JP7404581B1 - Chronic nephropathy subtype mining system based on self-supervised graph clustering - Google Patents

Abstract

The present invention provides a chronic nephropathy subtype mining system based on self-supervised graph clustering. [Solution] A data collection module for collecting structured data in chronic kidney disease medical records, and a data extraction module for extracting and preprocessing the structured data to obtain an entity set and a consultation set. and a preprocessing module, a chronic nephropathy subtype mining module for constructing a chronic nephropathy subtype mining model using the entity set and the consultation set, and a chronic nephropathy subtype mining module for evaluating the chronic nephropathy subtype mining model. The present invention includes a chronic nephropathy subtype evaluation module and a chronic nephropathy subtype prediction module for predicting structured data of patients. The present invention solves the problem that the process mining method cannot handle the coexistence of multi-granularity information such as event information within a single visit and event information between multiple visits in longitudinal electronic medical record data. [Selection diagram] Figure 1

Description

The present invention relates to the technical field of medical health information, and more particularly to a chronic nephropathy subtype mining system based on self-supervised graph clustering.

Chronic kidney disease is an important public health problem, affecting 10% of China's population. According to clinical guidelines, chronic nephropathy is graded by the patient's estimated glomerular filtration rate (eGFR) and urinary albumin-creatinine ratio (UACR). Although eGFR and UACR can be used for screening tests and monitoring measurements of chronic nephropathy, eGFR and UACR alone cannot express differences in disease phenotype between individuals of chronic nephropathy patients. Chronic kidney disease is a highly heterogeneous disease and is closely related to systemic diseases and conditions such as diabetes, hypertension, autoimmune diseases, genetic predispositions or congenital abnormalities. There are clear differences between individuals with chronic kidney disease, and these differences can be explained by disease phenotype, such as laboratory tests, medical history, medication history, and social factors. Due to the differences in the initial phenotype of patients with chronic kidney disease, the treatment process and complications vary widely among individuals. A rational phenotypic classification of chronic kidney disease should distinguish between different subgroups of patients and indicate the disease characteristics and underlying disease pathology of different subgroups, thereby identifying the different mechanisms of disease deterioration and progression. Contribute to better understanding.

Traditional chronic nephropathy subtype classification methods are mainly clustering analysis based on patients' initial static phenotypic data. These methods mainly utilize multidimensional data such as patient demographics, biomarkers, and clinical characteristics collected at the beginning of the study, and use commonly used clustering algorithms such as hierarchical clustering and consensus clustering to identify chronic kidney disease. Mining the phenotypic classification of disease patients. However, patients with chronic kidney disease have a long disease process and many complications, so there are large differences in the treatment process between individual patients. Clinical course data may implicitly contain important information that distinguishes between different phenotypes in patients with chronic kidney disease. Event information such as surgeries, examinations, tests, and drug treatments performed on a specific patient, as well as the times at which these events occurred, can be extracted from patient treatment process data collected and stored in the electronic medical record system. Clustering patients' clinical course data to study the disease phenotype mode of patients has important implications for identifying and studying the characteristics of different subgroups of patients. Regarding mining of disease treatment process data, commonly used methods are as follows. The first method is a process mining method, in which information is extracted from event logs generated during the patient treatment process and arranged in chronological order to form a treatment event sequence. Next, the disease phenotype of the patient is classified by mining different modes in the medical event sequence as different treatment processes of the disease. This method has difficulty in utilizing co-occurrence information between events, and cannot process the relational relationship and sequential ranking of events in data of multiple visits in a longitudinal electronic medical record. The mined medical treatment process is complex and has low representativeness and coverage. The second method is the tensor decomposition method, which combines three-dimensional information of patient, time, and phenotype into a third-order tensor, and mines the patient's latent phenotype classification by decomposing the third-order tensor. do. The method only considers changes in disease phenotype between consecutive visits and cannot process information on changes in phenotype over the course of long-term care.

Therefore, a chronic nephropathy subtype mining system based on self-supervised graph clustering is provided to solve the above technical problems.

In order to solve the above technical problems, the present invention provides a chronic nephropathy subtype mining system based on self-supervised graph clustering.

The technical scheme used by the present invention is as follows.

A chronic nephropathy subtype mining system based on self-supervised graph clustering, comprising:
a data collection module used to collect structured data in chronic nephropathy medical records;
a data extraction and preprocessing module used to extract and preprocess the structured data to obtain an entity set and a consultation set;
a chronic nephropathy subtype mining module used to construct a chronic nephropathy subtype mining model using the entity set and the consultation set;
a chronic nephropathy phenotype subtype evaluation module used to evaluate the chronic nephropathy subtype mining model;
and a chronic nephropathy subtype prediction module used to predict structured data of patients.

Furthermore, the structured data includes patient basic information, medical visit records, diagnosis during observation window, laboratory tests, medical tests, surgeries and/or medication data.

Further, the data extraction and pre-processing module specifically pre-processes the data set to extract patient basic information, consultation records, diagnosis during observation window, laboratory tests, medical tests, surgical data, medication data. The structured data in the chronic kidney disease medical record in the electronic medical record system is extracted, and the extracted structured data is preprocessed, and the laboratory test data is determined to be abnormal according to the normal reference range. It is concerned only with the test items of , divides abnormal test results into two types, too low and too high, maintains the names of abnormal test items and abnormal categories, and allows medical test and surgical data to be easily and naturally Processed using language processing technology, the examination site, category, and name of surgery are retained, and medication data is stored such as antihyperglycemic drugs, antihypertensive drugs, lipid regulators, nonsteroidal anti-inflammatory drugs, antiplatelet aggregating drugs, and steroids. We are only interested in the use of 6 types of drugs such as , classify the 6 types of drugs in medication data, maintain drug categories, diagnose set, medication set, surgery set, test set, number of diagnosis types, medication Obtain the number of types, the number of surgery types, the number of test types, and the number of consultation records, merge the diagnosis set, medication set, surgery set, and test set to form an entity set, and collect the patient consultation records. It is for configuring as a set.

