Medical behavior consistency identification model establishment method and risk identification method
By building a medical insurance risk identification model based on deep learning, using the correlation characteristics of the patient's medical time series data, the problem of difficulty in identifying new fraud methods in existing systems is solved, automatic identification and detection of new fraud patterns are realized, and new risks are provided. Identification perspective meets the needs of modern information technology.
Patent Information
- Application Number
- CN202111542828.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-16
- Publication Date
- 2024-12-13
- Estimated Expiration
- 2041-12-16
AI Technical Summary
The existing medical insurance risk control system is difficult to actively identify new fraud methods, and lacks positive and negative sample labels, and cannot directly locate violation characteristics through supervised learning. Traditional rules systems are difficult to meet the needs of modern information technology.
Based on the time series data of patients seeking medical treatment, data characteristics of medical behavior association are extracted, medical insurance risk identification model is constructed through deep learning models, data transformation and information fusion are used to generate medical information vectors, combining multi-head self-attention mechanism and loss reverse Disseminate optimization parameters to identify consistency in medical behavior to identify risks.
Provide automatic identification of emerging medical insurance fraud models, correlate and analyze the patient's previous medical information, realize purely big data-driven risk identification, provide a new perspective on detecting fraud risks, make full use of information in the prescription dimension, and do not rely on medical knowledge and policy terms.
Smart Images
Figure CN114357008B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an information processing technology, and in particular to a medical insurance risk identification model establishment method based on medical behavior consistency and a risk identification method. Background Art
[0002] At present, rule systems based on medical insurance policies and clinical knowledge are still the mainstream means of medical insurance risk control. The rule-based risk control engine can effectively screen predefined risk scenarios from the data and build the first line of defense for medical insurance funds. However, with the iteration of risk control systems, medical insurance fraud methods are also constantly updated. It is difficult for rule-based systems to actively locate new fraud methods. Therefore, during the online operation of the system, a large number of experts are still needed to continuously update the rules to deal with newly exposed fraud scenarios.
[0003] With the rapid development of modern information technology represented by big data and artificial intelligence, traditional rule systems can no longer meet all needs. Building a medical insurance risk control system with machine learning as the core has become the main development direction. Deep learning models combined with massive medical insurance settlement data make it possible to efficiently and accurately identify new types of fraud.
[0004] Existing medical insurance data lacks positive and negative sample labels, and it is impossible to directly locate illegal features through supervised learning methods. Therefore, a large number of studies have focused on unsupervised learning methods with clustering and outlier detection as the main means. Such methods often regard multiple visits to the same patient as relatively independent data, hoping to identify the internal risks of a patient's single visit to the doctor, but ignore the continuity and continuous correlation between the patient's visits to the doctor. In this context, the technology of the present invention proposes a fraud identification method based on patient medical time series data, which aims to discover inconsistencies within the patient's medical behavior sequence and thus indicate risks. Summary of the invention
[0005] Aiming at the problem of medical insurance risk control, a method for establishing a medical insurance risk identification model and a risk identification method based on the consistency of medical behavior were proposed. The data features associated with medical behavior were extracted from the time series data of patients' medical treatment for risk identification.
[0006] The technical solution of the present invention is: a method for establishing a medical insurance risk identification model based on medical behavior consistency, which specifically includes the following steps:
[0007] 1) Data preprocessing: Collect the patient information, medical information, and doctor's advice information, classify and count the collected medical information, and associate the items;
[0008] 2) Construct a single medical visit feature extraction model to infer prescription information: set category variables, embed data according to categories, and after forming various category information matrices, perform data conversion and information fusion on all category information matrices through the multi-layer fully connected layers of the neural network to generate the final single medical visit information vector as the information representation of a single medical visit; restore the original prescription information through the single medical visit vector, form a closed loop of the entire neural network, and optimize the parameters of the entire neural network through loss back propagation; the constructed single medical visit feature extraction model infers the use of single medical visits;
[0009] 3) Construct a feature extraction model for multiple medical visit sequences and identify the time period to which the medical visit sequence belongs: input each medical visit information into the single medical visit feature extraction model to obtain a single medical visit information vector, and horizontally splice the relative time information to obtain a vector of a single medical visit that incorporates time information; vertically splice multiple medical visit information vectors that incorporate time information to generate a medical visit sequence information matrix within a fixed time period; transform through a multi-head self-attention mechanism to generate a medical visit sequence information vector with the same dimension but further fully integrated and average pooled compression; combine the medical visit sequence information vector with relevant personal information to infer the absolute time information of the medical visit sequence, form a closed loop for the entire neural network, and optimize the parameters of the entire neural network by loss back propagation; the constructed multiple medical visit sequence feature extraction model infers the month of the input medical visit information;
[0010] 4) Construct a hospitalization and follow-up sequence consistency model to determine the consistency of the sequences before and after hospitalization; input the single hospitalization information into the single medical visit feature extraction model constructed in step 2) to generate a single hospitalization information vector for hospitalization; input all medical visit data within a fixed time period after discharge from a single hospitalization into the constructed multiple medical visit sequence feature extraction model to generate a follow-up sequence vector; combine the single hospitalization vector with the follow-up sequence vector to infer whether the follow-up sequence is a reasonable follow-up behavior for the hospitalization, form a closed loop for the entire neural network, and optimize the parameters of the entire neural network by loss back propagation; the constructed hospitalization and follow-up sequence consistency model identifies abnormal follow-up behavior.
