A discharge diagnosis and diagnosis and treatment process consistency quality inspection method, device and equipment

Abstract

The application discloses a discharge diagnosis and diagnosis and treatment process consistency quality inspection method, device and equipment, the method comprises the following steps: firstly, determining the target medical record to be inspected, and obtaining the target discharge diagnosis and target diagnosis and treatment process from the target medical record; then, obtaining the typical treatment medical record corresponding to the target discharge diagnosis from the pre-constructed typical treatment medical record library, and obtaining the typical diagnosis and treatment process from the typical treatment medical record; then, calculating the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process; and then, judging the consistency between the target discharge diagnosis and the target diagnosis and treatment process according to the calculation result of the similarity. Therefore, the pre-constructed typical treatment medical record library can be used to calculate the similarity between the target diagnosis and treatment process of the target medical record to be inspected and the typical diagnosis and treatment process of the typical treatment medical record corresponding to the target discharge diagnosis in the medical record library, and the quality inspection result with higher accuracy can be obtained according to the similarity calculation result.

Description

Technical Field

[0001] This application relates to the field of smart healthcare technology, and in particular to a quality inspection method, apparatus, and equipment for ensuring consistency between discharge diagnosis and treatment processes. Background Technology

[0002] The "discharge diagnosis" in a medical record typically refers to the first item on the discharge diagnosis list, i.e., the initial discharge visit, which represents the patient's primary medical purpose and the main disease treated during their hospitalization. Simultaneously, the patient's hospitalization data in the medical record also includes a "discharge summary," whose "treatment process" field fully records the patient's treatment process during this visit. The ability to accurately verify the consistency between the discharge diagnosis and the treatment process, thus ensuring the rationality of the discharge diagnosis, is particularly important for the effective implementation of Diagnosis Related Groups (DRG) payment.

[0003] Currently, there are generally two methods for quality control of the consistency between discharge diagnosis and treatment process in medical records: one is manual quality control, which can guarantee a very high accuracy rate, but has disadvantages such as high labor costs, slow quality control speed, and low efficiency; the other commonly used quality control method is quality control based on knowledge base rules. Although this method overcomes some of the problems of manual quality control, there has been relatively little research and attention on this level in the past, and the technical level is only a relatively superficial knowledge base with limited coverage. Therefore, quality control based on knowledge base can only cover a small part of the treatment methods involved, which will lead to inaccurate final quality control results. Summary of the Invention

[0004] The main objective of this application is to provide a quality inspection method, apparatus, and equipment for ensuring consistency between discharge diagnosis and treatment processes, thereby improving the accuracy of quality inspection results and providing technical support for the effective implementation of DRG payment.

[0005] This application provides a quality control method for ensuring consistency between discharge diagnosis and treatment process, including:

[0006] Identify the target medical records to be inspected, and obtain the target discharge diagnosis and target treatment process from the target medical records;

[0007] From a pre-built database of typical treatment medical records, obtain the typical treatment medical record corresponding to the target discharge diagnosis, and obtain the typical diagnosis and treatment process from the typical treatment medical record;

[0008] Calculate the similarity between the target treatment process and the typical treatment process;

[0009] Based on the similarity calculation results, the consistency between the target discharge diagnosis and the target treatment process is determined.

[0010] In one possible implementation, calculating the similarity between the target diagnostic and treatment process and the typical diagnostic and treatment process includes:

[0011] Construct a context graph corresponding to the typical diagnosis and treatment process; and construct a heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process.

[0012] The context graph corresponding to the typical diagnosis and treatment process, the heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process are jointly input into a pre-constructed graph network quality inspection model to identify the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process; the graph network quality inspection model is obtained by comparative training using a loss function based on positive sample medical record data and negative sample medical record data.

[0013] In one possible implementation, the graph network quality inspection model is constructed as follows:

[0014] Construct training sample medical record data, which includes positive sample medical record data and negative sample medical record data;

[0015] Construct a training context graph and a training heterogeneous graph corresponding to the diagnosis and treatment process in the medical record data of the training samples;

[0016] The training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in the training sample medical record data are input into the initial graph network quality inspection model, and the graph network quality inspection model is trained by adjusting the loss function.

