A medical integrity risk prevention and control system and method based on multi-source data fusion

Abstract

The application discloses a medical incorruptible risk prevention and control system and method based on multi-source data fusion, and relates to the technical field of information processing.The system operation log data, ICD coding data and medicine catalog data in a hospital system are acquired, the ICD coding data and the medicine catalog data are standardized, and an initial association relationship is established.A medical coder is constructed, input data is obtained, and corresponding feature vectors are obtained.The association strength of the coding and the medicine is calculated, an association feature vector is generated, and the compliant and non-compliant data are input into a support vector machine classifier to train an association verification model.The system operation log data is input into the model to obtain operation matching degrees, and if the operation matching degrees are lower than a threshold value, an early warning is triggered.The operation time sequence of an operator is acquired, the deviation of the time interval and the historical average interval is calculated, and if the deviation exceeds a threshold value, an early warning is triggered.The method realizes the accurate prevention and control of medical incorruptible risks and improves the supervision efficiency through multi-source data fusion and intelligent models.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and in particular to a medical integrity risk prevention and control system and method based on multi-source data fusion. BACKGROUND

[0002] In recent years, with the continuous strengthening of the government's efforts in the field of people's livelihood, medical integrity issues have shifted from stage-specific rectification to full-chain normalized supervision, and the monitoring technology of medical integrity risks is upgrading from traditional administrative supervision to digital and intelligent direction. This transformation not only reflects the modernization needs of the governance system, but also reflects the higher requirements of the medical integrity risk monitoring system for regulatory technology.

[0003] However, in the current medical system, various types of data have independent terminology systems and semantic expression differences due to different sources, forming a semantic gap, which makes it difficult for computers to automatically recognize the associated logic between medical data, such as medical ICD diagnosis coding, drug directory, and operation logs. At the same time, there are coding structures and text structures, which makes it difficult to directly process these data through computer systems, and further makes it impossible to accurately locate the implicit abnormal medical patterns through cross-domain data association analysis, so the evaluation of medical integrity risks often relies on static rule bases and manual audits, and it is difficult to achieve unified risk judgment across systems. SUMMARY

[0004] The purpose of the present application is to provide a medical integrity risk prevention and control system and method based on multi-source data fusion to solve the problems in the prior art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] Step S1, obtaining system operation log data, ICD coding data and drug directory data in a hospital system; standardizing the ICD coding data and the drug directory data, and establishing an initial association relationship between the ICD coding data and the drug directory data;

[0007] Further, step S1 further comprises:

[0008] Step S1-1, directly obtaining ICD coding data, drug directory data and system operation log data through a hospital information system;

[0009] The ICD coding data includes ICD codes corresponding to diseases and hierarchical structures; the hierarchical structure includes chapters, categories, sub-categories and sub-categories of ICD codes;

[0010] The drug catalog data comprises drug basic identification and drug treatment attribute; the drug basic identification comprises drug generic name and drug ATC classification code; the drug treatment attribute is a treatment relationship of the drug and the ICD code corresponding to the disease;

[0011] The system operation log data comprises operator identification, operation time and operation data; the operation data comprises patient number and ICD code data of the disease of the patient, and drug circulation record; the drug circulation record is the warehouse-outbound destination of the drug;

[0012] Step S1-2, standardizing the obtained ICD code data and drug catalog data:

[0013] The standardization processing of the ICD code data comprises: verifying the structure of the ICD code by using regular expression; extracting chapter, category, subcategory and item information in the ICD code, and establishing a directed acyclic graph according to the hierarchical structure of the chapter, category, subcategory and item information with the ICD code as a node; analyzing case data to count the co-occurrence frequency of different codes appearing in the same medical record at the same time, establishing co-occurrence connection for the code pairs with the co-occurrence frequency exceeding a preset threshold, and taking the ratio of the co-occurrence frequency to the maximum co-occurrence frequency in all code pairs as the weight of the co-occurrence connection;

[0014] The standardization processing of the drug catalog data comprises: mapping different trade names of the same drug to a unique generic name; converting the drug treatment attribute into a standard code list by using an ICD official mapping tool;

[0015] Step S1-3, setting a drug statistical period, obtaining the use frequency of all drugs corresponding to a disease in the drug statistical period according to the drug treatment attribute, and calculating the association strength of a drug and the disease as the initial association relationship of the ICD code data and the drug catalog data according to the proportion of the use frequency of the drug in the use frequency of all drugs.