Furthermore, the chronic nephropathy subtype mining module specifically includes:
a consultation network construction unit used for constructing a consultation network using the consultation set and the entity set;
constructing an entity co-occurrence matrix using the entity set; obtaining an initial embedding representation of an entity node and an initial embedding representation of a visited node using the entity co-occurrence matrix; an embedding representation construction unit used to construct an initial embedding representation of the node;
a clustering network construction unit used for constructing an adjacency matrix according to the relationships between nodes in the visited network, and training a clustering network model of visited nodes based on self-supervised graph clustering by the adjacency matrix and the initial embedding representation of the nodes; ,
a chronic nephropathy subtype mining model construction unit used for constructing a chronic nephropathy subtype mining model by the clustering network model of the visited nodes based on the self-supervised graph clustering.

Furthermore, the consultation network construction unit specifically:
The consultation set and the entity set are used to configure a node set;
used to construct an edge set based on node co-occurrence relationships in the node set;
The method includes using the node set and the edge set to construct a consultation network.

Furthermore, the embedded representation construction unit specifically includes:
being used to construct an entity co-occurrence matrix by the entity set;
used to calculate and obtain an initial embedding representation of each entity node by a GloVe algorithm based on the entity co-occurrence matrix;
Obtain the initial embedding representation of the visited node by calculating the average value of the initial embedding representation of the entity node of all neighboring entity nodes, and combine the initial embedding representation of the visited node and the initial embedding representation of the entity node to used in constructing an initial embedded representation.

Furthermore, the clustering network construction unit specifically includes:
constructing an adjacency matrix according to the relationships between nodes in the visited network, and inputting the adjacency matrix and the initial embedding representation of the nodes to the clustering network model of visited nodes based on the self-supervised graph clustering to perform graph attention training; used to obtain an embedded representation of a node including an embedded representation of a consultation node and an embedded representation of an entity node;
used to reconstruct the consultation network using the embedded representation of the nodes and calculate a reconstruction error of the consultation network;
inputting the embedded representation of the entity node to a decoder of a neural network for training, and using the output of the final layer of the decoder as the reconstructed embedded representation of the entity node to calculate a reconstruction error of the entity node;
performing a softmax regression operation on the embedded representation of the visited node to obtain a probability distribution of the visited node, and using the method to calculate a clustering loss based on the probability distribution of the visited node;
used to construct an overall loss function of a clustering network model of the visited nodes based on the self-supervised graph clustering based on the reconstruction error of the visited network, the reconstruction error of the entity nodes, and the clustering loss; include.

Furthermore, the chronic nephropathy subtype mining model construction unit specifically includes:
The clustering distribution of the visited nodes obtained by the clustering network model of the visited nodes based on the self-supervised graph clustering is set as the category distribution of the visited nodes, the category with the highest probability in the category distribution is selected as the category tag of the visited nodes, and each used for chronologically arranging all visit nodes of a patient;
Decide to merge or keep the visited nodes separately by calculating the cosine similarity between the category distributions of consecutive visited nodes with the same category tag, and construct an event matrix by arranging the visited nodes. to be used and
Search for frequent event determination nodes, connect consultation nodes in order to configure an event process, and select events whose event frequency in each column from the first column of the event matrix is greater than a threshold value as frequent events. Let the event be a node in the event process, the remaining events will directly enter the end node, and each event in the frequent events will be the starting node for the next search, and the corresponding event vectors will be extracted and combined into a new event matrix, and the first column After removing , perform a similar search operation for frequent events, and extend the event process by connecting the node obtained by searching each time to the start node, and the frequent event becomes null or the length of the event process becomes longer than the event process. and is used to obtain a chronic nephropathy subtype mining model by terminating the iterations until a maximum length of .

Furthermore, the chronic nephropathy subtype prediction module specifically includes:
Pre-processing the patient's structured data and inputting it into the clustering network model of visit nodes based on the self-supervised graph clustering for prediction to obtain a probability distribution of visit nodes of the patient;
determining a clustering category of the visited nodes based on the probability distribution of the visited nodes, and using the clustering category to construct a visited event sequence;
Input the consultation event sequence into the chronic nephropathy subtype mining model, sequentially fit nodes in the chronic nephropathy subtype mining model to obtain one event process, and determine which chronic nephropathy subtype by the event process. This includes being used to determine whether a person belongs to a group or not.

The beneficial effects of the present invention are as follows. The present invention provides a chronic nephropathy subtype mining system based on self-supervised graph clustering. First, longitudinal electronic medical record data from multiple patient visits, including multidimensional patient treatment event information such as consultations, diagnoses, laboratory tests, medical tests, surgeries, and medication, is constructed into a visit network. Next, a vector representation of the medical event is obtained based on the co-occurrence information of the medical event. Consultation events are clustered by a clustering network model of consultation nodes based on self-supervised graph clustering, and each consultation event is tagged. Additionally, in terms of consultations, we will mine the patient's treatment process to obtain different subtypes of chronic nephropathy phenotypes. Finally, we provide a method for assessing phenotypic subtypes, including a set of comprehensive indicators such as patient demographics, medications, comorbidities, and survival rates, to provide clinical interpretation for the different subtypes mined. Assess whether there are possible differences.

Among them, event information such as diagnosis, laboratory tests, medical tests, surgeries, and medications in each visit is first trained by a clustering network model of visit nodes based on self-supervised graph clustering, and category tags for each visit are determined. In this process, the low-level, fine-grained information is collected into high-level, coarse-grained comprehensive information, and the category tags of the consultation are used to mine the medical care process. Solved the problem of not being able to handle the coexistence of multi-granularity information such as event information within a single visit and event information between multiple visits in electronic medical record data.

Obtaining the event vector representation based on the co-occurrence information and using it in the graph model can effectively solve the problem that process mining methods have difficulty using event co-occurrence information, and can easily handle cross-sectional and longitudinal electronic medical record data. By using this simultaneously, it is possible to perform sufficient feature mining for diseases.

The provided self-supervised graph clustering algorithm simultaneously incorporates the patient's multiple consultation information into the clustering network model of the consultation nodes based on self-supervised graph clustering and the embedded representation of the training nodes, and incorporates the phenotypic change information over the long-term treatment process. Can be processed. Next, supervised learning is performed on different nodes and relationships in each visit network. The embedded representation of the low-level node is reconstructed by the decoder, the reconstruction error of the node is computed by the L2 norm, the reconstruction error of the graph relationship is computed by the cross entropy, and the clustering error of the visited node is computed by the KL divergence.

Based on the event tag distribution similarity of the visited nodes, similar adjacent events are merged, the process mining method is optimized, the mined medical process is simplified, and the representativeness and coverage rate of the medical process are improved.