[0017] The initial graph network quality inspection model includes a graph convolutional layer, a new graph construction layer, and a feature output layer.

[0018] In one possible implementation, the loss function is a hinge loss function; the loss function is used to shorten the distance between positive sample medical record data representations and to widen the distance between negative sample medical record data representations.

[0019] In one possible implementation, constructing the context graph corresponding to the typical diagnosis and treatment process includes:

[0020] The typical diagnosis and treatment process is segmented into words to obtain the segmented words contained in the typical diagnosis and treatment process;

[0021] The typical diagnostic and treatment process is traversed through each segmented word by a sliding window, and the dependency relationship between each segmented word is calculated using the mutual information PMI formula.

[0022] Based on the dependencies between the segmented words, a context graph corresponding to the typical diagnosis and treatment process is generated.

[0023] In one possible implementation, constructing a heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process includes:

[0024] The target treatment process and the typical treatment process are segmented into words to obtain the segmented words contained in each of the target treatment process and the typical treatment process.

[0025] Calculate the word vectors corresponding to each segmented word contained in the target diagnosis and treatment process and the typical diagnosis and treatment process;

[0026] The word vectors corresponding to each segmented word in the target treatment process and the word vectors corresponding to each segmented word in the typical treatment process are compared pairwise to calculate the adjacency matrix of the heterogeneous graphs corresponding to the target treatment process and the typical treatment process.

[0027] In one possible implementation, the method further includes:

[0028] Obtain the test medical records, and extract the test discharge diagnosis and test treatment process from the test medical records;

[0029] From the pre-built typical treatment medical record library, obtain the typical treatment medical record corresponding to the test discharge diagnosis, and obtain the typical diagnosis and treatment process from the typical treatment medical record;

[0030] Construct a context graph corresponding to the typical diagnostic and treatment process of the test; and construct a heterogeneous graph corresponding to the diagnostic and treatment process of the test and the typical diagnostic and treatment process of the test.

[0031] The context graph corresponding to the typical diagnostic and treatment process, the heterogeneous graph corresponding to the diagnostic and treatment process and the typical diagnostic and treatment process are input into the graph network quality inspection model to identify the test similarity between the diagnostic and treatment process and the typical diagnostic and treatment process.

[0032] When the test similarity does not exceed the preset threshold, the test medical records are used again as training sample medical records to update the graph network quality inspection model.

[0033] In one possible implementation, the graph network quality inspection model includes a first graph convolutional layer, a first new graph construction layer, a second graph convolutional layer, a second new graph construction layer, and a feature readout layer; the step of inputting the context graph corresponding to the typical diagnosis and treatment process, the heterogeneous graph corresponding to the target diagnosis and treatment process, and the typical diagnosis and treatment process into the pre-constructed graph network quality inspection model to identify the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process includes:

[0034] The context graph corresponding to the typical diagnosis and treatment process, the target diagnosis and treatment process, and the heterogeneous graph corresponding to the typical diagnosis and treatment process are jointly input into the first graph convolutional layer of the graph network quality inspection model for convolution processing to obtain the first initial feature vector after convolution processing.

[0035] The first initial feature vector is input into the first new graph construction layer to extract the features of the core words and obtain the first feature vector.

[0036] The first feature vector is input into the second graph convolutional layer for convolution processing to obtain the second initial feature vector after convolution processing.

[0037] The second initial feature vector is input into the second new graph construction layer, and the core words are extracted again to obtain the second feature vector;

[0038] The feature vectors corresponding to the heterogeneous graphs of the target diagnosis and treatment process and the typical diagnosis and treatment process, the first feature vector, and the second feature vector are concatenated to obtain a concatenated feature vector. The concatenated feature vector is then input into the feature readout layer for dimensionality reduction to obtain a dimensionality-reduced feature vector.

[0039] Based on the reduced-dimensional feature vector, the similarity between the target treatment process and the typical treatment process is determined.

[0040] This application also provides a quality inspection device for the consistency of discharge diagnosis and treatment process, including:

[0041] The first acquisition unit is used to determine the target medical record to be inspected and to acquire the target discharge diagnosis and target treatment process from the target medical record.