[0016] Step S2, constructing a medical encoder to input the ICD code data and the drug catalog data to obtain corresponding feature vectors;

[0017] Further, step S2 further comprises:

[0018] Constructing a medical encoder, the medical encoder comprising an ICD encoder and a drug encoder:

[0019] The construction of the ICD encoder comprises: vectorizing the chapter, category, subcategory and item information in the ICD code by one-hot encoding, and converting the clinical attribute into a numerical vector by the method of word vector; generating a disease semantic vector V e ;

[0020] The construction of the drug encoder includes: performing a word segmentation operation on the generic name of the drug to split continuous text into words or phrases with independent meanings; using a pre-trained language model to process the generic name after the word segmentation operation to output a fixed-dimensional vector, the fixed-dimensional vector being used to fuse semantic information and context relationships of the words or phrases with independent meanings in the generic name of each drug; calculating a mean of disease semantic vectors of all diseases corresponding to each drug to obtain a mean vector; inputting the fixed-dimensional vector and the mean vector into a fully connected layer of a convolutional neural network to perform linear and nonlinear transformation to obtain a drug feature vector V d ;

[0021] The disease semantic vector V e and the drug feature vector V d are converted into the same latitude when generated.

[0022] Step S3, calculating the association strength of the ICD code data and the drug directory data according to the feature vectors by a cosine similarity method to generate an association feature vector; obtaining compliant data and non-compliant data according to the association strength, inputting the compliant data and the non-compliant data into a support vector machine classifier to train an association verification model;

[0023] Further, step S3 further includes:

[0024] Step S3-1, calculating the association strength of the ICD code data and the drug directory data by a cosine similarity method:

[0025]

[0026] wherein V e represents the disease semantic vector, V d represents the drug feature vector, S represents the association strength of the ICD code data and the drug directory data, · represents the dot product operation of the vector, and ||| represents the modulus of the vector;

[0027] Step S3-2, pairing the disease semantic vector, the drug feature vector and the association strength to generate an association feature vector F=[V e , V d , S];

[0028] Step S3-3, set a correlation strength threshold, take the ICD code data and drug catalog data pair with the correlation strength greater than or equal to the correlation strength threshold as compliance data, and take the ICD code data and drug catalog data pair with the correlation strength less than the correlation strength threshold as non-compliance data; input the compliance data as positive samples and the non-compliance data as negative samples into a support vector machine classifier, optimize the penalty function parameters and the kernel function parameters through grid search, and train a correlation verification model by cross-validation; the correlation verification model outputs the matching degree of the ICD code data and the drug catalog data according to the input ICD code and drug generic name.

[0029] Step S4, input the system operation log data into the correlation verification model to obtain an operation matching degree; if the operation matching degree is less than a set matching degree threshold, trigger a medical risk warning; obtain the continuous operation time sequence of each operator in the system operation log data, calculate the deviation of the time interval of the continuous operation time sequence from the historical average time interval of the operator, set a time interval threshold, and when the deviation exceeds the time interval threshold, trigger a medical risk warning.

[0030] Further, step S4 further comprises:

[0031] Step S4-1, pair the ICD code of the patient's disease in the system operation log data with the drug generic name delivered to the patient, input the pair into the correlation verification model to obtain the operation matching degree of the pair; set a matching degree threshold, and when the operation matching degree is less than the matching degree threshold, trigger a medical risk warning and record the operator identifier and operation time of the operator corresponding to the delivery operation of the operator;

[0032] Step S4-2, group the system operation log data according to the operator identifier, and sort each group of data according to the operation time to form a continuous operation time sequence of each operator; the continuous operation time sequence includes the pair information of each delivery operation of the operator and the continuous operation time interval for the same patient;

[0033] Step S4-3, pre-train a long short-term memory network model using continuous operation time sequences without medical risks; the long short-term memory network model is used to calculate the historical average continuous operation time interval of each operator; set a time interval threshold, and if the deviation of the current continuous operation time interval of the operator from the historical average continuous operation time interval exceeds the time interval threshold, trigger a medical risk warning and record the operator identifier of the operator.

[0034] A medical integrity risk prevention and control system based on multi-source data fusion, the system comprises a data processing module, a feature coding module, an association model module and a risk early warning module; the data processing module is used for acquiring data in the hospital system for standardized processing, and establishing an initial association relationship between ICD coding data and drug directory data; the feature coding module is used for constructing a medical encoder, converting the ICD coding data and the drug directory data into corresponding feature vectors; the association model module is used for calculating the association strength of the ICD coding data and the drug directory data and training an association verification model; the risk early warning module is used for triggering medical risk early warning according to operation matching degree and deviation;

[0035] The output end of the data processing module is electrically connected to the input end of the feature coding module; the output end of the feature coding module is electrically connected to the input end of the association model module; and the output end of the association model module is electrically connected to the input end of the risk early warning module.