FIG. 1 is a schematic structural diagram of a chronic nephropathy subtype mining system based on self-supervised graph clustering according to the present invention. FIG. 2 is a schematic diagram showing the functional process of the chronic nephropathy subtype mining system based on self-supervised graph clustering according to the present invention. FIG. 3 is a diagram showing a medical examination network according to an embodiment of the present invention. FIG. 4 is a diagram showing a co-occurrence matrix according to an embodiment of the present invention. FIG. 5 is a structural diagram showing a clustering network model of visiting nodes based on self-supervised graph clustering according to an embodiment of the present invention.

The following description of at least one exemplary embodiment is illustrative in nature and is not intended to limit the invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without any creative efforts fall within the protection scope of the present invention.

Referring to FIG. 1, a chronic nephropathy subtype mining system based on self-supervised graph clustering, comprising:
a data collection module used to collect structured data in chronic nephropathy medical records;
a data extraction and preprocessing module used to extract and preprocess the structured data to obtain an entity set and a consultation set;
a chronic nephropathy subtype mining module used to construct a chronic nephropathy subtype mining model using the entity set and the consultation set;
a chronic nephropathy phenotype subtype evaluation module used to evaluate the chronic nephropathy subtype mining model;
and a chronic nephropathy subtype prediction module used to predict structured data of patients.

Referring to FIG. 2, the functional process of the chronic nephropathy subtype mining system based on self-supervised graph clustering includes the following steps S1 to S5.

In step S1, a data collection module collects structured data in chronic nephropathy medical records to construct a dataset, and the structured data includes patient basic information, consultation records, diagnosis during observation window period, and laboratory tests. , including medical test, surgical and/or medication data;
In step S2, the data extraction and preprocessing module preprocesses the data set to obtain a visit set and an entity set, which specifically preprocesses the data set to obtain patient basic information, visit record, Extract structured data in the chronic kidney disease medical record in the electronic medical record system, including diagnosis, laboratory tests, medical tests, surgical data, and medication data during the observation window period, and pre-process the extracted structured data. Regarding the laboratory test data, we are concerned only with the abnormal test items according to the normal reference range, and we divide the abnormal test results into two types: too low and too high. medical examination and surgical data are processed using simple natural language processing technology, and examination site and category, surgical name are retained, and medication data is We are only interested in the use of six types of drugs, such as hyperglycemic drugs, antihypertensive drugs, lipid-regulating drugs, nonsteroidal anti-inflammatory drugs, antiplatelet aggregants, and steroids, and we classify the six types of drugs in the medication data. The categories of the diagnosis set, medication set, surgery set, test set, number of diagnosis types, number of medication types, number of surgery types, number of test types, and number of consultation records are obtained. merging a set, a medication set, a surgery set, and a test set into an entity set, and configuring a patient's visit record as a visit set.

In step S3, input the consultation set and the entity set to a chronic nephropathy subtype mining module, and construct a chronic nephropathy subtype mining model by the chronic nephropathy subtype mining module;
In step S31, a consultation network is constructed using the consultation set and the entity set,
In step S311, a node set is configured by the consultation set and the entity set,
In step S312, an edge set is constructed based on the node co-occurrence relationship in the node set,
In step S313, a consultation network is constructed using the node set and the edge set.

In step S32, an entity co-occurrence matrix is constructed using the entity set, an initial embedding representation of the entity node and an initial embedding representation of the visited node are obtained using the entity co-occurrence matrix, and an initial embedding representation of the entity node and the visited node are obtained. Construct the initial embedding representation of the node with the initial embedding representation of
In step S321, constructing an entity co-occurrence matrix using the entity set;
In step S322, an initial embedding representation of each entity node is calculated and obtained by the GloVe algorithm based on the entity co-occurrence matrix;
In step S323, the initial embedding representation of the visited node is obtained by calculating the average value of the initial embedding representation of the entity node of all adjacent entity nodes, and the initial embedding representation of the visited node and the initial embedding representation of the entity node are obtained. and constitute the initial embedding representation of the node.

In step S33, constructing an adjacency matrix according to the relationships between nodes in the visited network, and training a clustering network model of visited nodes based on self-supervised graph clustering by the adjacency matrix and the initial embedding representation of the nodes;
In step S331, an adjacency matrix is constructed according to the relationships between nodes in the visited network, and the adjacency matrix and the initial embedding representation of the nodes are input into the clustering network model of visited nodes based on the self-supervised graph clustering to perform graph attention training. to obtain the embedded representation of the node including the embedded representation of the consultation node and the embedded representation of the entity node,
In step S332, reconstructing the consultation network using the embedded representation of the node and calculating a reconstruction error of the consultation network;
In step S333, the embedded representation of the entity node is input to a decoder of a neural network for training, and the output of the final layer of the decoder is used as the reconstructed embedded representation of the entity node to calculate the reconstruction error of the entity node;
In step S334, a softmax regression operation is performed on the embedded representation of the visited node to obtain a probability distribution of the visited node, and a clustering loss is calculated based on the probability distribution of the visited node;
In step S335, an overall loss function of the clustering network model of the visited nodes based on the self-supervised graph clustering is constructed based on the reconstruction error of the visited network, the reconstruction error of the entity nodes, and the clustering loss.

In step S34, a chronic nephropathy subtype mining model is constructed using the clustering network model of the visited nodes based on the self-supervised graph clustering.

In step S341, the clustering distribution of the visited nodes obtained by the clustering network model of the visited nodes based on the self-supervised graph clustering is set as the category distribution of the visited nodes, and the category with the highest probability in the category distribution is set as the category tag of the visited nodes. Select and arrange all visit nodes for each patient in chronological order,
In step S342, it is determined to merge or keep the visited nodes separately by calculating the cosine similarity between the category distributions of successive visited nodes having the same category tag, and the event matrix is created by arranging the visited nodes. build,
In step S343, a frequent event determination node is searched, and an event process is configured by sequentially connecting consultation nodes, and events whose event frequency in each column from the first column of the event matrix is greater than a threshold value are determined as frequent events. Select and make the frequent events nodes in the event process, the remaining events directly enter the end node, and each event in the frequent events as the start node for the next search, extract the corresponding event vectors and combine them into a new event matrix. , after removing the first column, perform a similar search operation for frequent events, and extend the event process by connecting the nodes obtained by searching each time to the start node, so that the frequent events become null or the length of the event process Terminate the iteration until the maximum length of the event process is reached to obtain the chronic nephropathy subtype mining model.