[0042] The second acquisition unit is used to acquire typical treatment records corresponding to the target discharge diagnosis from a pre-built typical treatment record database, and to acquire typical diagnosis and treatment processes from the typical treatment records.

[0043] A calculation unit is used to calculate the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process;

[0044] The quality control unit is used to determine the consistency between the target discharge diagnosis and the target treatment process based on the similarity calculation results.

[0045] In one possible implementation, the computational unit includes:

[0046] Construct sub-units for constructing a context graph corresponding to the typical diagnosis and treatment process; and constructing a heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process;

[0047] The identification subunit is used to input the context graph corresponding to the typical diagnosis and treatment process, the heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process into a pre-constructed graph network quality inspection model to identify the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process; the graph network quality inspection model is obtained by comparative training using a loss function based on positive sample medical record data and negative sample medical record data.

[0048] In one possible implementation, the device further includes:

[0049] The first construction unit is used to construct training sample medical record data, which includes positive sample medical record data and negative sample medical record data.

[0050] The second construction unit is used to construct the training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in the training sample medical record data;

[0051] The training unit is used to input the training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in the training sample medical record data into the initial graph network quality inspection model, and to train the graph network quality inspection model by adjusting the loss function.

[0052] The initial graph network quality inspection model includes a graph convolutional layer, a new graph construction layer, and a feature output layer.

[0053] In one possible implementation, the loss function is a hinge loss function; the loss function is used to shorten the distance between positive sample medical record data representations and to widen the distance between negative sample medical record data representations.

[0054] In one possible implementation, the building subunit includes:

[0055] The first word segmentation subunit is used to perform word segmentation processing on the typical diagnosis and treatment process to obtain the various word segments contained in the typical diagnosis and treatment process;

[0056] The first calculation subunit is used to traverse each segmented word contained in the typical diagnosis and treatment process through a sliding window, and to calculate the dependency relationship between each segmented word using the mutual information PMI formula.

[0057] A subunit is generated to generate a context graph corresponding to the typical diagnosis and treatment process based on the dependency relationships between the segmented words.

[0058] In one possible implementation, the building subunit includes:

[0059] The second word segmentation subunit is used to segment the target diagnosis and treatment process and the typical diagnosis and treatment process into words, respectively, to obtain the word segments contained in the target diagnosis and treatment process and the typical diagnosis and treatment process.

[0060] The second calculation subunit is used to calculate the word vectors corresponding to each segmented word contained in the target diagnosis and treatment process and the typical diagnosis and treatment process.

[0061] The third calculation subunit is used to perform pairwise similarity calculations between the word vectors corresponding to each segmented word in the target diagnosis and treatment process and the word vectors corresponding to each segmented word in the typical diagnosis and treatment process, so as to obtain the adjacency matrix of the heterogeneous graphs corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process.

[0062] In one possible implementation, the device includes:

[0063] The third acquisition unit is used to acquire test medical records and to acquire test discharge diagnoses and test treatment processes from the test medical records.

[0064] The fourth acquisition unit is used to acquire the test typical treatment medical record corresponding to the test discharge diagnosis from the pre-built typical treatment medical record library, and to acquire the test typical diagnosis and treatment process from the test typical treatment medical record;

[0065] The third construction unit is used to construct a context graph corresponding to the typical diagnostic and treatment process of the test; and to construct a heterogeneous graph corresponding to the diagnostic and treatment process of the test and the typical diagnostic and treatment process of the test.

[0066] The identification unit is used to input the context graph corresponding to the typical diagnostic and treatment process and the heterogeneous graph corresponding to the diagnostic and treatment process and the typical diagnostic and treatment process into the graph network quality inspection model, and identify the test similarity between the diagnostic and treatment process and the typical diagnostic and treatment process.

[0067] The update unit is used to update the graph network quality inspection model by reusing the test medical records as training sample medical records when the test similarity does not exceed the preset threshold.

