is calculated according to the acquired learning data of the local learning layer; and the global learning loss of the global learning layer is calculated according to the acquired learning data of the global learning layer;

[0114] Calculate the overall loss of the multi-view spectral clustering model based on the local learning loss and the global learning loss;

[0115] According to the overall loss, the performance index of the multi-view spectral clustering model is evaluated; and whether the performance index reaches the preset standard performance index is determined;

[0116] When it is determined that the performance index does not reach the standard performance index, the overall loss is optimized to obtain the optimized overall loss until the performance index of the multi-view spectrum clustering model reaches the preset standard performance index;

[0117] When it is determined that the performance index reaches the standard performance index, it is determined that the local learning layer and the global learning layer have completed learning.

[0118] In the embodiment of the present invention, the weight of the loss in each view can be adjusted in the local learning layer and the weight of the loss between multiple views can be adjusted in the global learning layer to optimize the overall loss.

[0119] It can be seen that this optional implementation method can calculate the local learning loss of the local learning layer according to the acquired learning data of the local learning layer, and calculate the global learning loss of the global learning layer according to the acquired learning data of the global learning layer, and calculate the overall loss of the multi-view spectral clustering model according to the local learning loss and the global learning loss, thereby improving the calculation accuracy and reliability of the overall loss of the model. Subsequently, the performance index of the multi-view spectral clustering model is evaluated according to the overall loss, thereby improving the evaluation accuracy and efficiency of the model performance index, and judging whether the performance index reaches the preset standard performance index. When it is judged that the performance index does not reach the standard performance index, the overall loss is optimized to obtain the optimized overall loss until the performance index of the multi-view spectral clustering model reaches the preset standard performance index, which can improve the timeliness and accuracy of the optimization of the overall loss, thereby achieving the accurate achievement of the model performance index. When it is judged that the performance index reaches the standard performance index, it is determined that the local learning layer and the global learning layer have completed learning, and the training accuracy of the local learning layer and the global learning layer contained in the model can be improved through the performance index evaluated based on the overall loss.

[0120] Embodiment 2

[0121] See also Figure 2 , Figure 2 : is a flow chart of a multimodal hemodynamic analysis method for sepsis disclosed in an embodiment of the present invention. Figure 2 The multimodal hemodynamic analysis method for sepsis described above can be applied to a multimodal hemodynamic analysis device for sepsis, wherein the device may include an analysis device or an analysis server, wherein the analysis server may include a cloud server or a local server, which is not limited in the embodiments of the present invention. Figure 2 As shown, the multimodal hemodynamic analysis method applied to sepsis may include the following operations:

[0122] 201.Monitoring multimodal data of each septic patient in multiple septic patients.

[0123] 202. According to a preset deep neural network model, feature extraction is performed on the sepsis data in each of the multiple views contained in the multimodal data of each sepsis patient to obtain feature data of each sepsis patient in each view.

[0124] In the embodiment of the present invention, specifically, for the sepsis data in each of the multiple views included in the multimodal data of each sepsis patient, feature extraction is performed on the sepsis data according to a preset deep neural network model corresponding to the data type of the sepsis data to obtain feature data in the view. The data type of each sepsis data may include one of a medical indicator type, a diagnostic text data type, and an image data data type.

[0125] 203. Based on the preset nearest neighbor analysis algorithm, a sample set corresponding to each sepsis patient is constructed according to the feature data of all sepsis patients in each view.

[0126] In an embodiment of the present invention, each sample in the sample set corresponds to a sample type, and the sample type includes a positive sample type or a negative sample type. The sample set corresponding to each sepsis patient may include a positive sample set corresponding to each sepsis patient and a negative sample set corresponding to each sepsis patient. All samples in the positive sample set are mutually positive samples, and all samples in the negative sample set are mutually negative samples.

[0127] 204. According to the feature data of the sample set corresponding to each sepsis patient in each view, the mapping function included in the pre-constructed target network is optimized.

[0128] In the embodiment of the present invention, the target network is a twin network with the same network branches. Specifically, the pre-constructed target network is subjected to nearest neighbor learning through contrast loss to optimize the mapping function in the target network.

[0129] 205. Calculate the adjacency matrix of each view according to the optimized mapping function and the feature data of the sample set corresponding to each sepsis patient in each view.