In step S4, the chronic nephropathy subtype mining model is evaluated by a chronic nephropathy phenotypic subtype evaluation module,
In step S5, the patient's structured data is predicted by the chronic nephropathy subtype prediction module;
In step S51, the structured data of the patient is preprocessed and then input into the clustering network model of the consultation nodes based on the self-supervised graph clustering to predict, thereby obtaining the probability distribution of the consultation nodes of the patient;
In step S52, a clustering category of the visited node is determined based on the probability distribution of the visited node, and a visited event sequence is constructed;
In step S53, the consultation event sequence is input to the chronic nephropathy subtype mining model, nodes in the chronic nephropathy subtype mining model are sequentially fitted to obtain one event process, and the event process determines which chronic nephropathy Determine whether it belongs to the nephropathy subtype.

Embodiment A chronic nephropathy subtype mining system based on self-supervised graph clustering, which includes a data collection module, a data extraction and preprocessing module, a chronic nephropathy subtype mining module, a consultation network construction unit, an embedded representation construction unit, and a clustering network. It includes a construction unit, a chronic nephropathy subtype mining model construction unit, a chronic nephropathy phenotypic subtype evaluation module, and a chronic nephropathy subtype prediction module.

The data collection module is for constructing a dataset by collecting structured data in chronic kidney disease medical records, and the structured data includes patient basic information, consultation records, diagnosis during the observation window period, and laboratory data. including examination, medical examination, surgical and/or medication data;
The data extraction and preprocessing module is for extracting and preprocessing the structured data to obtain a consultation set and an entity set, and specifically, preprocessing the data set, Extract and extract structured data in the chronic kidney disease medical records in the electronic medical record system, including patient basic information, consultation records, observation window period diagnosis, laboratory tests, medical tests, surgical data, and medication data. For the laboratory test data, we are interested only in abnormal test items according to the normal reference range, and classify abnormal test item results into two categories: too low and too high. The name of the abnormal test item and the abnormal category are stored, the medical test and surgical data are processed using simple natural language processing technology, the test site and category, the name of the surgery are stored, and the medication data is We are only interested in the use of six types of drugs, including antihyperglycemic drugs, antihypertensive drugs, lipid-regulating drugs, nonsteroidal anti-inflammatory drugs, antiplatelet aggregants, and steroids, and we analyze the six types of drugs in the medication data. classify, maintain drug categories, obtain diagnosis sets, medication sets, surgery sets, test sets, number of diagnosis types, number of medication types, number of surgery types, number of test types, and number of consultation records; This is for merging the diagnosis set, medication set, surgery set, and test set to form an entity set, and configuring a patient's medical examination record as a medical examination set.

であり、ここで、Ｎ^Ｖが受診数を示す。Ｄ、Ｍ、Ｐ、Ｌがそれぞれ診断セット、服薬セット、手術セット、試験セットであり、

、

、

、

であり、ここで、Ｎ^Ｄ、Ｎ^Ｍ、Ｎ^Ｐ、Ｎ^Ｌがそれぞれ診断種類の数、服薬種類の数、手術種類の数、試験種類の数を示す。Ｄ、Ｍ、Ｐ、Ｌがエンティティセット

を構成し、エンティティセット種類の数がＮ^Ｓ＝Ｎ^Ｄ＋Ｎ^Ｍ＋Ｎ^Ｐ＋Ｎ^Ｌである。 The chronic nephropathy subtype mining module is for inputting the consultation set and the entity set into the chronic nephropathy subtype mining module, and constructing a chronic nephropathy subtype mining model by the chronic nephropathy subtype mining module. can be,
The consultation network construction unit is for constructing a consultation network using the consultation set and the entity set,
The consultation set and the entity set constitute a node set,
A medical examination set

, where ^NV indicates the number of consultations. D, M, P, and L are a diagnosis set, a medication set, a surgery set, and a test set, respectively;

,

,

,

Here, N ^D , N ^M , N ^P , and N ^L respectively represent the number of diagnosis types, the number of medication types, the number of surgery types, and the number of test types. D, M, P, L are entity sets

, and the number of entity set types is N ^S =N ^D +N ^M +N ^P +N ^L.

を構成し、ノードの個数がＮ^Ｎ＝Ｎ^Ｖ＋Ｎ^Ｓ＝Ｎ^Ｖ＋Ｎ^Ｄ＋Ｎ^Ｍ＋Ｎ^Ｐ＋Ｎ^Ｌであり、
前記ノードセットにおけるノード共起関係によってエッジセットを構築するためのものであり、
同一回の受診（Ｖ_ｉ）に現れるエンティティをエンティティサブセット

に構成し、ｊがエンティティサブセットＳ（Ｖ_ｉ）におけるエンティティの数を示し、

である。各エンティティサブセットがその対応する受診とともに１つの受診リンクサブセット

を構成する。１つの前記受診リンクサブセットには１つの受診ノード及び今回の受診におけるすべてのエンティティノードが含まれており、１つの前記受診リンクサブセットにおけるすべてのノードに共起関係があり、ノードが２つずつ接続してエッジサブセットを構成し、すべての前記エッジサブセットがエッジセットを構成し、前記エッジセットが

であり、
前記ノードセット及び前記エッジセットによって受診ネットワークＧ＝（Ｎ，Ｅ）を構築するためのものである。 Entity set is node set along with visit set

and the number of nodes is N ^N = N ^V + N ^S = N ^V + N ^D + N ^M + N ^P + N ^L ,
for constructing an edge set based on node co-occurrence relationships in the node set,
Entities that appear in the same medical visit (V _i ) are defined as an entity subset.

, where j denotes the number of entities in the entity subset S(V _i ), and

It is. Each entity subset has one visit link subset with its corresponding visit

Configure. One consultation link subset includes one consultation node and all entity nodes in the current consultation, and all the nodes in one consultation link subset have a co-occurrence relationship, and two nodes are connected. constitute an edge subset, all said edge subsets constitute an edge set, and said edge set

and
This is for constructing a consultation network G=(N,E) using the node set and the edge set.