[0068] In one possible implementation, the graph network quality inspection model includes a first graph convolutional layer, a first new graph construction layer, a second graph convolutional layer, a second new graph construction layer, and a feature readout layer; the recognition subunit includes:

[0069] The first convolutional subunit is used to input the context graph corresponding to the typical diagnosis and treatment process, the target diagnosis and treatment process and the heterogeneous graph corresponding to the typical diagnosis and treatment process into the first graph convolutional layer of the graph network quality inspection model for convolution processing to obtain the first initial feature vector after convolution processing.

[0070] The first extraction subunit is used to input the first initial feature vector into the first new graph construction layer to extract the features of core words and obtain the first feature vector.

[0071] The second convolutional subunit is used to input the first feature vector into the second graph convolutional layer for convolution processing to obtain the second initial feature vector after convolution processing.

[0072] The second extraction subunit is used to input the second initial feature vector into the second new graph construction layer, and perform feature extraction of core words again to obtain the second feature vector.

[0073] The dimensionality reduction subunit is used to concatenate the feature vectors corresponding to the heterogeneous graphs of the target diagnosis and treatment process and the typical diagnosis and treatment process, the first feature vector, and the second feature vector to obtain the concatenated feature vector, and input the concatenated feature vector into the feature readout layer for dimensionality reduction processing to obtain the dimensionality-reduced feature vector.

[0074] A sub-unit is defined to determine the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process based on the dimensionality-reduced feature vector.

[0075] This application embodiment also provides a quality inspection device for consistency between discharge diagnosis and treatment process, including: processor, memory, and system bus;

[0076] The processor and the memory are connected via the system bus;

[0077] The memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform any of the above-described quality control methods for consistency of discharge diagnosis and treatment processes.

[0078] This application also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any one of the above-described quality inspection methods for consistency between discharge diagnosis and treatment processes.

[0079] This application also provides a computer program product, which, when run on a terminal device, enables the terminal device to execute any one of the above-described quality inspection methods for consistency between discharge diagnosis and treatment processes.

[0080] This application provides a quality inspection method, apparatus, and equipment for ensuring the consistency of discharge diagnosis and treatment process. First, a target medical record to be inspected is identified, and the target discharge diagnosis and treatment process are obtained from the target medical record. Then, typical treatment records corresponding to the target discharge diagnosis are obtained from a pre-built typical treatment medical record database, and typical treatment processes are obtained from these typical treatment medical records. Next, the similarity between the target treatment process and the typical treatment processes is calculated. Based on the similarity calculation result, the consistency between the target discharge diagnosis and the target treatment process can be determined. Therefore, since this application utilizes a pre-built typical treatment medical record database to calculate the similarity between the target diagnosis and treatment process of the target medical record to be inspected and the typical treatment processes of all typical treatment medical records corresponding to the target discharge diagnosis in the database, and determines whether the target discharge diagnosis and the target treatment process are consistent based on the obtained similarity calculation result, the accuracy of the quality inspection results can be improved, thereby providing technical support for the effective implementation of DRG payment. Attached Figure Description

[0081] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0082] Figure 1 A flowchart illustrating a quality control method for ensuring consistency between discharge diagnosis and treatment processes, provided in an embodiment of this application;

[0083] Figure 2 A schematic diagram illustrating the calculation of the similarity between the target diagnosis and treatment process and a typical diagnosis and treatment process, provided for embodiments of this application;

[0084] Figure 3 One of the quality control flowcharts illustrating the consistency of discharge diagnosis and treatment processes provided in this application embodiment;

[0085] Figure 4 A second example diagram illustrating the quality control process for ensuring consistency between discharge diagnosis and treatment procedures in this application embodiment;

[0086] Figure 5 This is a schematic diagram of the composition of a quality inspection device for ensuring consistency between discharge diagnosis and treatment process, provided in an embodiment of this application. Detailed Implementation

[0087] To improve the quality of healthcare services, the "pay-per-disease" model, represented by Diagnosis Related Groups (DRG), is becoming increasingly popular. In the DRG grouping process, cases are first categorized into "major diagnostic categories" based on the "discharge diagnosis" in the settlement statement; then, based on the "discharge diagnosis" and "major surgery," cases are further categorized into "core disease-related groups"; finally, other diagnostic information is used to determine if there are comorbidities or accompanying diseases, resulting in the final grouping. Therefore, the "discharge diagnosis" is the most crucial and essential information in the DRG grouping payment method. Thus, accurately verifying the consistency between the discharge diagnosis and the treatment process during diagnosis and treatment, and achieving the goal of checking the rationality of the discharge diagnosis, is particularly important for the effective implementation of DRG payment.