[0036] The data processing module further comprises a data acquisition unit and a data association unit; the data acquisition unit is used for acquiring system operation log data, ICD coding data and drug directory data in the hospital system; and the data association unit is used for standardizing the ICD coding data and the drug directory data, and establishing an initial association relationship therebetween according to the frequency of drug use.

[0037] The feature coding module further comprises an ICD coding unit and a drug coding unit; the ICD coding unit is used for constructing an ICD encoder, generating a disease semantic vector through one-hot coding, word vector conversion and graph convolution network processing; and the drug coding unit is used for constructing a drug encoder, generating a drug feature vector through word segmentation operation, pre-trained language model and convolutional neural network processing.

[0038] The association model module further comprises a strength calculation unit and a model training unit; the strength calculation unit is used for calculating the association strength of the ICD coding data and the drug directory data through the cosine similarity method, and generating an association feature vector; and the model training unit is used for inputting compliance and non-compliance data into a support vector machine classifier, and training an association verification model.

[0039] The risk early warning module further comprises a matching degree early warning unit and a deviation early warning unit; the matching degree early warning unit is used for inputting the system operation log data into the association verification model, and triggering medical risk early warning if the operation matching degree is less than a matching degree threshold; and the deviation early warning unit is used for analyzing the continuous operation time interval of the operator, and triggering medical risk early warning if the deviation of the operation time interval exceeds a time interval threshold.

[0040] Compared with the prior art, the present application has the following beneficial effects:

[0041] 1、The present application constructs a directed acyclic graph and establishes co-occurrence connection by standardizing the ICD code data and drug directory data, and combines the medical encoder to convert the data into a feature vector, and then calculates the correlation strength through cosine similarity and trains the correlation verification model, effectively solving the semantic gap problem caused by the independent term system and the semantic difference of medical data from different sources in the prior art, enabling the computer to automatically identify the association logic between medical data, accurately positioning the implicit abnormal medical mode, changing the limitations of relying on static rule base and manual review, and improving the accuracy and efficiency of medical integrity risk assessment.

[0042] 2、The present application inputs the patient disease ICD code in the system operation log data and the generic name of the out-of-stock drug into the correlation verification model to obtain the operation matching degree, and simultaneously analyzes the deviation of the continuous operation time interval of the operator from the historical average interval by using the long short-term memory network model, sets a double early warning mechanism, and is more intelligent than the traditional method. The correlation verification model can accurately judge the matching rationality of the code and the drug, the time interval analysis can capture abnormal operation rhythm, and the double dimension synergistic effect greatly reduces the missed and mistaken judgment probability, so that the medical risk early warning is more timely and accurate, and provides reliable technical support for medical integrity risk prevention and control.

[0043] 3、The present application integrates system operation log, ICD code and drug directory multi-source data to form a complete process from data acquisition, standardization processing, feature extraction, correlation model training to risk early warning, and realizes the whole chain supervision of medical integrity risk. The process covers each link from the establishment of the initial correlation relationship to the real-time operation monitoring, upgrades from single data processing to multi-source data fusion analysis, adapts to the transformation demand of medical integrity supervision from the stage special rectification to the whole chain normalization, and comprehensively improves the systematicness, continuity and effectiveness of medical integrity risk prevention and control in the medical field. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The flowchart of the medical integrity risk prevention and control method based on multi-source data fusion of the present application;

[0045] Figure 2 The structure diagram of the medical integrity risk prevention and control system based on multi-source data fusion of the present application. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0047] Embodiment one: as shown, the present application provides a technical solution, a medical integrity risk prevention and control method based on multi-source data fusion, the medical integrity risk prevention and control method comprises the following steps: Figure 1

[0048] Step S1, obtain system operation log data, ICD coding data and drug directory data in the hospital system; standardize the ICD coding data and the drug directory data, and establish the initial association relationship of the ICD coding data and the drug directory data;

[0049] Step S1 also includes:

[0050] Step S1-1, directly obtain ICD coding data, drug directory data and system operation log data through a hospital information system;

[0051] The ICD coding data includes ICD coding and hierarchical structure corresponding to the disease; the hierarchical structure includes chapter, category, subcategory and detail of ICD coding;

[0052] The drug directory data includes drug basic identification and drug treatment attribute; the drug basic identification includes drug generic name and drug ATC classification code; the drug treatment attribute is the treatment relationship of the drug and the ICD coding corresponding to the disease;