[0011] Furthermore, the category information matrix construction method is as follows: the category variables are set, and the discrete data of medical behavior are first one-hot encoded according to the category variables, and then embedded through the embedding matrix, converted into the form of information vectors, and horizontally spliced with other information in the category to obtain a complete single item category information matrix.
[0012] Furthermore, the information matrix performs information fusion and information compression to obtain the information vector as follows:
[0013] First, the multi-head self-attention mechanism transformation:
[0014] A1) Input the information matrix X into three neural network fully connected layers in sequence to obtain the QUERY matrix Q, KEY matrix K, and VALUE matrix V;
[0015] A2) Transform Q, K, and V into self-attention mechanism according to the following formula:
[0016]
[0017] A3) Repeat the self-attention mechanism transformation in A2) multiple times, and horizontally splice the transformation results and perform linear transformation to obtain the result of the multi-head self-attention mechanism transformation. The formula is as follows:
[0018] MultiHead(Q,K,V)=Concat(head 1 ,...,head n )W O
[0019] Where head i represents the i-th result of the self-attention mechanism transformation in A2); W O The parameter matrix representing the linear transformation;
[0020] A4) Compress the information matrix transformed by the multi-head self-attention mechanism by average pooling to obtain the information vector.
[0021] Furthermore, the prescription information restored in step 2) includes:
[0022] B1) In the full list of items, whether each item was used during the current visit, inference method: using a single medical information vector as input, it is transformed through a fully connected layer of a neural network to generate a vector with a dimension equal to the length of the item dictionary, and then activated by the sigmoid function to standardize all values in the vector to between 0 and 1; if the value is less than 0.5, the model determines that the item at the corresponding position in the item dictionary is not used; if the value is greater than or equal to 0.5, the model determines that the item at the corresponding position in the item dictionary is used during this medical visit.
[0023] B2) If the model believes that the item is used, it is also necessary to infer the item's usage quantity and usage amount in the current visit. The inference method is as follows: take a single medical information vector as input, transform it through a neural network fully connected layer, generate an item amount or quantity prediction vector with a dimension equal to the length of the item dictionary, select the items actually used in this visit, locate their position in the item dictionary, and thus locate their corresponding position in the item amount or quantity prediction vector; calculate the square error loss based on the actual usage amount and quantity of the item and the predicted amount and quantity given by the item amount and quantity prediction vector, and continuously optimize the loss through back propagation; after multiple rounds of iterations, the model will gradually have the ability to infer the item usage amount and quantity.
[0024] Furthermore, the specific implementation steps for determining the time period to which the medical treatment sequence belongs in step 3) are as follows:
[0025] C1) Generate personal basic information vector: Personal basic information includes age, gender, and insurance type. Gender and insurance type are categorical variables. They need to be input into the model in the form of one-hot encoding and converted into dense information vectors through embedding operations. After the two are horizontally spliced, the age information is horizontally spliced to obtain the personal basic information vector;
[0026] C2) On the basis of the medical treatment sequence information vector, further horizontally splicing the personal basic information vector to generate an initial medical treatment sequence information vector containing personal information;
[0027] C3) The initial medical treatment sequence information vector containing personal information is input into the multi-layer fully connected layer of the neural network for data conversion and information fusion to generate the final medical treatment sequence information vector containing personal information as the information representation of the individual's medical treatment behavior over a period of time;
[0028] C4) Based on the medical treatment sequence information vector generated in C3), a multi-classification task is performed to infer the month information of the medical treatment sequence. The inference method is as follows: the medical treatment sequence information vector is used as input, and is converted through a fully connected layer of a neural network to generate a vector with a dimension of the total number of months. The vector is then activated through a softmax function and standardized so that the value of the vector elements is greater than or equal to 0 and the sum of the vector elements is equal to 1. Finally, the month corresponding to the position of the largest element in the vector is the month inference result of the model.