[0130] In the embodiment of the present invention, specifically, a Gaussian kernel function is used to calculate the final adjacency matrix, where δ is a width parameter in the Gaussian kernel function, and δ is usually greater than 0. The Gaussian kernel function is used to map data to a high-dimensional feature space and then calculate the distance between samples. The calculation formula of the adjacency matrix is ​​as follows:

[0131] Neighbors

[0132] in, represents the feature data of the i-th sepsis patient in the v-th view, represents the mapping function in the twin network, which is used to embed sample data into the latent space; the calculated adjacency matrix can be expressed as [S 1 ,S 2 ,...,S v] for spectral embedding learning based on knowledge transfer.

[0133] Based on the dynamic fusion algorithm of the preset multi-view spectral clustering model, the adjacency matrices of all views are embedded to obtain a multi-view embedding matrix.

[0134] 207. Based on the preset clustering algorithm, all sepsis patients are clustered and analyzed according to all multi-view spectrum embedding matrices to obtain clustering results.

[0135] In the embodiment of the present invention, for other descriptions of step 201, step 206 and step 207, please refer to the detailed description of step 101, step 103 and step 104 in the first embodiment, and the embodiment of the present invention will not be repeated.

[0136] It can be seen that the implementation Figure 2The multimodal hemodynamic analysis method for sepsis described herein can monitor the multimodal data of each sepsis patient among multiple sepsis patients, such as two or more of medical indicator data, diagnostic text data, and imaging data data, and construct an adjacency matrix of each view according to the sepsis data in each view in multiple views contained in the multimodal data of each sepsis patient, thereby improving the accuracy and efficiency of constructing the adjacency matrix of each view, and then embedding the adjacency matrices of all views based on the dynamic fusion algorithm of the preset multi-view spectral clustering model to obtain a multi-view embedding matrix, thereby improving the processing accuracy and efficiency. The accuracy and reliability of the multi-view embedding matrix are obtained; then based on the preset clustering algorithm, all sepsis patients are clustered and analyzed according to all multi-view spectrum embedding matrices to obtain clustering results, so that all sepsis patients in each sepsis patient group included in the clustering results correspond to a hemodynamic feature, which can improve the classification accuracy of sepsis patients based on multimodal data, and is beneficial to improving the analysis accuracy of the hemodynamic features of each type of sepsis patient, thereby facilitating the generation accuracy of the treatment strategy for each type of patient based on the accurately analyzed hemodynamic features, and further facilitating achieving good treatment effects. In addition, according to the preset deep neural network model, feature extraction can be performed on the sepsis data in each view of multiple views contained in the multimodal data of each sepsis patient to obtain the feature data of each sepsis patient in each view, thereby improving the accuracy and efficiency of extracting the feature data of each view, and based on the preset nearest neighbor analysis algorithm, a sample set corresponding to each sepsis patient is constructed according to the feature data of all sepsis patients in each view, thereby improving the accuracy and efficiency of constructing the sample set corresponding to each sepsis patient, and according to the feature data of the sample set corresponding to each sepsis patient in each view, the mapping function contained in the pre-constructed target network is optimized, and the mapping function in the target network can be accurately optimized by using the output features obtained by using the feature data as the input of the target network, and then the adjacency matrix of each view is calculated according to the optimized mapping function and the feature data of the sample set corresponding to each sepsis patient in each view, thereby improving the calculation accuracy and reliability of the adjacency matrix of each view.

[0137] In an optional embodiment, the aforementioned step 203, based on a preset nearest neighbor analysis algorithm, constructs a sample set corresponding to each sepsis patient according to the feature data of all sepsis patients in each view, which may include:

[0138] Based on a preset distance measurement algorithm, the similarity between each sepsis patient is calculated according to the feature data of all sepsis patients in each view;

[0139] Based on the preset nearest neighbor analysis algorithm, a neighbor set corresponding to each sepsis patient is constructed according to the similarity between all sepsis patients;

[0140] According to the determined sample type of each sepsis patient, a sample set corresponding to each sepsis patient is determined.

[0141] In the embodiment of the present invention, optionally, the distance measurement algorithm may include a Euclidean distance algorithm, a cosine similarity algorithm, or any other algorithm that can equally calculate the distance / similarity between each sepsis patient. The k nearest neighbors of each sepsis patient can be found through the nearest neighbor analysis algorithm as the neighbor set corresponding to the sepsis patient, which is not limited in the embodiment of the present invention.