[0053] The system operation log data includes operator identification, operation time and operation data; the operation data includes the number of patients and the ICD coding data of the disease of the patient, and the drug circulation record; the drug circulation record is the destination of the drug out of the warehouse;

[0054] For example:

[0055] Three types of data are obtained from the hospital information system:

[0056] ICD coding data: select "type 2 diabetes" (ICD-10 code E11) and "hypertension" (ICD-10 code I10). The hierarchical structure of E11 is: chapter (fourth chapter "endocrine, nutrition and metabolic diseases") → category (E11 "type 2 diabetes") → subcategory (E11.9 "unspecified type 2 diabetes"), no detail (the subsequent co-occurrence connection is not affected by the absence of detail).

[0057] Drug directory data: select "metformin hydrochloride tablets" (generic name) and "nifedipine controlled-release tablets" (generic name). The ATC classification code of metformin is A10BA02, and the treatment attribute corresponds to E11 (type 2 diabetes); the ATC classification code of nifedipine is C08CA01, and the treatment attribute corresponds to I10 (hypertension).

[0058] ​System operation log data: select 3 operation records of operator Dr001.

[0059] Step S1-2, standardize the obtained ICD coding data and drug directory data:

[0060] The standardization processing of ICD coding data is: using regular expression to verify the structure of ICD coding; extracting chapter, category, subcategory and item information in ICD coding, establishing directed acyclic graph according to the hierarchical structure of chapter, category, subcategory and item with ICD coding as node; analyzing case data to count the co-occurrence frequency of different codes appearing in the same medical record at the same time, establishing co-occurrence connection for the code pairs with co-occurrence frequency exceeding the preset threshold, and taking the ratio of the co-occurrence frequency to the maximum co-occurrence frequency in all code pairs as the weight of the co-occurrence connection;

[0061] The standardization processing of drug directory data is: mapping different trade names of the same drug to a unique generic name; converting the drug treatment attribute to a standard code list through ICD official mapping tool;

[0062] For example:

[0063] ICD coding standardization:

[0064] Verify the structure compliance of E11 and I10 (both are "letter + number + decimal point + number" format) using regular expression; establish directed acyclic graph, add co-occurrence connection for E11 and I10 because their co-occurrence frequency is 0.3 (exceeding the preset threshold 0.2) in the medical record, and the weight is 0.3 / 0.5 (0.5 is the maximum co-occurrence frequency of all code pairs) = 0.6.

[0065] Drug directory standardization:

[0066] Map "Gua Zhi" (trade name) to the generic name "metformin hydrochloride tablets"; confirm the treatment relationship between metformin and E11 and the treatment relationship between nifedipine and I10 through ICD official tool.

[0067] Initial association relationship:

[0068] Statistical drug use frequency of E11 corresponding to 2024Q1 (drug statistical period): metformin 800 times, other drugs 1200 times, total frequency 2000 times. Then the initial association strength between metformin and E11 = 800 / 2000 = 0.4.

[0069] Step S1-3, set a drug statistical period, obtain the use frequency of all drugs corresponding to a disease in the drug statistical period according to the drug treatment attribute, and calculate the association strength of a drug and the disease as the initial association relationship of ICD coding data and drug directory data according to the proportion of the use frequency of the drug in the use frequency of all drugs.

[0070] Step S2, constructing a medical encoder, inputting ICD coding data and drug directory data to obtain corresponding feature vectors;

[0071] Step S2 further comprises:

[0072] Constructing a medical encoder, the medical encoder comprising an ICD encoder and a drug encoder:

[0073] The construction of the ICD encoder comprises: converting chapter, category, subcategory and item information in the ICD code into a vector representation through one-hot encoding, and converting clinical attributes into numerical vectors through the method of word vectors; adopting a graph convolution network to process ICD coding data to generate a disease semantic vector V e ;

[0074] The construction of the drug encoder comprises: performing a word segmentation operation on the generic name of the drug to split continuous text into words or phrases with independent meanings; using a pre-trained language model to process the generic name after the word segmentation operation to output a fixed-dimensional vector, the fixed-dimensional vector being used to fuse semantic information and context relationships of words or phrases with independent meanings in the generic name of each drug; calculating the mean of disease semantic vectors of all diseases corresponding to each drug to obtain a mean vector; inputting the fixed-dimensional vector and the mean vector into a fully connected layer of a convolutional neural network to perform linear and nonlinear transformation to obtain a drug feature vector V d ;

[0075] The disease semantic vector V e and the drug feature vector V d are converted into the same latitude when generated.