[0029] Furthermore, the sequence consistency determination before and after hospitalization in step 4) is specifically implemented as follows:
[0030] D1) Based on the single hospitalization vector, further splice the follow-up sequence vector;
[0031] D2) Fusion of the re-visit sequence vector and the single hospitalization vector through a multi-layer fully connected neural network;
[0032] D3) Based on the fused information vector, a binary classification task is performed to infer whether the medical treatment sequence is a reasonable follow-up behavior for the hospitalization.
[0033] A risk identification method based on a medical behavior consistency identification model. In the task of distinguishing the time period to which the medical treatment sequence belongs in the medical behavior consistency identification model, the medical treatment sequence contains the unique medical treatment mode in the time period, and the model is accurately identified and trained. A certain threshold is set through the probability value output by the model to screen out the medical treatment sequence that can be accurately predicted as the core sequence data for subsequent analysis; based on the screened core sequence data, further cluster analysis is performed to cluster the medical treatment sequence vector-similar data into one category; the sequences within the cluster will be consistent in the dimensions of medical treatment time, institution, diagnosis, and medical category; the consistency is measured by the internal distance of the cluster, and the smaller the internal distance of the cluster, the higher the similarity of the sequences within the cluster, and the more worthy of attention; the class with a very small or even 0 internal distance of the cluster will be used as the risk warning result of the model, which means that the medical treatment behavior characteristics of the people in the class in a certain time period are highly similar, and only appear in this time period, and are inconsistent with the medical treatment habits of people in other time periods.
[0034] The beneficial effects of the present invention are as follows: the present invention provides a method for establishing a medical insurance risk identification model and a risk identification method based on consistency of medical behavior, providing automatic identification of emerging medical insurance fraud patterns; correlation analysis of patients' previous medical information, providing a new perspective for detecting fraud risks; extracting features from medical insurance data in a deep learning manner, making full use of information from prescription dimensions; and risk identification that is purely big data driven, independent of medical knowledge and policy terms. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 This is a schematic diagram of the feature extraction of a single medical visit in the present invention;
[0036] Figure 2 This is a schematic diagram of distinguishing the time period to which the medical sequence of the present invention belongs;
[0037] Figure 3 This is a schematic diagram of sequence consistency discrimination before and after hospitalization according to the present invention. DETAILED DESCRIPTION
[0038] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0039] The medical insurance risk identification method based on consistency of medical behavior includes the following steps:
[0040] 1. Data preprocessing:
[0041] 1.1、Input data:
[0042] (1) Personal information: age, gender, and insurance type;
[0043] (2) Medical information: name of medical institution, grade of medical institution, medical category, time of admission / discharge, disease diagnosis, length of hospital stay, total medical expenses, and medical insurance fund expenditure;
[0044] (3) Medical order information: item category, item name, time of issuance, quantity used, unit price, and amount used;
[0045] 1.2. Data category statistics and associations:
[0046] (1) Age: an integer between 0 and 110;
[0047] (2) Gender: male, female;
[0048] (3) Insurance types: urban employees, urban and rural residents;
[0049] (4) Medical institution level: no level, level 1, level 2, level 3;
[0050] (4) Medical treatment categories: outpatient, hospitalization, and drug purchase;
[0051] (5) Discharge / admission time: divided into four dimensions: year, month, day, and hour;
[0052] (6) Disease diagnosis: Map the disease diagnosis in the original data to the ICD10 coding level, and associate the ICD10 sub-level, category, and chapter information based on the ICD10 hierarchical catalog table as supplementary information for disease diagnosis and input into the model;
[0053] (7) Project types: medicines, diagnosis and treatment projects, medical materials;
[0054] (8) Project name: Standardized mapping to the three major medical insurance catalogues.