[0142] It can be seen that this optional embodiment can calculate the similarity between each sepsis patient based on the preset distance measurement algorithm and the feature data of all sepsis patients in each view, thereby improving the calculation accuracy of the similarity between each sepsis patient. Subsequently, based on the preset nearest neighbor analysis algorithm, a neighbor set corresponding to each sepsis patient is constructed according to the similarity between all sepsis patients, which can improve the construction accuracy and reliability of the neighbor set corresponding to each sepsis patient. Then, based on the determined sample type of each sepsis patient, the sample set corresponding to each sepsis patient is determined, thereby improving the determination accuracy of the sample set used as the input data sample of the twin network, which is beneficial to improving the subsequent optimization accuracy of the mapping function of the twin network.

[0143] In this optional embodiment, as an optional implementation, determining the sample set corresponding to each sepsis patient according to the determined sample type of each sepsis patient may include:

[0144] According to the determined sample type of each sepsis patient, samples having the same sample type as the sepsis patient are determined from the neighbor set corresponding to each sepsis patient as the positive sample set corresponding to the sepsis patient;

[0145] According to the sample type of each sepsis patient, samples that are different from the sample type of the sepsis patient are determined from the neighbor set corresponding to each sepsis patient as the negative sample set corresponding to the sepsis patient;

[0146] A positive sample set corresponding to each sepsis patient and a negative sample set corresponding to the sepsis patient are determined as the sample set corresponding to the sepsis patient.

[0147] In an embodiment of the present invention, specifically, if most of the neighbors of a sample are positive samples, then the sample is also determined to be a positive sample; if most of the neighbors of a sample are negative samples, then the sample is determined to be a negative sample; for samples of uncertain sample types, a probabilistic division strategy can be adopted, such as determining the probability of belonging to the positive / negative class based on the ratio of positive and negative neighbors.

[0148] It can be seen that this optional implementation can determine, according to the determined sample type of each sepsis patient, samples of the same sample type as the sepsis patient from the neighbor set corresponding to each sepsis patient as the positive sample set corresponding to the sepsis patient, thereby improving the determination accuracy of the positive sample set corresponding to each sepsis patient, and according to the sample type of each sepsis patient, determine, from the neighbor set corresponding to each sepsis patient, samples of a different sample type from the sepsis patient as the negative sample set corresponding to the sepsis patient, thereby improving the determination accuracy of the negative sample set corresponding to each sepsis patient, and then determine the positive sample set corresponding to each sepsis patient and the negative sample set corresponding to the sepsis patient as the sample set corresponding to the sepsis patient, thereby improving the determination accuracy and reliability of the sample set corresponding to each sepsis patient.

[0149] In another optional embodiment, the above step 204, based on the feature data of the sample set corresponding to each sepsis patient in each view, optimizes the mapping function included in the pre-constructed target network, which may include:

[0150] For each sepsis patient, the sample set is divided according to the sample types of all samples contained in the sample set corresponding to the sepsis patient to obtain multiple sample groups;

[0151] Input the feature data of each sample group in each view into the pre-built target network, and output the feature vector of the sample group through all network branches contained in the target network;

[0152] According to the feature vectors of all sample groups, the contrast loss of the target network is calculated, and the contrast loss of the target network is optimized to obtain the optimized contrast loss;

[0153] According to the optimized contrast loss, the mapping function contained in the target network is optimized.

[0154] In the embodiment of the present invention, specifically, the contrast loss can be minimized to optimize the above contrast loss. Each sample group has a corresponding discriminant label, and the discriminant label includes a label that the corresponding sample groups are positive samples of each other or a label that the corresponding sample groups are negative samples of each other. The expression of contrast loss can be specifically:

[0155]

[0156] Among them, P∈{0,1} is the discriminant label, that is, when P=1, it means and They are positive samples of each other, and P = 0 means and are each other’s negative samples; γ refers to the distance boundary (usually defined as 1).

[0157] It can be seen that this optional embodiment can divide the sample set for each sepsis patient according to the sample types of all samples contained in the sample set corresponding to the sepsis patient to obtain multiple sample groups, thereby improving the grouping accuracy of the sample set corresponding to each sepsis patient, and inputting the feature data of each sample group in each view into the pre-constructed target network, and outputting the feature vector of the sample group through all network branches contained in the target network, thereby improving the output accuracy of the sample feature vector based on the network branch, and then calculating the contrast loss of the target network according to the feature vectors of all sample groups, thereby improving the calculation accuracy and reliability of the contrast loss of the target network, and optimizing the contrast loss of the target network to obtain the optimized contrast loss, thereby improving the optimization accuracy and efficiency of the contrast loss, and then optimizing the mapping function contained in the target network according to the optimized contrast loss, thereby facilitating the accurate optimization of the mapping function by optimizing the contrast loss, thereby facilitating improving the accuracy of constructing the adjacency matrix of each subsequent view based on the accurately optimized mapping function.