が１つの受診リンクサブセットを構成し、受診ネットワークにおいてこの５つのノードが２つずつ接続している。受診Ｖ_４において、医者はＴＳＨ測定（Ｌ_３）を行ってから甲状腺機能低下症（Ｄ_４）の診断を下してレボチロキシンナトリウム錠剤（Ｍ_１）の薬を出す。そうすると、

も１つの受診リンクサブセットであり、受診ネットワークにおいてこの４つのノードが２つずつ接続している。Ｍ_１がＣ（Ｖ_１）及びＣ（Ｖ_４）に同時に現れるため、受診ネットワークにおいてＭ_１がこの２つの受診リンクサブセットにおける他のノードにいずれも接続している。 Referring to Figure 3, at visit _V1 , the doctor makes two diagnoses: goiter ( _D1 ) and thyroid nodule ( _D2 ), performs partial thyroidectomy ( _P1 ), and administers levothyroxine sodium. Dispense medicine in tablet form (M ₁ ). Then,

constitutes one consultation link subset, and these five nodes are connected two by two in the consultation network. At visit V ₄ , the doctor performs a TSH measurement (L ₃ ), then diagnoses hypothyroidism (D ₄ ) and prescribes the medicine levothyroxine sodium tablets (M ₁ ). Then,

is also one consultation link subset, and these four nodes are connected two by two in the consultation network. Since M ₁ appears simultaneously in C(V ₁ ) and C(V ₄ ), M ₁ is connected to both other nodes in these two visited link subsets in the visited network.

であり、
エンティティＳ_ｉとエンティティＳ_ｊが受診Ｖ_ｋにおいて同時に現れる場合、

が１に等しく、そうではない場合、０と記す。ここで、Ｓ（Ｖ_ｋ）が受診Ｖ_ｋにおいて現れるすべてのエンティティで構成される１つのエンティティサブセットである。エンティティ共起行列Ｘが対称であり、Ｘ_ｉｊとＸ_ｊｉが等しく、対角線上にあるものが同じエンティティの共起情報であり、０と記す。 The embedded representation construction unit constructs an entity co-occurrence matrix using the entity set, obtains an initial embedding representation of an entity node and an initial embedding representation of a visited node using the entity co-occurrence matrix, and combines the initial embedding representation of the entity node with the initial embedding representation of the visited node. This is for configuring the initial embedding representation of the node with the initial embedding representation of the consultation node,
for constructing an entity co-occurrence matrix by the entity set,
An entity co ^- occurrence matrix X is constructed by the entity set S, and with reference to FIG ^. 4, the dimensions of the entity co-occurrence matrix As a representative, X _ij indicates co-occurrence information of entity S _i and entity S _j . The calculation formula for X _ij is

and
If entity S _i and entity S _j appear simultaneously in visit V _k ,

is equal to 1, otherwise it is written as 0. Here, S(V _k ) is one entity subset consisting of all entities appearing in visit V _k . The entity co-occurrence matrix X is symmetrical, X _ij and X _ji are equal, and those on the diagonal are co-occurrence information of the same entity, which is written as 0.

と示され、
ここで、ｗ_ｉとｗ_ｊがそれぞれ最終的に求める必要のあるエンティティＳ_ｉ及びエンティティＳ_ｊのエンティティノードの初期埋め込み表現であり、１２８次元で－０．１～０．１間の値を取るランダムベクトルにランダムに初期化し、上付き文字Ｔが転置操作であり、ｂ_ｉとｂ_ｊがそれぞれ２つのエンティティノードの初期埋め込み表現のバイアス項であり、それらの初期値が０である。 for calculating and obtaining an initial embedding representation of each entity node by the GloVe algorithm based on the entity co-occurrence matrix;
The relationship between the initial embedding representation of entity nodes and the entity co-occurrence matrix is

It is shown that
Here, w _i and w _j are the initial embedding representations of entity nodes of entity S _i and entity S _j that need to be finally found, respectively, and take values between -0.1 and 0.1 in 128 dimensions. Randomly initialize to a random vector, the superscript T is the transpose operation, and b _i and b _j are the bias terms of the initial embedding representations of the two entity nodes, respectively, and their initial values are zero.

であり、
ここで、ＭＡＸが共起情報の閾値であり、αが指数パラメータである。 Construct an objective function J based on the relationship between the entity co-occurrence matrix and the initial embedding representation of the entity node,

and
Here, MAX is a threshold value of co-occurrence information, and α is an index parameter.

に対応するエンティティノードの初期埋め込み表現

を取得し、
すべての隣接するエンティティノードのエンティティノードの初期埋め込み表現の平均値を計算することにより受診ノードの初期埋め込み表現を取得し、前記受診ノードの初期埋め込み表現と前記エンティティノードの初期埋め込み表現とでノードの初期埋め込み表現を構成するためのものであり、
受診ノードＶ_ｉについては、そのすべての隣接するエンティティノードのセットが

であり、Ｖ_ｉノードの初期埋め込み表現は、

であり、
ここで、ｊがＳ（Ｖ_ｉ）におけるエンティティノードの数である。 If two entity nodes do not appear at the same time, ie if X _ij =0, they do not participate in the calculation of the objective function. Optimize the objective function by AdaDelta gradient descent algorithm until convergence, and each entity in the entity set

the initial embedding representation of the entity node corresponding to

and
Obtain the initial embedding representation of the visited node by calculating the average value of the initial embedding representation of the entity node of all neighboring entity nodes, and combine the initial embedding representation of the visited node and the initial embedding representation of the entity node to It is for configuring the initial embedding representation,
For the visited node V _i , the set of all its neighboring entity nodes is

, and the initial embedding representation of V _inode is

and
Here, j is the number of entity nodes in S(V _i ).

であり、Ｂ_Ｖが受診ノードの初期埋め込み表現であり、Ｂ_Ｓがエンティティノードの初期埋め込み表現である。 Initial embedding representation of the node

, B _V is the initial embedding representation of the visited node, and B _S is the initial embedding representation of the entity node.

The clustering network construction unit is for constructing an adjacency matrix according to the relationships between nodes in the visiting network, and training a clustering network model of visiting nodes based on self-supervised graph clustering by the adjacency matrix and the initial embedding representation of the nodes. Referring to FIG. 5, the clustering network model of visiting nodes based on self-supervised graph clustering is composed of three parts: graph attention, autoencoder, and self-supervision.