[0055] 2. Construct a single medical visit feature extraction model:
[0056] like Figure 1 As shown in the figure, a fully connected network in deep learning is used to extract the features of a single medical visit, which specifically includes the following steps:
[0057] 2.1. Set category variables, embed data according to categories, and communicate various category information matrices:
[0058] Since the name of the medical order item, the name of the medical institution, the level of the medical institution, the medical category, and the ICD10 disease information are all discrete data and cannot be directly calculated numerically, they need to be embedded in the embedding matrix after one-hot encoding (the categorical variable is represented as a binary vector) and converted into the form of an information vector as the input of the subsequent layer.
[0059] (1) Medical order information list: The medical order information list contains the name of the prescribed item, the item category, and the quantity, unit price, and amount of each item. The item category and item name of a single item are categorical variables and are input into the model in the form of one-hot encoding. After the two are converted into dense vectors through the embedding operation, they are horizontally spliced to obtain the item category and name information vector; then the item category and name information vector and the corresponding item quantity, unit price, and amount of use data are directly horizontally spliced to obtain a complete single item information vector. Since the medical order item list generally contains multiple items, the medical order item list will obtain multiple item information vectors after the embedding operation, and the multiple item information vectors are vertically spliced to obtain the medical order item information matrix.
[0060] (2) Medical institution information: includes the name of the medical institution and the level of the medical institution. Both are categorical variables and are input into the model in the form of one-hot encoding. After being converted into a dense information vector through embedding operation, it is fused into a medical institution information vector through horizontal splicing.
[0061] (3) Medical category information: Categorical variables are input into the model in the form of one-hot encoding and can be directly converted into a dense vector through the embedding operation to form a medical category information vector.
[0062] (4) Disease diagnosis information: includes ICD10 details, ICD10 subheadings, ICD10 categories, and ICD10 chapters, all of which are categorical variables. They are input into the model in the form of one-hot encoding, converted into dense vectors through embedding operations, and then fused into disease diagnosis information vectors through horizontal splicing.
[0063] 2.2. Information Compression of Medical Order Item Information Matrix
[0064] In 2.1, after embedding and concatenating the medical order information list, the medical order item information matrix is obtained. Since the matrix is a second-order tensor with a higher order than the vector (first-order tensor), it is necessary to further perform information fusion and information compression on the matrix to obtain the medical order item information vector. The specific steps are as follows:
[0065] (1) The original medical order item information matrix is used as input and transformed by the multi-head self-attention mechanism to generate a medical order item information matrix with fully integrated information. The detailed steps of the multi-head self-attention mechanism transformation are as follows:
[0066] Step 1: Input the medical order item information matrix X into three neural network fully connected layers in sequence to obtain the QUERY matrix Q, KEY matrix K, and VALUE matrix V.
[0067] Step 2: Transform Q, K, and V into self-attention mechanism according to the following formula:
[0068]
[0069] Step 3: Repeat the self-attention mechanism transformation in step 2 multiple times, and horizontally splice the transformation results and perform linear transformation to obtain the result of the multi-head self-attention mechanism transformation. The formula is as follows:
[0070] MultiHead(Q,K,V)=Concat(head 1 ,...,head n )W O
[0071] Where head i represents the i-th result of the self-attention mechanism transformation in step 2; W O A parameter matrix representing a linear transformation.
[0072] (2) Using the average pooling method, the medical order item information matrix transformed by the multi-head self-attention mechanism is compressed to obtain the medical order item information vector.
[0073] 2.3. Single medical information vector generation:
[0074] After processing in 2.1 and 2.2, the information vector of the medical order item, medical institution information vector, medical category information vector and disease diagnosis information vector of a single medical visit are obtained. The four information vectors need to be further fused to generate the core single medical visit information vector. The specific steps are as follows:
[0075] (1) The medical order item information vector, medical institution information vector, medical category information vector and disease diagnosis information vector are horizontally spliced to generate the initial single medical information vector.
[0076] (2) Since simple horizontal splicing of vectors cannot fully integrate the information between the vectors, the initial single medical information vector needs to be further input into the multi-layer fully connected layer of the neural network for data conversion and information fusion to generate the final single medical information vector as the information representation of a single medical visit.
[0077] 2.4 Prescription Information Inference
[0078] The ultimate task of the single medical prescription feature extraction model is to restore the original prescription information through the single medical vector, forming a closed loop of the entire neural network, so that the parameters of the entire neural network can be optimized through loss back propagation. The specific prescription information to be restored includes:
[0079] (1) In the full list of items, whether each item was used during the current visit. Inference method: Take the single medical information vector in 2.3 as input, transform it through a fully connected layer of a neural network, generate a vector with a dimension equal to the length of the item dictionary, and then activate it through the sigmoid function to standardize all values in the vector to between 0 and 1. If the value is less than 0.5, the model determines that the item at the corresponding position in the item dictionary is not used; if the value is greater than or equal to 0.5, the model determines that the item at the corresponding position in the item dictionary is used during this visit.