[0158] Embodiment 3

[0159] See also Figure 3 , Figure 3 : is a schematic diagram of a multimodal hemodynamic analysis device for sepsis disclosed in an embodiment of the present invention. Figure 3 The multimodal hemodynamic analysis device for sepsis described may include an analysis device or an analysis server, wherein the analysis server may include a cloud server or a local server, which is not limited in the embodiment of the present invention. Figure 3 As shown, the multimodal hemodynamic analysis device applied to sepsis may include:

[0160] A monitoring module 301 is used to monitor multimodal data of each sepsis patient among a plurality of sepsis patients, where the multimodal data includes two or more of medical indicator data, diagnostic text data, and imaging data data;

[0161] A construction module 302 is used to construct an adjacency matrix of each view according to the sepsis data in each view of the multiple views included in the multimodal data of each sepsis patient;

[0162] The processing module 303 is used to embed the adjacency matrices of all views based on the dynamic fusion algorithm of the preset multi-view spectral clustering model to obtain a multi-view embedding matrix;

[0163] The clustering module 304 is used to perform cluster analysis on all sepsis patients based on a preset clustering algorithm and according to all multi-view spectrum embedding matrices to obtain clustering results, wherein the clustering results include multiple sepsis patient groups, and all sepsis patients in each sepsis patient group correspond to a hemodynamic feature; the hemodynamic feature is used as a basis for generating a treatment strategy for all sepsis patients in the corresponding sepsis patient group.

[0164] It can be seen that the implementation Figure 3 The multimodal hemodynamic analysis device for sepsis described herein can monitor the multimodal data of each sepsis patient among multiple sepsis patients, such as two or more of medical indicator data, diagnostic text data, and imaging data data, and construct an adjacency matrix of each view according to the sepsis data in each view in multiple views contained in the multimodal data of each sepsis patient, which can improve the accuracy and efficiency of constructing the adjacency matrix of each view, and then embed the adjacency matrices of all views based on the dynamic fusion algorithm of the preset multi-view spectral clustering model to obtain a multi-view embedding matrix, which can improve the processing accuracy and efficiency. The accuracy and reliability of the multi-view embedding matrix are obtained; then based on the preset clustering algorithm, all sepsis patients are clustered and analyzed according to all multi-view spectrum embedding matrices to obtain clustering results, so that all sepsis patients in each sepsis patient group included in the clustering results correspond to a hemodynamic feature, which can improve the classification accuracy of sepsis patients based on multimodal data, and is beneficial to improving the analysis accuracy of the hemodynamic features of each type of sepsis patient, thereby facilitating the generation accuracy of the treatment strategy for each type of patient based on the accurately analyzed hemodynamic features, and further facilitating achieving good treatment effects.

[0165] In an optional embodiment, the construction module 302 constructs the adjacency matrix of each view according to the sepsis data in each view of the multiple views included in the multimodal data of each sepsis patient, and the method may specifically include:

[0166] According to a preset deep neural network model, feature extraction is performed on the sepsis data in each of the multiple views contained in the multimodal data of each sepsis patient to obtain feature data of each sepsis patient in each view;

[0167] Based on the preset nearest neighbor analysis algorithm, a sample set corresponding to each sepsis patient is constructed according to the characteristic data of all sepsis patients in each view, and each sample in the sample set corresponds to a sample type, which includes a positive sample type or a negative sample type;

[0168] According to the characteristic data of the sample set corresponding to each sepsis patient in each view, the mapping function contained in the pre-built target network is optimized, and the target network is a twin network with the same network branches;

[0169] According to the optimized mapping function and the feature data of the sample set corresponding to each sepsis patient in each view, the adjacency matrix of each view is calculated.

[0170] It can be seen that this optional embodiment can perform feature extraction on the sepsis data in each view of multiple views contained in the multimodal data of each sepsis patient according to the preset deep neural network model, obtain the feature data of each sepsis patient in each view, improve the extraction accuracy and efficiency of the feature data of each view, and construct a sample set corresponding to each sepsis patient according to the feature data of all sepsis patients in each view based on the preset nearest neighbor analysis algorithm, improve the construction accuracy and efficiency of the sample set corresponding to each sepsis patient, and optimize the mapping function contained in the pre-constructed target network according to the feature data of the sample set corresponding to each sepsis patient in each view, and can accurately optimize the mapping function in the target network by using the output features obtained by taking the feature data as the input of the target network, and then calculate the adjacency matrix of each view according to the optimized mapping function and the feature data of the sample set corresponding to each sepsis patient in each view, thereby improving the calculation accuracy and reliability of the adjacency matrix of each view.