であり、
ここで、

がｒｅｌｕ活性化関数であり、Ｗ^ｌが第ｌ層のグラフ注意重みである。

であり、Ａが正規化された隣接行列であり、Ｉが単位行列であり、

である。Ｌ層のグラフ注意訓練を行った後、ノードの埋め込み表現Ｚ^Ｌを取得する。Ｚ^Ｌはノードの初期埋め込み表現Ｂと同様に、更新後の受診ノードの埋め込み表現Ｚ_Ｖ ^Ｌとエンティティノードの埋め込み表現Ｚ_Ｓ ^Ｌとで構成され、

である。 Constructing an adjacency matrix A according to the relationships between nodes in the visited network, and inputting the adjacency matrix A and the initial embedding representation B of the nodes into the clustering network model of visited nodes based on the self-supervised graph clustering to perform graph attention training. The embedding representation of the node in the lth layer is ^Zl , and the calculation method is as follows.

and
here,

is the relu activation function and W ^l is the graph attention weight of the lth layer.

, A is the normalized adjacency matrix, I is the identity matrix,

It is. After performing graph attention training for the L layer, obtain the embedded representation Z ^L of the node. Similar to the initial embedded representation B of the node, Z ^L is composed of the updated embedded representation Z _V ^L of the visited node and the embedded representation Z _S ^L of the entity node,

It is.

は、

であり、
ここで、（Ｚ^Ｌ）^ＴがＺ^Ｌの転置行列であり、

がｓｉｇｍｏｉｄ活性化関数である。 for reconstructing the consultation network using the embedded representation of the nodes and calculating a reconstruction error of the consultation network;
Adjacency matrix after reconstruction

teeth,

and
Here, (Z ^L ) ^T is the transposed matrix of Z ^L ,

is the sigmoid activation function.

であり、
ここで、

である。 Calculate the reconstruction error L _rec−G of the consultation network,

and
here,

It is.

であり、
ここで、Ｗ_ｄ ^ｙが第ｙ層のデコーダネットワーク重みであり、ｂ_ｄ ^ｙが偏差であり、デコーダの入力がＨ^０＝Ｚ_Ｓ ^Ｌである。デコーダの最終層の出力をエンティティノードの再構築埋め込み表現

としてエンティティノードの再構築誤差Ｌ_{ｒｅｃ－Ｓ}を計算し、

であり、
受診ノードの埋め込み表現Ｚ_Ｖ ^Ｌに対してｓｏｆｔｍａｘ回帰動作を行って、受診ノードの確率分布を取得するためのものであり、

であり、
ここで、Ｚ_Ｖ ^Ｒの次元がＮ^Ｖ×Ｋであり、Ｋがデフォルトのクラスタリングセンター数即ち受診ノードカテゴリ数であり、経験によって３、５、１０を試して結果がより良いカテゴリ数を選択する。

はｉ番目のサンプルがｊカテゴリに属する確率を示す。 This is for training by inputting the embedded representation _ZSL of the entity node to the decoder of the Y-layer neural network, and the representation of the node ⁱⁿ the y-layer decoder is H ^y , which was obtained by the calculation formula below. can be,

and
Here, W _dy is the decoder network weight of the y-th layer, ^{b dy} _is the deviation, and the input of ^the decoder is H ⁰ =Z _S ^L. The output of the final layer of the decoder is the reconstructed embedding representation of the entity node.

Calculate the reconstruction error L _rec−S of the entity node as

and
This is for performing a softmax regression operation on the embedded representation Z _V ^L of the visited node to obtain the probability distribution of the visited node,

and
Here, the dimension of Z _V ^R is N ^V ×K, K is the default number of clustering centers, that is, the number of visited node categories, and depending on experience, try 3, 5, and 10 and select the number of categories that gives the best result. .

indicates the probability that the i-th sample belongs to the j category.

であり、
ここで、ｑ_ｉｊはｉ番目のサンプルがｊ番目のクラスタに属する確率である。

をすべてのサンプルクラスタリング分布のセットとして設定する。クラスタリング分布Ｑを取得した後、目標分布Ｐを計算し、目標分布Ｐは一層高い信頼度のサンプル割り当てを有し、従って、Ｐに基づいてデータ分布を最適化してデータをクラスタリングセンターに更に近づけさせることができる。ＰとＱの次元がＮ^Ｖ×Ｋである。目標分布Ｐにおける各要素ｐ_ｉｊの計算公式は、

であり、
ここで、

である。目標分布Ｐにおいて、Ｑにおける各分布がいずれも二乗されるため、Ｐは一層高い信頼度を有する。クラスタリング損失の計算公式は、

であり、
受診ネットワークの再構築誤差Ｌ_{ｒｅｃ－Ｇ}、エンティティノードの再構築誤差Ｌ_{ｒｅｃ－Ｓ}及びクラスタリング損失Ｌ_ｃｌｕに基づいて、前記自己監督グラフクラスタリングに基づく受診ノードのクラスタリングネットワークモデルの全体損失関数を構築するためのものである。前記全体損失関数は、

であり、
ここで、γ、βが異なる損失項の重要性を調整するハイパーパラメータであり、０．１としてデフォルト設定される。 Calculating a clustering loss based on the probability distribution of the visited nodes;
For the i-th visit sample and the j-th cluster, the degree of similarity between the data representation z _i and the clustering center μ _j is determined using the student t distribution. z _i is the i-th row of Z _V ^R , μ _j is the clustering center initialized by the K-means method based on the probability distribution Z _V ^R of the visited nodes, and v is the degree of freedom of the student t distribution. Yes, the calculation formula for q _ij is

and
Here, q _ij is the probability that the i-th sample belongs to the j-th cluster.

Set as the set of all sample clustering distributions. After obtaining the clustering distribution Q, calculate a target distribution P, the target distribution P has a higher confidence sample assignment, and therefore optimize the data distribution based on P to bring the data closer to the clustering center. be able to. The dimensions of P and Q are N ^V ×K. The calculation formula for each element p _ij in the target distribution P is:

and
here,

It is. In the target distribution P, each distribution in Q is squared, so P has higher confidence. The formula for calculating clustering loss is:

and
Based on the reconstruction error L _rec-G of the visited network, the reconstruction error L _rec-S of the entity node, and the clustering loss L _clu , construct the overall loss function of the clustering network model of the visited nodes based on the self-supervised graph clustering. It is for. The overall loss function is

and
Here, γ and β are hyperparameters that adjust the importance of different loss terms, and are set as 0.1 by default.

The chronic nephropathy subtype mining model construction unit is for constructing a chronic nephropathy subtype mining model by the clustering network model of the visiting nodes based on the self-supervised graph clustering.