[0088] Currently, the quality control methods for ensuring consistency between discharge diagnosis and treatment process in medical records include the following two:

[0089] The first method is manual quality inspection. Specifically, staff from the medical insurance bureau can mobilize professional insurance company inspectors to form a sampling inspection team, directly go to grassroots hospitals to collect data, and conduct inspections on multiple items, including the rationality of the initial discharge diagnosis.

[0090] While this manual quality inspection method guarantees a very high accuracy rate, it suffers from drawbacks such as high labor costs, slow inspection speed, and low efficiency. This is because each manual inspection requires a professional team, whose members are mostly drawn from hospitals or insurance companies, making the human resource composition inherently inconvenient. Furthermore, manual quality inspection is slow; a complete review of a medical record takes approximately 20 minutes, and even checking the consistency between the discharge diagnosis and treatment process takes about five minutes. Since DRG payment is to be widely implemented in numerous hospitals, involving a large volume of data processing, relying on manual quality inspection is clearly unsustainable and unenhancing.

[0091] The second commonly used quality control method is based on knowledge base rules. However, previous research and attention at this level have been relatively limited, and the technical aspects have only involved relatively superficial knowledge bases. This typically involves constructing a knowledge base of the main treatment methods for a disease, followed by character-level comparison for judgment. While this method overcomes some of the problems of manual quality control, the limited coverage of the knowledge base means that, based on past analysis and experience, doctors essentially modify and refine medical records based on the hospital's information system. Firstly, there are numerous information system vendors, each developing different medical record templates; secondly, there is no unified standard for doctors writing medical records, relying primarily on the habits and experience of their hospital and department. Therefore, quality control based on a knowledge base can only cover a small portion of the treatment methods involved, leading to potentially inaccurate final quality control results.

[0092] To address the aforementioned deficiencies, this application provides a quality inspection method for ensuring the consistency of discharge diagnoses and treatment processes. First, a target medical record to be inspected is identified, and the target discharge diagnosis and treatment process are obtained from this record. Then, typical treatment records corresponding to the target discharge diagnosis are retrieved from a pre-built database of typical treatment records, and typical treatment processes are obtained from these records. Next, the similarity between the target treatment process and the typical treatment processes is calculated. Based on the similarity calculation result, the consistency between the target discharge diagnosis and the target treatment process can be determined. Therefore, since this application utilizes a pre-built database of typical treatment records to calculate the similarity between the target diagnosis and treatment process of the target medical record and the typical treatment processes of all typical treatment records corresponding to the target discharge diagnosis in the database, and determines whether the target discharge diagnosis and the target treatment process are consistent based on the obtained similarity calculation result, the accuracy of the quality inspection results can be improved, thereby providing technical support for the effective implementation of DRG payment.

[0093] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0094] First Embodiment

[0095] See Figure 1 This is a flowchart illustrating a quality control method for ensuring consistency between discharge diagnosis and treatment process, as provided in this embodiment. The method includes the following steps:

[0096] S101: Identify the target medical records to be inspected and obtain the target discharge diagnosis and target treatment process from the target medical records.

[0097] In this embodiment, any medical record used for quality control of the consistency between discharge diagnosis and treatment process according to this embodiment is defined as the target medical record to be inspected. It should be noted that this embodiment does not limit the language type of the target medical record; for example, the target medical record can be a Chinese medical record or an English medical record, etc. This embodiment also does not limit the length of the target medical record; for example, the target medical record can be a sentence text or a paragraph text, etc. This embodiment also does not limit the type of disease to which the target medical record belongs; for example, the target medical record can be a medical record with a discharge diagnosis of glaucoma or a medical record with a discharge diagnosis of gastric malignancy, etc.