[0171] In this optional embodiment, as an optional implementation, the construction module 302 constructs a sample set corresponding to each sepsis patient based on the characteristic data of all sepsis patients in each view based on a preset nearest neighbor analysis algorithm, which specifically includes:

[0172] Based on a preset distance measurement algorithm, the similarity between each sepsis patient is calculated according to the feature data of all sepsis patients in each view;

[0173] Based on the preset nearest neighbor analysis algorithm, a neighbor set corresponding to each sepsis patient is constructed according to the similarity between all sepsis patients;

[0174] According to the determined sample type of each sepsis patient, a sample set corresponding to each sepsis patient is determined.

[0175] It can be seen that this optional implementation can calculate the similarity between each sepsis patient based on the preset distance measurement algorithm and the feature data of all sepsis patients in each view, thereby improving the calculation accuracy of the similarity between each sepsis patient. Subsequently, based on the preset nearest neighbor analysis algorithm, a neighbor set corresponding to each sepsis patient is constructed according to the similarity between all sepsis patients, which can improve the construction accuracy and reliability of the neighbor set corresponding to each sepsis patient. Then, based on the determined sample type of each sepsis patient, the sample set corresponding to each sepsis patient is determined, thereby improving the determination accuracy of the sample set used as the input data sample of the twin network, which is beneficial to improving the subsequent optimization accuracy of the mapping function of the twin network.

[0176] In this optional implementation, optionally, the construction module 302 determines the sample set corresponding to each sepsis patient according to the determined sample type of each sepsis patient, specifically including:

[0177] According to the determined sample type of each sepsis patient, samples having the same sample type as the sepsis patient are determined from the neighbor set corresponding to each sepsis patient as the positive sample set corresponding to the sepsis patient;

[0178] According to the sample type of each sepsis patient, samples that are different from the sample type of the sepsis patient are determined from the neighbor set corresponding to each sepsis patient as the negative sample set corresponding to the sepsis patient;

[0179] A positive sample set corresponding to each sepsis patient and a negative sample set corresponding to the sepsis patient are determined as the sample set corresponding to the sepsis patient.

[0180] It can be seen that this optional implementation can also determine, according to the determined sample type of each sepsis patient, samples of the same sample type as the sepsis patient from the neighbor set corresponding to each sepsis patient as the positive sample set corresponding to the sepsis patient, thereby improving the determination accuracy of the positive sample set corresponding to each sepsis patient, and according to the sample type of each sepsis patient, determine, from the neighbor set corresponding to each sepsis patient, samples of a different sample type from the sepsis patient as the negative sample set corresponding to the sepsis patient, thereby improving the determination accuracy of the negative sample set corresponding to each sepsis patient, and then determine the positive sample set corresponding to each sepsis patient and the negative sample set corresponding to the sepsis patient as the sample set corresponding to the sepsis patient, thereby improving the determination accuracy and reliability of the sample set corresponding to each sepsis patient.

[0181] In this optional embodiment, as another optional implementation, the construction module 302 optimizes the mapping function included in the pre-constructed target network according to the feature data of the sample set corresponding to each sepsis patient in each view, and specifically includes:

[0182] For each sepsis patient, the sample set is divided according to the sample types of all samples contained in the sample set corresponding to the sepsis patient to obtain multiple sample groups, each sample group has a corresponding discriminant label, and the discriminant label includes a label that the corresponding sample groups are positive samples for each other or a label that the corresponding sample groups are negative samples for each other;

[0183] Input the feature data of each sample group in each view into the pre-built target network, and output the feature vector of the sample group through all network branches contained in the target network;

[0184] According to the feature vectors of all sample groups, the contrast loss of the target network is calculated, and the contrast loss of the target network is optimized to obtain the optimized contrast loss;

[0185] According to the optimized contrast loss, the mapping function contained in the target network is optimized.