[0076] For example:

[0077] ICD encoder construction:

[0078] One-hot encoding is performed on E11: chapter (Chapter IV) → [1, 0, 0,...], category (E11) → [0, 1, 0,...], subcategory (E11.9) → [0, 0, 1,...].

[0079] Combining the co-occurrence frequency in clinical cases (E11 and I10 co-occurrence frequency 0.3), a disease semantic vector is generated through a graph convolution network (GCN):

[0080] Ve (E11) = [0.2, 0.3, 0.1, 0.4] (4-dimensional vector, example value).

[0081] Drug encoder construction:

[0082] The "metformin hydrochloride tablets" are segmented into "hydrochloric acid", "metformin", and "tablets", and a pre-trained language model (such as BERT) is used to generate a text vector: [0.1, 0.2, 0.5, 0.2] (4-dimensional).

[0083] Calculate the semantic vector mean of the disease (E11) corresponding to this drug (i.e. V e (E11)): [0.2, 0.3, 0.1, 0.4].

[0084] After processing by the convolutional neural network full connection layer, the drug feature vector is generated:

[0085] V d (metformin) = [0.15, 0.25, 0.3, 0.3] (4-dimensional vector, example value).

[0086] Step S3, according to the feature vector, the correlation strength between the ICD code data and the drug directory data is calculated by the cosine similarity method, and the correlation feature vector is generated; according to the correlation strength, the compliant data and the non-compliant data are obtained, and the compliant data and the non-compliant data are input into a support vector machine classifier to train an association verification model;

[0087] Step S3 also includes:

[0088] Step S3-1, calculate the correlation strength between the ICD code data and the drug directory data by the cosine similarity method:

[0089]

[0090] where V e represents the disease semantic vector, V d represents the drug feature vector, S represents the correlation strength between the ICD code data and the drug directory data, · represents the dot product operation of the vector, and |||| represents the length of the vector.

[0091] For example, the correlation strength between E11 and metformin is calculated by the cosine similarity formula:

[0092] Dot product: V e · V d = 0.225;

[0093] Length: ||V e || ≈ 0.547; ||V d || ≈ 0.515;

[0094] Correlation strength: S ≈ 0.255 / (0.547*0.515) ≈ 0.904;

[0095] Step S3-2, pair the disease semantic vector, the drug feature vector and the correlation strength to generate a correlation feature vector F = [V e , V d , S];

[0096] Step S3-3, set a correlation strength threshold, take the ICD code data and drug catalog data pairs with a correlation strength greater than or equal to the correlation strength threshold as compliance data, and take the ICD code data and drug catalog data pairs with a correlation strength less than the correlation strength threshold as non-compliance data; input the compliance data as positive samples and the non-compliance data as negative samples into a support vector machine classifier, optimize the penalty function parameters and kernel function parameters through grid search, and train a correlation verification model using cross-validation; the correlation verification model outputs the matching degree of ICD code data and drug catalog data according to the input ICD code and drug generic name.

[0097] For example:

[0098] Positive samples (compliance data): select pairs with a correlation strength ≥ 0.7, a total of 600 pairs. For example:

[0099] (E11, metformin, S = 0.904), (I10, nifedipine, S = 0.88), etc.

[0100] Negative samples (non-compliance data): select pairs with a correlation strength < 0.7, a total of 400 pairs. For example:

[0101] (E11, nifedipine, S = 0.2), (I10, metformin, S = 0.3), etc.

[0102] Support vector machine (SVM) training:

[0103] Input features: correlation feature vector F = [Ve, Vd, S]. For example, the F of (E11, metformin) is:

[0104] [0.2, 0.3, 0.1, 0.4, 0.15, 0.25, 0.3, 0.3, 0.904] (9-dimensional vector).

[0105] Parameter optimization: use grid search, set the candidate values of the penalty parameter C as [1, 10, 100], the candidate values of the kernel function gamma as [0.1, 1, 10], and evaluate the model performance (indicator is accuracy) through 5-fold cross-validation.

[0106] The final model has an accuracy of 91% on the test set and outputs an operation matching degree (range 0-1).

[0107] Step S4, input the system operation log data into the association verification model to obtain an operation matching degree; if the operation matching degree is less than a set matching degree threshold, trigger a medical risk warning; obtain the continuous operation time sequence of each operator in the system operation log data, calculate the deviation of the time interval of the continuous operation time sequence from the historical average time interval of the operator, set a time interval threshold, and when the deviation exceeds the time interval threshold, trigger a medical risk warning.