[0080] (2) If the model believes that the item has been used, it is also necessary to infer the amount and amount of use of the item in the current visit. Inference method: Taking the amount of item used as an example, the same applies to the amount of use. The single medical information vector in 2.3 is used as input, and converted through a fully connected layer of a neural network to generate a project amount prediction vector with a dimension equal to the length of the project dictionary. Select the items actually used in this visit, locate their position in the project dictionary, and then locate their corresponding position in the project amount prediction vector. Based on the actual amount of use of the item and the predicted amount given by the project amount prediction vector, calculate the squared error loss, and continuously optimize the loss through back propagation. After multiple rounds of iterations, the model will gradually have the ability to infer the amount of item use.
[0081] 3. Construct a feature extraction model for multiple medical visits sequence - determine the time period to which the medical visit sequence belongs:
[0082] like Figure 2 As shown in the figure, the fully connected network in deep learning is used to extract the features of multiple medical visits and identify the time period to which the medical visit sequence belongs. The specific steps include:
[0083] 3.1 Generation of medical sequence information matrix
[0084] Select all medical data of the insured person within a fixed time period (specific month), input each medical information into the single medical feature extraction model to obtain a single medical information vector, and horizontally splice the relative time information to obtain a single medical vector that incorporates time information. Splice multiple medical information vectors that incorporate time information vertically to generate a medical sequence information matrix within a fixed time period.
[0085] Note: Relative time information refers to the relative time relationship between visits, such as the length of time between the second visit and the first visit, but does not include the specific dates and times of each visit.
[0086] 3.2 Generation of medical sequence information vector
[0087] After obtaining the medical treatment sequence information matrix in 3.1, since the matrix is a second-order tensor, it is necessary to further fuse and compress the information to obtain the medical treatment sequence information vector of the first-order tensor. The specific steps are as follows:
[0088] (1) The original medical treatment sequence information matrix is used as input and transformed by the multi-head self-attention mechanism to generate a medical treatment sequence information matrix with the same dimension but with further information fusion. The detailed steps of the multi-head self-attention mechanism transformation are as follows:
[0089] Step 1: Input the medical sequence information matrix X into three neural network fully connected layers in sequence to obtain the QUERY matrix Q, KEY matrix K, and VALUE matrix V.
[0090] Step 2: Transform Q, K, and V into self-attention mechanism according to the following formula:
[0091]
[0092] Step 3: Repeat the self-attention mechanism transformation in step 2 multiple times, and horizontally splice the transformation results and perform linear transformation to obtain the result of the multi-head self-attention mechanism transformation. The formula is as follows:
[0093] MultiHead(Q,K,V)=Concat(head 1 ,...,head n )W O
[0094] Where head i represents the i-th result of the self-attention mechanism transformation in step 2; W O A parameter matrix representing a linear transformation.
[0095] (2) Using the average pooling method, the medical treatment sequence information matrix after the multi-head self-attention mechanism is compressed to obtain the medical treatment sequence information vector.
[0096] 3.3 Absolute time information inference
[0097] The final task of the model is to merge relevant personal information through the medical sequence information vector, infer the absolute time information (month) of the medical sequence, and form a closed loop of the entire neural network, so that the parameters of the entire neural network can be optimized through loss back propagation. The specific steps are described as follows:
[0098] (1) Generate personal basic information vector: Personal basic information includes age, gender, and insurance type. Gender and insurance type are categorical variables. Both need to be input into the model in the form of one-hot encoding and converted into a dense information vector through embedding operation. The two are horizontally spliced, and then the age information is horizontally spliced to obtain the personal basic information vector.
[0099] (2) On the basis of the medical treatment sequence information vector, the personal basic information vector is further horizontally spliced to generate the initial medical treatment sequence information vector containing personal information.
[0100] (3) Since simple horizontal splicing of vectors cannot fully integrate the information between vectors, the initial medical sequence information vector containing personal information needs to be further input into the multi-layer fully connected layer of the neural network for data conversion and information fusion to generate the final medical sequence information vector containing personal information as the information representation of an individual's medical behavior over a period of time.