[0186] It can be seen that this optional implementation method can divide the sample set for each sepsis patient according to the sample types of all samples contained in the sample set corresponding to the sepsis patient to obtain multiple sample groups, thereby improving the grouping accuracy of the sample set corresponding to each sepsis patient, and inputting the feature data of each sample group in each view into the pre-constructed target network, and outputting the feature vector of the sample group through all network branches contained in the target network, thereby improving the output accuracy of the sample feature vector based on the network branch, and then calculating the contrast loss of the target network according to the feature vectors of all sample groups, thereby improving the calculation accuracy and reliability of the contrast loss of the target network, and optimizing the contrast loss of the target network to obtain the optimized contrast loss, thereby improving the optimization accuracy and efficiency of the contrast loss, and then optimizing the mapping function contained in the target network according to the optimized contrast loss, thereby facilitating the accurate optimization of the mapping function by optimizing the contrast loss, thereby facilitating improving the accuracy of constructing the adjacency matrix of each subsequent view based on the accurately optimized mapping function.

[0187] In another optional embodiment, the multi-view spectral clustering model includes multiple deep neural network layers, and each deep neural network layer includes one of a local learning layer, a global learning layer, and an orthogonal constraint layer. And, the processing module 303 embeds the adjacency matrix of all views based on the dynamic fusion algorithm of the preset multi-view spectral clustering model, and the method of obtaining the multi-view embedding matrix may specifically include:

[0188] Based on the local learning layer, the sepsis data of each sepsis patient in each view and the adjacency matrix of each view are processed to obtain the single view feature information of each view;

[0189] Based on the global learning layer, the single-view feature information of all views is processed to obtain the spectral embedding matrix of each view;

[0190] Based on the orthogonal constraint layer, the spectral embedding matrices of all views are orthogonalized to obtain the multi-view spectral embedding matrix.

[0191] It can be seen that this optional embodiment can process the sepsis data of each sepsis patient in each view and the adjacency matrix of each view based on the local learning layer to obtain the single view feature information of each view, which can improve the extraction accuracy and efficiency of the single view features of each view, and based on the global learning layer, the single view feature information of all views is processed to obtain the spectral embedding matrix of each view, which can improve the accuracy and reliability of the processed spectral embedding matrix of each view, and then based on the orthogonal constraint layer, the spectral embedding matrices of all views are orthogonalized to obtain a multi-view spectral embedding matrix, which can improve the accuracy and reliability of the processed multi-view spectral embedding matrix.

[0192] In this optional embodiment, as an optional implementation, as Figure 4 As shown, Figure 4 is a schematic diagram of the structure of another multimodal hemodynamic analysis device for sepsis disclosed in an embodiment of the present invention, wherein the device may also include:

[0193] The calculation module 305 is used to calculate the local learning loss of the local learning layer according to the acquired learning data of the local learning layer during the learning process using the local learning layer and the global learning layer; and calculate the global learning loss of the global learning layer according to the acquired learning data of the global learning layer.

[0194] The calculation module 305 is further used to calculate the overall loss of the multi-view spectrum clustering model according to the local learning loss and the global learning loss.

[0195] The evaluation module 306 is used to evaluate the performance index of the multi-view spectrum clustering model according to the overall loss.

[0196] The judgment module 307 is used to judge whether the performance index reaches a preset standard performance index.

[0197] The optimization module 308 is used to optimize the overall loss to obtain the optimized overall loss when the judgment module 307 judges that the performance index does not reach the standard performance index, until the performance index of the multi-view spectrum clustering model reaches the preset standard performance index.

[0198] The determination module 309 is used to determine that the learning of the local learning layer and the global learning layer is completed when the judgment module 307 determines that the performance index reaches the standard performance index.

[0199] It can be seen that this optional implementation method can calculate the local learning loss of the local learning layer according to the acquired learning data of the local learning layer, and calculate the global learning loss of the global learning layer according to the acquired learning data of the global learning layer, and calculate the overall loss of the multi-view spectral clustering model according to the local learning loss and the global learning loss, thereby improving the calculation accuracy and reliability of the overall loss of the model. Subsequently, the performance index of the multi-view spectral clustering model is evaluated according to the overall loss, thereby improving the evaluation accuracy and efficiency of the model performance index, and judging whether the performance index reaches the preset standard performance index. When it is judged that the performance index does not reach the standard performance index, the overall loss is optimized to obtain the optimized overall loss until the performance index of the multi-view spectral clustering model reaches the preset standard performance index, which can improve the timeliness and accuracy of the optimization of the overall loss, thereby achieving the accurate achievement of the model performance index. When it is judged that the performance index reaches the standard performance index, it is determined that the local learning layer and the global learning layer have completed learning, and the training accuracy of the local learning layer and the global learning layer contained in the model can be improved through the performance index evaluated based on the overall loss.