[0108] Step S4 also includes:

[0109] Step S4-1, pair the ICD code of the patient's disease in the system operation log data with the generic name of the drug dispensed to the patient, input the pair into the association verification model to obtain the operation matching degree of the pair; set a matching degree threshold, and when the operation matching degree is less than the matching degree threshold, trigger a medical risk warning and record the operator identifier and operation time of the operator corresponding to the dispensing operation;

[0110] Step S4-2, group the system operation log data by operator identifier, and sort each group of data by operation time to form a continuous operation time sequence for each operator; the continuous operation time sequence includes the pair information of each dispensing operation of the operator and the continuous operation time interval for the same patient;

[0111] Step S4-3, pre-train a long short-term memory network model using continuous operation time sequences of operators without medical risks; the long short-term memory network model is used to calculate the historical average continuous operation time interval of each operator; set a time interval threshold, and if the deviation of the current continuous operation time interval of the operator from the historical average continuous operation time interval exceeds the time interval threshold, trigger a medical risk warning and record the operator identifier of the operator.

[0112] For example:

[0113] Select the continuous operation time sequence of an operator without medical risks (such as Dr002), and extract the time interval:

[0114] [5 minutes, 7 minutes, 6 minutes, 8 minutes, 5 minutes] (a total of 5 consecutive operations, the interval is the time difference between two operations).

[0115] Input: time interval sequence (normalized to the range [0, 1]);

[0116] Output: predict the historical average time interval of the operator;

[0117] Training process:

[0118] Initial history average calculation: (5+7+6+8+5) / 5 = 6.2 minutes;

[0119] Loss function: Mean Squared Error (MSE), initial loss 0.8;

[0120] After 10 iterations, the loss decreased to 0.1, the model converged, and the output history average time interval was stable.

[0121] Operation matching degree early warning:

[0122] Dr001 is the patient P003 (ICD code E11) discharging "nifedipine controlled release tablets", inputting the association verification model, the output matching degree is 0.3, which is greater than the matching degree threshold 0.7, triggering the early warning.

[0123] Time interval early warning:

[0124] The continuous operation time interval of Dr001 is 15 minutes, which is 8.8 minutes different from the history average output by the LSTM model, which is 6.2 minutes, and is greater than the time interval threshold 5 minutes, triggering the early warning.

[0125] Example two, as Figure Two shown, the application provides a medical integrity risk prevention and control system based on multi-source data fusion, which comprises a data processing module, a feature coding module, an association model module and a risk early warning module; the data processing module is used for acquiring data in the hospital system for standardization processing, and establishing an initial association relationship between ICD code data and drug directory data; the feature coding module is used for constructing a medical encoder to convert the ICD code data and the drug directory data into corresponding feature vectors; the association model module is used for calculating the association strength of the ICD code data and the drug directory data and training an association verification model; the risk early warning module is used for triggering medical risk early warning according to operation matching degree and deviation;

[0126] The output end of the data processing module is electrically connected to the input end of the feature coding module; the output end of the feature coding module is electrically connected to the input end of the association model module; and the output end of the association model module is electrically connected to the input end of the risk early warning module.

[0127] The data processing module further comprises a data acquisition unit and a data association unit; the data acquisition unit is used for acquiring system operation log data, ICD code data and drug directory data in the hospital system; and the data association unit is used for standardizing the ICD code data and the drug directory data, and establishing an initial association relationship between the two according to the drug use frequency.

[0128] The feature coding module further comprises an ICD coding unit and a drug coding unit; the ICD coding unit is used to construct an ICD encoder to generate a disease semantic vector through one-hot coding, word vector conversion and graph convolution network processing; the drug coding unit is used to construct a drug encoder to generate a drug feature vector through word segmentation operation, pre-trained language model and convolutional neural network processing.

[0129] The association model module further comprises a strength calculation unit and a model training unit; the strength calculation unit is used to calculate the association strength of the ICD coding data and the drug directory data through the cosine similarity method to generate an association feature vector; the model training unit is used to input the compliance and non-compliance data into a support vector machine classifier to train an association verification model.

[0130] The risk early warning module further comprises a matching degree early warning unit and a deviation early warning unit; the matching degree early warning unit is used to input the system operation log data into the association verification model, and if the operation matching degree is less than the matching degree threshold, a medical risk early warning is triggered; the deviation early warning unit is used to analyze the continuous operation time interval of the operator, and if the deviation of the operation time interval exceeds the time interval threshold, a medical risk early warning is triggered.