に対応するカテゴリタグが

である。単回受診の場合の１番目の医療記録の記録時間を受診ノードの開始時間とし、最後の医療記録の記録時間を受診ノードの終了時間とし、各患者のすべての受診ノードを時間順序で配列する。 The clustering distribution Q of the visited nodes obtained by the clustering network model of the visited nodes based on the self-supervised graph clustering is set as the category distribution of the visited nodes, and the category with the highest probability in the category distribution is selected as the category tag of the visited nodes. and the consultation node

The category tag corresponding to

It is. In the case of a single visit, the recording time of the first medical record is the start time of the visit node, the recording time of the last medical record is the end time of the visit node, and all visit nodes for each patient are arranged in chronological order. .

であり、
ここで、

がイベントＶ_ｉ、Ｖ_ｊのカテゴリ分布である。 Decide to merge or keep the visited nodes separately by calculating the cosine similarity between the category distributions of consecutive visited nodes with the same category tag, and construct an event matrix by arranging the visited nodes. It is a thing,
Calculate the cosine similarity between the category distributions of V _i and V _j for two consecutive visited nodes V _i and V _j having the same category tag,

and
here,

is the category distribution of events V _i and V _j .

であり、そうではない場合に２つの受診ノードを別個に保持する。同じカテゴリタグを有する複数の連続受診ノードの場合、配列順序で前から後まで２つごとにコサイン類似度の判断を行って、マージ又は別個に保持することを決定する。 Two consultation nodes with cosine similarity greater than 0.8 are merged into one consultation node, and the consultation node category distribution after merging is

, and if not, the two visited nodes are maintained separately. In the case of a plurality of consecutively visited nodes having the same category tag, the cosine similarity is determined for every two nodes from front to rear in the arrangement order, and it is determined whether to merge or keep them separately.

に配列し、ｋが受診ノードの一番多い患者のノード数であり、ノード数がｋ未満の患者の場合に０でイベントベクトルを充填する。すべての患者のイベントベクトルをイベント行列Ｈに組み合わせ、前記イベント行列Ｈは、

であり、
ここで、Ｈの次元がｎ×ｋであり、ｎが患者の総数である。 Finally, each patient's visit node is an event vector

where k is the number of nodes of the patient with the largest number of visited nodes, and if the number of nodes is less than k, the event vector is filled with 0. The event vectors of all patients are combined into an event matrix H, where the event matrix H is

and
Here, the dimension of H is n×k, where n is the total number of patients.

Search for frequent event determination nodes, connect consultation nodes in order to configure an event process, and select events whose event frequency in each column from the first column of the event matrix is greater than a threshold value as frequent events. Let the event be a node in the event process, the remaining events will directly enter the end node, and each event in the frequent events will be the starting node for the next search, and the corresponding event vectors will be extracted and combined into a new event matrix, and the first column After removing , perform a similar search operation for frequent events, and extend the event process by connecting the node obtained by searching each time to the start node, and the frequent event becomes null or the length of the event process becomes longer than the event process. This is to obtain the chronic nephropathy subtype mining model by terminating the iterations until the maximum length of .

The chronic nephropathy phenotype subtype evaluation module evaluates the chronic nephropathy subtype mining model,
By comparing the differences between patients with different phenotypic subtypes and testing whether there are statistical differences in the mined different subtype features, the disease subtypes obtained by the phenotypic subtype mining method can be This is to evaluate whether it has any meaning. The specific evaluation scheme is as follows.

Calculate indicators such as gender, age, and glomerular filtration rate of patients with different phenotypic subtypes to determine whether there are differences in clinical symptoms of patients with different phenotypic subtypes by statistical testing methods.

We will statistically analyze whether there are any differences in important medication data such as the usage of recombinant human erythropoietin, metformin, candesartan, and pravastatin among patients with different subtypes, and use statistical testing methods.

We calculated the number of patients with various comorbidities in each type of subtype, including cardiac weakness, coronary heart disease, hypertension, diabetes, and hyperlipidemia, calculated the proportion of each comorbidity, and calculated the prevalence of each comorbidity in patients with different subtypes. To test whether there are differences in comorbidity rates.

The total number of patients of each subtype and the number of survivors at different time points will be compiled to compare the survival rates of patients with different subtypes. Differences in survival rates of different patient subtypes over time are observed and analyzed by Log-rank tests.

If there are significant differences in the characteristics of more than 50% of patient groups of different subtypes, the mined subtype is said to have better clinical use value.

The chronic nephropathy subtype prediction module is for predicting patient structured data,
The structured data of the patient is preprocessed and then input into the clustering network model of the consultation nodes based on the self-supervised graph clustering for prediction to obtain the probability distribution of the consultation nodes of the patient,
determining a clustering category of the visited nodes based on the probability distribution of the visited nodes and constructing a visiting event sequence;
Input the consultation event sequence into the chronic nephropathy subtype mining model, sequentially fit nodes in the chronic nephropathy subtype mining model to obtain one event process, and determine which chronic nephropathy subtype by the event process. This is to determine whether it belongs to the .

The present invention provides a clustering network model of visited nodes based on self-supervised graph clustering, adds a decoder for reconstructing the embedded representation of nodes to graph attention training, and provides a self-supervised loss for training the clustering model. In addition, the clustering network model of consultation nodes based on self-supervised graph clustering collects low-level, fine-grained chronic kidney disease patient information into high-level, coarse-grained comprehensive information, and uses it to mine the medical treatment process. solves the problem of not being able to handle the coexistence of multi-grained information such as event information within a single visit and event information between multiple visits in longitudinal electronic medical record data, and analyzes the patient's single visit based on a self-supervised graph clustering method. In addition to fully integrating multidimensional medical information within a visit and sequence information between multiple visits, sufficient feature mining is performed on electronic medical record data from two dimensions, cross-sectional and longitudinal, and event tags of visit nodes are Merging similar adjacent events based on distribution similarity, optimizing the process mining method, simplifying the mined medical process, and improving the representativeness and coverage rate of the medical process.