[0098] Once the target medical record to be inspected is identified, the discharge diagnosis and treatment process can be extracted from the target medical record and defined as the target discharge diagnosis and target treatment process (and represented by a query), so that accurate quality inspection of the consistency between the target discharge diagnosis and the target treatment process can be achieved through subsequent steps S102-S104.

[0099] S102: Obtain the typical treatment medical record corresponding to the target discharge diagnosis from the pre-built typical treatment medical record database, and obtain the typical diagnosis and treatment process from the typical treatment medical record.

[0100] It should be noted that, in order to improve the accuracy of quality control regarding the consistency between discharge diagnoses and treatment processes, this application employs a typical medical record method to construct a database of typical treatment medical records related to the diagnoses of various typical diseases. This database contains relevant typical treatment medical records corresponding to the diagnoses of various typical diseases, such as the treatment medical records corresponding to the discharge diagnosis of "glaucoma." Each medical record records a detailed diagnosis and treatment process for the discharge diagnosis of "glaucoma," such as: "After admission, relevant examinations were completed, surgical contraindications were ruled out, and on February 5, 2018, a 'complex trabeculectomy of the left eye' was performed under local anesthesia. The surgery was successful, and postoperatively, the anterior chamber of the patient was well formed, and aqueous humor outflow was smooth. The patient's vital signs were stable during and after the surgery, and the patient returned to the ward safely, receiving anti-inflammatory and other symptomatic treatments." In other words, in the typical treatment medical record database, the discharge diagnoses and treatment processes contained in each medical record are consistent.

[0101] Based on this, in order to accurately check the consistency between the target discharge diagnosis and the target treatment process, after extracting the target discharge diagnosis from the target medical record, the typical treatment medical records corresponding to the target discharge diagnosis can be further selected from the pre-built typical treatment medical record library. For example, character matching can be used to select the diagnosis result corresponding to the target discharge diagnosis from the pre-built typical treatment medical record library, and then the typical treatment medical record to which the diagnosis result belongs can be obtained. The typical treatment process can be obtained from each typical treatment medical record (and represented using a document) to execute the subsequent step S103.

[0102] For example, if the target discharge diagnosis is "glaucoma", then the typical treatment records corresponding to "glaucoma" can be screened from the pre-built typical treatment record database, and then the typical diagnosis and treatment process can be screened from each typical treatment record to execute the subsequent step S103.

[0103] S103: Calculate the similarity between the target treatment process and the typical treatment process.

[0104] In this embodiment, after obtaining the target diagnosis and treatment process query from the target medical record in step S101 and the typical diagnosis and treatment process document from the typical treatment medical record in step S102, the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process can be calculated using existing or future similarity calculation methods, so as to execute the subsequent step S104.

[0105] A preferred implementation is to use a graph construction method that combines the target treatment process query and the typical treatment process document, and then use a pre-built graph network quality inspection model to calculate the similarity between the target treatment process and the typical treatment process, thereby improving the accuracy of the calculation results. Specifically, the implementation process of step S103 may include the following steps S1031-S1032:

[0106] Step S1031: Construct a context graph corresponding to a typical diagnosis and treatment process; and construct a heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process.

[0107] In this implementation, to accurately calculate the similarity between the target treatment process query and the typical treatment process document, it is first necessary to construct a context graph corresponding to the typical treatment process query (such as the "typical medical record context graph" in 2(a)) and a heterogeneous graph corresponding to the target treatment process query and the typical treatment process document (such as... Figure 2 (b) "query-doc heterogeneous graph" in the text.

[0108] Specifically, such as Figure 2As shown, to construct the context graph corresponding to a typical diagnosis and treatment process document, it is first necessary to segment the document using existing or future segmentation methods. For example, existing tool libraries can be used for jieba segmentation to obtain the segmented words contained in the document. Then, a sliding window can be used to traverse the segmented words in the document, and the dependency relationships between the segmented words can be calculated using the Pointwise Mutual Information (PMI) formula. Based on these dependencies, a context graph corresponding to the typical diagnosis and treatment process is generated. If two segmented words frequently appear as contexts of each other, an edge is constructed for them, with a weight of a large PMI index. If two segmented words do not frequently appear as contexts, the weight of their edge is small or below the minimum threshold, and no edge is generated. The generated context graph corresponding to the typical diagnosis and treatment process document can be denoted as D∈R. n×n ,like Figure 2 As shown, n represents the number of word segments contained in the document.