[0101] (4) Based on the medical treatment sequence information vector generated in (3), a multi-classification task is performed to infer the month information of the medical treatment sequence. Inference method: The medical treatment sequence information vector is used as input, and is transformed through a fully connected layer of a neural network to generate a vector with a dimension of the total number of months. The vector is then activated by a softmax function and normalized to a value greater than or equal to 0 and the sum of the vector elements is equal to 1. The month corresponding to the position of the largest element in the final vector is the month inference result of the model.
[0102] 3.4 Risk Identification Plan
[0103] In the task of distinguishing the time period to which the medical visit sequence belongs, the model cannot accurately predict the specific month of occurrence of most normal medical visit sequences. However, for a small number of sequences, the model can accurately predict the time of occurrence, indicating that these medical visit sequences contain unique medical visit patterns in this time period, and can therefore be accurately identified by the model. After the model training is completed, a certain threshold is set based on the probability value output by the model, so as to screen out the medical visit sequences that can be accurately predicted as the core sequence data for subsequent analysis.
[0104] Based on the core sequence data selected above, cluster analysis is further performed to cluster the data with similar medical sequence vectors into one category. The sequences within the cluster will have high consistency in dimensions such as medical time, institution, diagnosis, and medical category. The consistency is measured by the distance within the cluster. The smaller the distance within the cluster, the higher the similarity of the sequences within the cluster, and the more worthy of attention. Classes with very small or even 0 internal distances in the cluster will be used as risk warning results of the model. The meaning of their representation is that the medical behavior characteristics of the people in the class within a certain period of time are highly similar, and only appear in this period of time, and are inconsistent with the medical habits of people in other time periods.
[0105] 4. Constructing a model for consistency of hospitalization and follow-up visits - Discrimination of consistency of sequences before and after hospitalization
[0106] like Figure 3 As shown in the figure, the fully connected network in deep learning is used to extract the features of multiple medical visits and to judge the consistency of the sequences before and after hospitalization, which specifically includes the following steps:
[0107] 4.1 Single hospitalization information vector generation
[0108] The single hospitalization information is input into the single medical visit feature extraction model to generate a single hospitalization information vector of the hospitalization.
[0109] 4.2 Generation of follow-up sequence information matrix
[0110] Select all medical data of the insured person within a fixed time period after a single hospitalization and discharge, convert each medical visit into a single medical visit information vector, and splice the relative time information horizontally to obtain a vector of a single medical visit that incorporates time information. Splice the medical information vectors that incorporate time information vertically to generate a follow-up visit sequence information matrix within a fixed time period after the insured person is discharged from hospital.
[0111] Note: Relative time information only provides the relative time relationship between visits, such as the length of time between the second visit and the first visit, but does not provide the specific date and time of each visit.
[0112] 4.3 Generation of Follow-up Sequence Vector
[0113] After obtaining the medical treatment sequence matrix in 4.2, it is necessary to further compress the information through pooling operation to obtain the follow-up sequence vector. The specific steps are as follows:
[0114] (1) The original follow-up sequence matrix is used as input and transformed through a multi-head self-attention mechanism to generate a high-order follow-up sequence matrix with the same dimension but further integrated information.
[0115] (2) Compress the high-order follow-up sequence matrix using average pooling to obtain the follow-up sequence vector.
[0116] 4.4 Inference of consistency of hospitalization and follow-up visit sequences
[0117] The final task of the model is to combine the single hospitalization vector with the follow-up sequence vector, infer whether the follow-up sequence is a reasonable follow-up behavior for the hospitalization, and form a closed loop of the entire neural network, so that the parameters of the entire neural network can be optimized through loss back propagation. The specific steps are described as follows:
[0118] (1) Based on the single hospitalization vector, the follow-up sequence vector is further spliced.
[0119] (2) The follow-up sequence vector and single hospitalization vector are further fused through a multi-layer fully connected neural network.
[0120] (3) Based on the fused information vector, a binary classification task is performed to infer whether the medical treatment sequence is a reasonable follow-up behavior for the hospitalization.