The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention, and those skilled in the art can make various modifications and changes to the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims (4)

A chronic nephropathy subtype mining system based on self-supervised graph clustering, comprising a data collection module, a data extraction and preprocessing module, a chronic nephropathy subtype mining module, a chronic nephropathy phenotypic subtype evaluation module, and a chronic nephropathy subtype evaluation module. Contains a type prediction module,
The data collection module is used to collect structured data in chronic nephropathy medical records,
The chronic nephropathy subtype mining module is used to extract and preprocess structured data in the chronic nephropathy medical record to obtain an entity set and a consultation set,
The chronic nephropathy subtype mining module is used to construct a chronic nephropathy subtype mining model using the entity set and the consultation set,
The chronic nephropathy phenotype subtype evaluation module is used to evaluate the chronic nephropathy subtype mining model,
The chronic nephropathy subtype prediction module is used to predict structured data of a patient,
Specifically, the data extraction and preprocessing module preprocesses the structured data in the chronic kidney disease medical records to extract patient basic information, consultation records, diagnosis of observation window period, laboratory tests, and medical tests. , extract structured data from the chronic nephropathy medical records in the electronic medical record system, including surgical data and medication data, preprocess the extracted structured data from the chronic nephropathy medical records, and extract laboratory test data. According to the normal reference range, we are concerned only with abnormal test items, and divide abnormal test results into two types: too low and too high, and maintain the names of abnormal test items and abnormal categories. The system processes medical test and surgical data using simple natural language processing technology, retains the test site and category, and name of the surgery, and records medication data such as antihyperglycemic drugs, antihypertensive drugs, lipid-regulating drugs, and We are only interested in the use of six types of drugs: steroidal anti-inflammatory drugs, antiplatelet aggregants, and steroids, and we classify the six types of drugs in medication data, maintain drug categories, and create diagnostic sets, medication sets, and surgeries. set, test set, number of diagnosis types, number of medication types, number of surgery types, number of test types, and number of consultation records, and merge the diagnosis set, medication set, surgery set, and test set into an entity. It is used to configure patient consultation records as a consultation set.
The chronic nephropathy subtype mining module specifically includes a consultation network construction unit, an embedded representation construction unit, a clustering network construction unit, and a chronic nephropathy subtype mining model construction unit,
The consultation network construction unit is used to construct a consultation network using the consultation set and the entity set,
The embedded representation construction unit constructs an entity co-occurrence matrix using the entity set, obtains an initial embedding representation of an entity node and an initial embedding representation of a visited node using the entity co-occurrence matrix, and combines the initial embedding representation of the entity node with the initial embedding representation of the visited node. used to configure an initial embedding representation of the node with the initial embedding representation of the consultation node,
The clustering network construction unit constructs an adjacency matrix according to the relationship between nodes in the visited network, and trains a clustering network model of visited nodes based on self-supervised graph clustering by the adjacency matrix and the initial embedding representation of the nodes. used,
The chronic nephropathy subtype mining model construction unit is used to construct a chronic nephropathy subtype mining model by a clustering network model of consultation nodes based on the self-supervised graph clustering,
Specifically, the chronic nephropathy subtype mining model construction unit:
The clustering distribution of the visited nodes obtained by the clustering network model of the visited nodes based on the self-supervised graph clustering is set as the category distribution of the visited nodes, the category with the highest probability in the category distribution is selected as the category tag of the visited nodes, and each used for chronologically arranging all visit nodes of a patient;
Decide to merge or keep the visited nodes separately by calculating the cosine similarity between the category distributions of consecutive visited nodes with the same category tag, and construct an event matrix by arranging the visited nodes. to be used and
Search for frequent event determination nodes, connect consultation nodes in order to configure an event process, and select events whose event frequency in each column from the first column of the event matrix is greater than a threshold value as frequent events. Let the event be a node in the event process, the remaining events will directly enter the end node, and each event in the frequent events will be the starting node for the next search, and the corresponding event vectors will be extracted and combined into a new event matrix, and the first column After removing , perform a similar search operation for frequent events, and extend the event process by connecting the node obtained by searching each time to the start node, and the frequent event becomes null or the length of the event process becomes longer than the event process. used to obtain a chronic nephropathy subtype mining model by terminating the iterations until a maximum length of
Specifically, the chronic nephropathy subtype prediction module
The structured data of the patient is pre-processed and then input into a clustering network model of visit nodes based on the self-supervised graph clustering for prediction to obtain a probability distribution of visit nodes of the patient. ,
determining a clustering category of the visited nodes based on the probability distribution of the visited nodes, and using the clustering category to construct a visited event sequence;
Input the consultation event sequence into the chronic nephropathy subtype mining model, sequentially fit nodes in the chronic nephropathy subtype mining model to obtain one event process, and determine which chronic nephropathy subtype by the event process. including that it is used to determine whether it belongs to
A chronic nephropathy subtype mining system based on self-supervised graph clustering. Specifically, the consultation network construction unit:
The consultation set and the entity set are used to configure a node set;
used to construct an edge set based on node co-occurrence relationships in the node set;
The chronic nephropathy subtype mining system based on self-supervised graph clustering according to claim 1, further comprising: using the node set and the edge set to construct a consultation network. Specifically, the embedded representation construction unit:
being used to construct an entity co-occurrence matrix by the entity set;
used to calculate and obtain an initial embedding representation of each entity node by a GloVe algorithm based on the entity co-occurrence matrix;
The initial embedding representation of the visited node is obtained by calculating the average value of the initial embedding representation of the entity node of all neighboring entity nodes, and the initial embedding representation of the visited node and the initial embedding representation of the entity node are The chronic nephropathy subtype mining system based on self-supervised graph clustering as claimed in claim 1, further comprising: being used to construct an initial embedding representation. Specifically, the clustering network construction unit includes:
constructing an adjacency matrix according to the relationships between nodes in the visited network, and inputting the adjacency matrix and the initial embedding representation of the nodes to the clustering network model of visited nodes based on the self-supervised graph clustering to perform graph attention training; used to obtain an embedded representation of a node including an embedded representation of a consultation node and an embedded representation of an entity node;
used to reconstruct the consultation network using the embedded representation of the nodes and calculate a reconstruction error of the consultation network;
inputting the embedded representation of the entity node to a decoder of a neural network for training, and using the output of the final layer of the decoder as the reconstructed embedded representation of the entity node to calculate a reconstruction error of the entity node;
performing a softmax regression operation on the embedded representation of the visited node to obtain a probability distribution of the visited node, and using the method to calculate a clustering loss based on the probability distribution of the visited node;
used to construct an overall loss function of a clustering network model of the visited nodes based on the self-supervised graph clustering based on the reconstruction error of the visited network, the reconstruction error of the entity nodes, and the clustering loss; The chronic nephropathy subtype mining system based on self-supervised graph clustering as claimed in claim 1.

Images