[0200] Embodiment 4

[0201] See also Figure 5 , Figure 5 FIG. 1 is a schematic diagram of the structure of another multimodal hemodynamic analysis device for sepsis disclosed in an embodiment of the present invention. Figure 5 As shown, the multimodal hemodynamic analysis device applied to sepsis may include:

[0202] A memory 401 storing executable program codes;

[0203] a processor 402 coupled to the memory 401;

[0204] The processor 402 calls the executable program code stored in the memory 401 to execute the steps of the multimodal hemodynamic analysis method applied to sepsis described in the first embodiment of the present invention or the second embodiment of the present invention.

[0205] Embodiment 5

[0206] An embodiment of the present invention discloses a computer storage medium, which stores computer instructions. When the computer instructions are called, they are used to execute the steps of the multimodal hemodynamic analysis method applied to sepsis described in Embodiment 1 or Embodiment 2 of the present invention.

[0207] Embodiment 6

[0208] An embodiment of the present invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to execute the steps of the multimodal hemodynamic analysis method for sepsis described in Embodiment 1 or Embodiment 2.

[0209] The device embodiments described above are only illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, i.e., they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0210] Through the specific description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution can be essentially or partly contributed to the prior art in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0211] Finally, it should be noted that the multimodal hemodynamic analysis method and device for sepsis disclosed in the embodiment of the present invention discloses only the preferred embodiment of the present invention, which is only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal hemodynamic analysis method for sepsis, characterized in that: The method comprises: Monitoring multimodal data of each of the plurality of sepsis patients, wherein the multimodal data includes two or more of medical indicator data, diagnostic text data, and imaging data data; constructing an adjacency matrix of each of the views according to the sepsis data in each of the multiple views included in the multimodal data of each of the sepsis patients; Based on a dynamic fusion algorithm of a preset multi-view spectral clustering model, the adjacency matrices of all the views are embedded to obtain a multi-view embedding matrix; Based on a preset clustering algorithm, cluster analysis is performed on all the sepsis patients according to all the multi-view spectrum embedding matrices to obtain a clustering result, wherein the clustering result includes multiple sepsis patient groups, and all sepsis patients in each sepsis patient group correspond to a hemodynamic feature; the hemodynamic feature is used as a basis for generating a treatment strategy for all sepsis patients in the corresponding sepsis patient group.

2. The multimodal hemodynamic analysis method for sepsis according to claim 1, characterized in that: The step of constructing an adjacency matrix of each of the views according to the sepsis data in each of the multiple views included in the multimodal data of each of the sepsis patients comprises: According to a preset deep neural network model, feature extraction is performed on the sepsis data in each of the multiple views contained in the multimodal data of each sepsis patient to obtain feature data of each sepsis patient in each of the views; Based on a preset nearest neighbor analysis algorithm, and according to the feature data of all the sepsis patients in each of the views, a sample set corresponding to each of the sepsis patients is constructed, wherein each sample in the sample set corresponds to a sample type, and the sample type includes a positive sample type or a negative sample type; According to the characteristic data of the sample set corresponding to each of the sepsis patients in each of the views, optimizing the mapping function included in the pre-constructed target network, wherein the target network is a twin network with the same network branches; The adjacency matrix of each of the views is calculated according to the optimized mapping function and the feature data of the sample set corresponding to each of the sepsis patients in each of the views.

3. The multimodal hemodynamic analysis method for sepsis according to claim 2, characterized in that: The method based on the preset nearest neighbor analysis algorithm and according to the characteristic data of all the sepsis patients in each of the views, constructs a sample set corresponding to each of the sepsis patients, including: Based on a preset distance measurement algorithm, calculating the similarity between each of the sepsis patients according to the feature data of all the sepsis patients in each of the views; Based on a preset nearest neighbor analysis algorithm, and according to the similarities between all the sepsis patients, a neighbor set corresponding to each of the sepsis patients is constructed; According to the determined sample type of each sepsis patient, a sample set corresponding to each sepsis patient is determined.