[0121] 4.5 Training of hospitalization and follow-up sequence consistency model
[0122] During the model training process, the positive samples are the hospitalization data and the follow-up visit information sequence within the following three months, which means that most of the follow-up visits that actually occurred are reasonable. The model also requires a large number of negative samples to counter the positive samples, so as to give the model the ability to identify abnormal follow-up visits in real data. Negative samples need to be constructed based on a certain method, and the logic is as follows: for each hospitalization, record its discharge time T; randomly select 10 patients from the entire population as negative sample patients; from time T, select the medical information of the 10 negative sample patients in the next 3 months to generate 10 medical visit sequences as negative samples for follow-up visits. That is, the final ratio of positive samples to negative samples is about 1:10. The positive and negative sample data are mixed and input into the model. After multiple rounds of training, the model has the ability to identify abnormal follow-up visits.
[0123] 4.6 Risk Identification Plan
[0124] In the task of distinguishing the consistency of sequences before and after hospitalization, after the model training is completed, in order to use the model to identify risks, the real hospitalization information and its follow-up sequence information can be input into the model, and the model will directly output the consistency probability between the hospitalization information and the follow-up sequence information. For most of the real data, the model can accurately distinguish it as a normal follow-up; but there is a small part of the real data that will be marked as an abnormal follow-up, that is, the model believes that the hospitalization behavior and the follow-up behavior are inconsistent. The real data that is judged as a negative sample by the model with a probability higher than 95% will be detected as a model risk.
[0125] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.
Claims
1. A method for establishing a medical insurance risk identification model based on consistency of medical behavior, characterized in that: The specific steps include: 1) Data preprocessing: Collect the patient information, medical information, and doctor's advice information, classify and count the collected medical information, and associate the items; 2) Construct a single medical visit feature extraction model to infer prescription information: set category variables, embed data according to categories, and after forming various category information matrices, perform data conversion and information fusion on all category information matrices through the multi-layer fully connected layers of the neural network to generate the final single medical visit information vector as the information representation of a single medical visit; restore the original prescription information through the single medical visit vector, form a closed loop of the entire neural network, and optimize the parameters of the entire neural network through loss back propagation; the constructed single medical visit feature extraction model infers the use of single medical visits; 3) Construct a feature extraction model for multiple medical visit sequences and identify the time period to which the medical visit sequence belongs: input each medical visit information into the single medical visit feature extraction model to obtain a single medical visit information vector, and horizontally splice the relative time information to obtain a vector of a single medical visit that incorporates time information; vertically splice multiple medical visit information vectors that incorporate time information to generate a medical visit sequence information matrix within a fixed time period; transform through a multi-head self-attention mechanism to generate a medical visit sequence information vector with the same dimension but further fully integrated and average pooled compression; combine the medical visit sequence information vector with relevant personal information to infer the absolute time information of the medical visit sequence, form a closed loop for the entire neural network, and optimize the parameters of the entire neural network by loss back propagation; the constructed multiple medical visit sequence feature extraction model infers the month of the input medical visit information; 4) Construct a hospitalization and follow-up sequence consistency model to determine the consistency of the sequences before and after hospitalization; input the single hospitalization information into the single medical visit feature extraction model constructed in step 2) to generate a single hospitalization information vector for hospitalization; input all medical visit data within a fixed time period after discharge from a single hospitalization into the constructed multiple medical visit sequence feature extraction model to generate a follow-up sequence vector; combine the single hospitalization vector with the follow-up sequence vector to infer whether the follow-up sequence is a reasonable follow-up behavior for the hospitalization, form a closed loop for the entire neural network, and optimize the parameters of the entire neural network by loss back propagation; the constructed hospitalization and follow-up sequence consistency model identifies abnormal follow-up behavior; The method for constructing the category information matrix includes setting the category variables, first one-hot encoding the medical behavior discrete data according to the category variables, then embedding it through the embedding matrix, converting it into the form of information vector, horizontally splicing it with other information in the category, and obtaining a complete single item category information matrix; The information matrix performs information fusion and information compression to obtain the information vector implementation method as follows: First, the multi-head self-attention mechanism transformation: A1) Input the information matrix X into three neural network fully connected layers in sequence to obtain the QUERY matrix Q, KEY matrix K, and VALUE matrix V; A2) Transform Q, K, and V into self-attention mechanism according to the following formula: A3) Repeat the self-attention mechanism transformation in A2) multiple times, and horizontally splice the transformation results and perform linear transformation to obtain the result of the multi-head self-attention mechanism transformation. The formula is as follows: MultiHead(Q,K,V)=Concat(head1,...,head n )W O Where head i represents the i-th result of the self-attention mechanism transformation in A2); W O The parameter matrix representing the linear transformation; A4) Compress the information matrix transformed by the multi-head self-attention mechanism by average pooling to obtain the information vector.