[0131] It will be obvious to a person skilled in the art that the application is not limited to the details of the above-described exemplary embodiments but can be implemented in other embodiments without departing from the scope of the application. The scope of the application is defined by the claims and not by the description, therefore all changes and modifications that fall within the meaning and range of equivalents of the claims are intended to be embraced therein. Any reference signs in the claims should not be construed as limiting the claims concerned.

Claims

1. A method for preventing and controlling medical corruption risks based on multi-source data fusion, characterized in that: The method includes the following steps: Step S1: Obtain system operation log data, ICD code data, and drug catalog data from the hospital system; standardize the ICD code data and drug catalog data, and establish the initial association between the ICD code data and drug catalog data; Step S2: Construct a medical encoder by inputting ICD-coded data and drug catalog data to obtain the corresponding feature vectors; Step S3: Calculate the association strength between ICD-coded data and drug catalog data using the cosine similarity method based on the feature vector, and generate an association feature vector; obtain compliant and non-compliant data based on the association strength, and input the compliant and non-compliant data into a support vector machine classifier to train an association verification model; Step S4: Input the system operation log data into the association verification model to obtain the operation matching degree; if the operation matching degree is less than the set matching degree threshold, trigger a medical risk warning; obtain the continuous operation time sequence of each operator in the system operation log data, calculate the deviation between the time interval of the continuous operation time sequence and the historical average time interval of the operator, set the time interval threshold, and trigger a medical risk warning when the deviation exceeds the time interval threshold. Step S2 includes: Construct a medical encoder, which includes an ICD encoder and a drug encoder: The construction of the ICD encoder includes: representing the chapter, category, subcategory, and detail information in the ICD encoding as vectors using one-hot encoding; converting clinical attributes into numerical vectors using word vectors; and using a graph convolutional network to process the ICD encoded data to generate a disease semantic vector V. e ; The construction of the drug encoder includes: performing word segmentation on the generic name of the drug to split continuous text into words or phrases with independent meanings; using a pre-trained language model to process the segmented generic name and output a fixed-dimensional vector, which is used to fuse the semantic information and contextual relationships of words or phrases with independent meanings in the generic name of each drug; calculating the mean of the disease semantic vectors of all diseases corresponding to each drug to obtain the mean vector; and inputting the fixed-dimensional vector and the mean vector into the fully connected layer of a convolutional neural network to perform linear and nonlinear transformations to obtain the drug feature vector V. d ; The disease semantic vector V e and the drug feature vector V d Converted to the same dimensions during generation; Step S3 includes: Step S3-1: Calculate the association strength between ICD-coded data and drug catalog data using the cosine similarity method: ； Where V e V represents the semantic vector of the disease. d The vector represents the feature vector of the drug, S represents the correlation strength between the ICD encoded data and the drug catalog data, · represents the dot product operation of the vector, and || represents the magnitude of the orientation quantity. Step S3-2: Pair the disease semantic vector, the drug feature vector, and the association strength to generate an association feature vector F=[V e V d ,S]; Step S3-3: Set an association strength threshold. ICD-coded data and drug catalog data pairs with an association strength greater than or equal to the threshold are considered compliant data, while those with an association strength less than the threshold are considered non-compliant data. The compliant data is used as positive samples, and the non-compliant data as negative samples, which are then input into a support vector machine classifier. The penalty function parameters and kernel function parameters are optimized using grid search, and a cross-validation model is trained. The association validation model outputs the matching degree between the ICD-coded data and the drug catalog data based on the input ICD code and drug generic name.

2. The method for preventing and controlling medical corruption risks based on multi-source data fusion according to claim 1, characterized in that: Step S1 includes: Step S1-1: Directly obtain ICD code data, drug catalog data, and system operation log data through the hospital information system; The ICD coding data includes the ICD codes and hierarchical structure corresponding to the diseases; the hierarchical structure includes the chapters, categories, subcategories, and details of the ICD codes; The drug catalog data includes basic drug identifiers and therapeutic attributes; the basic drug identifiers include the generic name of the drug and its ATC classification code; the therapeutic attributes are the therapeutic relationship between the drug and the corresponding ICD codes of the disease. The system operation log data includes operator identification, operation time, and operation data; the operation data includes the patient's number and the ICD code data of the patient's disease, as well as the drug circulation record; the drug circulation record shows the destination of the drugs after they leave the warehouse. Step S1-2: Standardize the obtained ICD code data and drug catalog data: The standardization process for ICD code data is as follows: Regular expressions are used to verify the structure of the ICD codes; chapter, category, subcategory, and detail information is extracted from the ICD codes; a directed acyclic graph is constructed based on the hierarchical structure of chapters, categories, subcategories, and details, using the ICD codes as nodes; the co-occurrence frequency of different codes appearing simultaneously in the same medical record is statistically analyzed using case data; co-occurrence connections are established for code pairs whose co-occurrence frequency exceeds a preset threshold; and the ratio of the co-occurrence frequency to the maximum co-occurrence frequency among all code pairs is used as the weight of the co-occurrence connection. The standardization process for drug catalog data involves mapping different brand names of the same drug to a unique generic name; and converting the therapeutic attributes of the drug into a standard coded list using the official ICD mapping tool. Steps S1-3: Set the drug statistical period, obtain the usage frequency of all drugs corresponding to a certain disease within the drug statistical period according to the drug treatment attributes, and calculate the correlation strength between the drug and the disease based on the proportion of the usage frequency of a drug in the usage frequency of all drugs as the initial correlation relationship between ICD coding data and drug catalog data.