4. The multimodal hemodynamic analysis method for sepsis according to claim 3, characterized in that: Determining a sample set corresponding to each sepsis patient according to the determined sample type of each sepsis patient includes: According to the determined sample type of each of the sepsis patients, samples having the same sample type as that of the sepsis patient are determined from a neighbor set corresponding to each of the sepsis patients as a positive sample set corresponding to the sepsis patient; According to the sample type of each sepsis patient, determine, from the neighbor set corresponding to each sepsis patient, samples that are different from the sample type of the sepsis patient as a negative sample set corresponding to the sepsis patient; A positive sample set corresponding to each of the sepsis patients and a negative sample set corresponding to the sepsis patient are determined as the sample set corresponding to the sepsis patient.

5. The multimodal hemodynamic analysis method for sepsis according to claim 2, characterized in that: The step of optimizing the mapping function included in the pre-constructed target network according to the feature data of the sample set corresponding to each of the sepsis patients in each of the views comprises: For each of the sepsis patients, the sample set is divided according to the sample types of all samples contained in the sample set corresponding to the sepsis patient to obtain multiple sample groups, each of the sample groups has a corresponding discriminant label, and the discriminant label includes a label that the corresponding sample groups are positive samples of each other or a label that the corresponding sample groups are negative samples of each other; Inputting feature data of each sample group in each view into a pre-constructed target network, and outputting a feature vector of the sample group through all network branches included in the target network; Calculating the contrast loss of the target network according to the feature vectors of all the sample groups, and optimizing the contrast loss of the target network to obtain an optimized contrast loss; According to the optimized contrast loss, the mapping function included in the target network is optimized.

6. The multimodal hemodynamic analysis method for sepsis according to any one of claims 1 to 5, characterized in that: The multi-view spectral clustering model includes a plurality of deep neural network layers, and each of the deep neural network layers includes one of a local learning layer, a global learning layer and an orthogonal constraint layer; And, the dynamic fusion algorithm based on the preset multi-view spectral clustering model embeds the adjacency matrices of all the views to obtain a multi-view embedding matrix, including: Based on the local learning layer, the sepsis data of each sepsis patient in each view and the adjacency matrix of each view are processed to obtain single view feature information of each view; Based on the global learning layer, single view feature information of all the views is processed to obtain a spectral embedding matrix of each of the views; Based on the orthogonal constraint layer, the spectrum embedding matrices of all the views are orthogonalized to obtain a multi-view spectrum embedding matrix.

7. The multimodal hemodynamic analysis method for sepsis according to claim 6, characterized in that: The method further comprises: In the process of learning using the local learning layer and the global learning layer, a local learning loss of the local learning layer is calculated according to the acquired learning data of the local learning layer; and a global learning loss of the global learning layer is calculated according to the acquired learning data of the global learning layer; Calculating the overall loss of the multi-view spectral clustering model according to the local learning loss and the global learning loss; According to the overall loss, evaluating the performance index of the multi-view spectrum clustering model; and determining whether the performance index reaches a preset standard performance index; When it is determined that the performance index does not reach the standard performance index, the overall loss is optimized to obtain an optimized overall loss, until the performance index of the multi-view spectrum clustering model reaches a preset standard performance index; When it is determined that the performance indicator reaches the standard performance indicator, it is determined that the learning of the local learning layer and the global learning layer is completed.

8. A multimodal hemodynamic analysis device for sepsis, characterized in that: The device comprises: A monitoring module, used to monitor multimodal data of each of the multiple sepsis patients, wherein the multimodal data includes two or more of medical indicator data, diagnostic text data and imaging data; A construction module, configured to construct an adjacency matrix of each of the views according to the sepsis data in each of the multiple views included in the multimodal data of each of the sepsis patients; A processing module, configured to embed the adjacency matrices of all the views based on a dynamic fusion algorithm of a preset multi-view spectral clustering model to obtain a multi-view embedding matrix; A clustering module is used to perform cluster analysis on all the sepsis patients based on a preset clustering algorithm and according to all the multi-view spectrum embedding matrices to obtain a clustering result, wherein the clustering result includes multiple sepsis patient groups, and all sepsis patients in each sepsis patient group correspond to a hemodynamic feature; the hemodynamic feature is used as a basis for generating a treatment strategy for all sepsis patients in the corresponding sepsis patient group.

9. A multimodal hemodynamic analysis device for sepsis, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the multimodal hemodynamic analysis method for sepsis according to any one of claims 1 to 7.

10. A computer storage medium, characterized in that: The computer storage medium stores computer instructions, and when the computer instructions are called, they are used to execute the multimodal hemodynamic analysis method for sepsis according to any one of claims 1 to 7.

Images