2. The method for establishing a medical insurance risk identification model based on medical behavior consistency according to claim 1 is characterized in that: The prescription information restored in step 2) includes: B1) In the full list of items, whether each item was used during the visit, inference method: take a single medical information vector as input, transform it through a neural network fully connected layer, generate a vector with a dimension equal to the length of the item dictionary, and then activate it through the sigmoid function to standardize all values in the vector to between 0 and 1; if the value is less than 0.5, the model determines that the item at the corresponding position in the item dictionary is not used; if the value is greater than or equal to 0.5, the model determines that the item at the corresponding position in the item dictionary is used during the visit; B2) If the model believes that the item is used, it is also necessary to infer the item's usage quantity and usage amount in the current visit. The inference method is as follows: take a single medical information vector as input, transform it through a neural network fully connected layer, generate an item amount or quantity prediction vector with a dimension equal to the length of the item dictionary, select the items actually used in this visit, locate their position in the item dictionary, and thus locate their corresponding position in the item amount or quantity prediction vector; calculate the square error loss based on the actual usage amount and quantity of the item and the predicted amount and quantity given by the item amount and quantity prediction vector, and continuously optimize the loss through back propagation; after multiple rounds of iterations, the model will gradually have the ability to infer the item usage amount and quantity.
3. The method for establishing a medical insurance risk identification model based on medical behavior consistency according to claim 1 is characterized in that: The specific implementation steps for determining the time period to which the medical treatment sequence belongs in step 3) are as follows: C1) Generate personal basic information vector: Personal basic information includes age, gender, and insurance type. Gender and insurance type are categorical variables. They need to be input into the model in the form of one-hot encoding and converted into dense information vectors through embedding operations. After the two are horizontally spliced, the age information is horizontally spliced to obtain the personal basic information vector; C2) On the basis of the medical treatment sequence information vector, further horizontally splicing the personal basic information vector to generate an initial medical treatment sequence information vector containing personal information; C3) The initial medical treatment sequence information vector containing personal information is input into the multi-layer fully connected layer of the neural network for data conversion and information fusion to generate the final medical treatment sequence information vector containing personal information as the information representation of the individual's medical treatment behavior over a period of time; C4) Based on the medical treatment sequence information vector generated in C3), a multi-classification task is performed to infer the month information of the medical treatment sequence. The inference method is as follows: the medical treatment sequence information vector is used as input, and is converted through a fully connected layer of a neural network to generate a vector with a dimension of the total number of months. The vector is then activated through a softmax function and standardized so that the value of the vector elements is greater than or equal to 0 and the sum of the vector elements is equal to 1. Finally, the month corresponding to the position of the largest element in the vector is the month inference result of the model.
4. The method for establishing a medical insurance risk identification model based on medical behavior consistency according to claim 1 is characterized in that: The specific implementation steps for determining the consistency of the sequences before and after hospitalization in step 4) are as follows: D1) Based on the single hospitalization vector, further splice the follow-up sequence vector; D2) Fusion of the re-visit sequence vector and the single hospitalization vector through a multi-layer fully connected neural network; D3) Based on the fused information vector, a binary classification task is performed to infer whether the medical treatment sequence is a reasonable follow-up behavior for the hospitalization.
5. A risk identification method based on a medical behavior consistency identification model, characterized in that: A medical behavior consistency identification model established by the method for establishing a medical insurance risk identification model based on medical behavior consistency as described in any one of claims 1 to 4 is used to extract the medical treatment pattern unique to the time period contained in the medical treatment sequence in the medical behavior consistency identification model, and accurately identify and train the model, set a certain threshold value through the probability value output by the model, and screen out the medical treatment sequence that can be accurately predicted as the core sequence data for subsequent analysis; Based on the screened core sequence data, further cluster analysis is performed to cluster the data with similar medical sequence vectors into one category; the sequences within the cluster will be consistent in the dimensions of medical treatment time, institution, diagnosis, and medical category; the consistency is measured by the internal distance of the cluster. The smaller the internal distance of the cluster, the higher the similarity of the sequences within the cluster, and the more worthy of attention; the class with a very small or even 0 internal distance of the cluster will be used as the risk warning result of the model, which means that the medical treatment behavior characteristics of the people in the class within a certain period of time are highly similar, and only appear in this period of time, and are inconsistent with the medical treatment habits of people in other time periods.