3. The method for preventing and controlling medical corruption risks based on multi-source data fusion according to claim 1, characterized in that: Step S4 includes: Step S4-1: Pair the patient's disease ICD code in the system operation log data with the generic name of the drug dispensed to the patient, input the pairing into the association verification model to obtain the operation matching degree of the pairing; set a matching degree threshold, and trigger a medical risk warning when the operation matching degree is less than the matching degree threshold, and record the operator's identifier and operation time corresponding to the dispensing operation; Step S4-2: Group the system operation log data according to the operator's identification, and sort each group of data according to the operation time to form a continuous operation time sequence for each operator; the continuous operation time sequence includes the pairing information of each outbound operation of the operator and the continuous operation time interval for the same patient. Step S4-3: Pre-train a long short-term memory network model using a continuous operation time sequence without medical risk; the long short-term memory network model is used to calculate the historical average continuous operation time interval for each operator; set a time interval threshold, if the deviation between the current continuous operation time interval of the operator and the historical average continuous operation time interval exceeds the time interval threshold, trigger a medical risk warning and record the operator's operator identifier.

4. A medical integrity risk prevention and control system based on multi-source data fusion, which is applied to the medical integrity risk prevention and control method based on multi-source data fusion as described in any one of claims 1-3, characterized in that: The system includes a data processing module, a feature encoding module, an association model module, and a risk warning module. The data processing module acquires data from the hospital system, performs standardization processing, and establishes an initial association between ICD-coded data and drug catalog data. The feature encoding module constructs a medical encoder, converting ICD-coded data and drug catalog data into corresponding feature vectors. The association model module calculates the association strength between ICD-coded data and drug catalog data and trains an association verification model. The risk warning module triggers medical risk warnings based on operational matching degree and deviation. The output of the data processing module is electrically connected to the input of the feature encoding module; the output of the feature encoding module is electrically connected to the input of the association model module; and the output of the association model module is electrically connected to the input of the risk warning module.

5. A medical integrity risk prevention and control system based on multi-source data fusion according to claim 4, characterized in that: The data processing module further includes a data acquisition unit and a data association unit; the data acquisition unit is used to acquire system operation log data, ICD code data and drug catalog data from the hospital system; the data association unit is used to standardize the ICD code data and drug catalog data and establish an initial association between the two based on the frequency of drug use.

6. A medical integrity risk prevention and control system based on multi-source data fusion according to claim 4, characterized in that: The feature encoding module further includes an ICD encoding unit and a drug encoding unit; the ICD encoding unit is used to construct an ICD encoder, which generates disease semantic vectors through one-hot encoding, word vector conversion and graph convolutional network processing; the drug encoding unit is used to construct a drug encoder, which generates drug feature vectors through word segmentation, pre-trained language models and convolutional neural network processing.

7. A medical integrity risk prevention and control system based on multi-source data fusion according to claim 4, characterized in that: The association model module further includes a strength calculation unit and a model training unit; the strength calculation unit is used to calculate the association strength between ICD coded data and drug catalog data using the cosine similarity method, and generate an association feature vector; the model training unit is used to input compliant and non-compliant data into a support vector machine classifier to train an association verification model.

8. A medical integrity risk prevention and control system based on multi-source data fusion according to claim 4, characterized in that: The risk warning module also includes a matching degree warning unit and a deviation warning unit; the matching degree warning unit is used to input system operation log data into the association verification model, and if the operation matching degree is less than the matching degree threshold, a medical risk warning is triggered. The deviation warning unit is used to analyze the continuous operation time interval of the operator. If the deviation of the operation time interval exceeds the time interval threshold, a medical risk warning is triggered.

Images