[0109] Similarly, after constructing the context graph corresponding to a typical diagnosis and treatment process document, such as Figure 2 As shown, a heterogeneous graph corresponding to the target treatment process query and the typical treatment process document can also be constructed. Specifically, firstly, existing or future word segmentation methods are still needed to segment the target treatment process query and the typical treatment process document. For example, jieba segmentation can still be performed on both to obtain the individual words contained in each. Then, existing or future word vector extraction methods can be used to extract the word vector representations corresponding to each word contained in the target treatment process query and the typical treatment process document. For example, the Word2vec word vector acquisition tool can be used to obtain the vector representation of each word in the query and document, and denoted as w respectively. q and w d The dimension of both can be 300 (or other values; this is just an example), i.e., w q ∈R 1×300 w d ∈R 1×300 Furthermore, pairwise similarity calculations can be performed between the word vectors corresponding to each segmented word in the target diagnosis process query and the word vectors corresponding to each segmented word in the typical diagnosis process document to obtain the adjacency matrix of the heterogeneous graphs corresponding to the query and document, denoted as A∈R.m×n ,like Figure 2 As shown, m represents the total number of segmented words in the query; n, as above, still represents the number of segmented words contained in the document; a ij Let A represent the edge weights between the i-th word in the query and the j-th word in the document. Therefore, the heterogeneous graph represented by A contains two types of nodes: nodes representing words in the query and nodes representing words in the document, such as... Figure 2 As shown in (b) "query-doc heterogeneous graph", there is only one type of edge, namely the edge between nodes of different types, and there is no edge between nodes of the same type.

[0110] Step S1032: Input the context graph corresponding to the typical diagnosis and treatment process, the heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process into the pre-built graph network quality inspection model to identify the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process; wherein, the graph network quality inspection model is obtained by comparative training using a loss function based on positive sample medical record data and negative sample medical record data.

[0111] In this implementation, after constructing the context graph corresponding to the typical diagnosis and treatment process query in step S1031, and the heterogeneous graph corresponding to the target diagnosis and treatment process query and the typical diagnosis and treatment process document, both can be input into the pre-constructed graph network quality inspection model, such as... Figure 2 As shown, the similarity between the target diagnostic process query and the typical diagnostic process document is identified in order to execute the subsequent step S104.

[0112] The graph network quality inspection model is obtained by comparing and training positive and negative sample medical records using a loss function. For details of the construction process, please refer to the following steps (1)-(3).

[0113] It should be noted that graph network quality inspection models can contain three types of network layers: graph convolutional layers, new graph construction layers, and feature readout layers. For example... Figure 2 As shown, the graph convolutional layer and the new graph construction layer can be deployed at multiple granularities according to the actual situation. This application does not limit this, that is, Figure 2 The graph convolutional layer and new graph construction layer in the graph are not limited to the single layer shown in the graph; they can also contain multiple graph convolutional layers and new graph construction layers. Figure 2 This is just an example; for instance, two graph convolutional layers and a new graph construction layer could be deployed. It should be noted that the post-text application will combine... Figure 2(The diagram only shows one graph convolutional layer and one new graph construction layer.) The graph network quality inspection model consisting of two graph convolutional layers, one new graph construction layer, and one feature readout layer will be used as an example for introduction. Other composition methods can be implemented in the same way, and will not be described in detail.

[0114] Specifically, one possible implementation is that the graph network quality inspection model can include a first graph convolutional layer, a first new graph construction layer, a second graph convolutional layer, a second new graph construction layer, and a feature readout layer; then the implementation process of "inputting the context graph corresponding to the typical diagnosis and treatment process, the heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process into the pre-constructed graph network quality inspection model to identify the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process" in step S1032 above can specifically include the following step AF:

[0115] Step A: Input the context graph corresponding to the typical diagnosis and treatment process, the heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process into the first graph convolutional layer of the graph network quality inspection model for convolution processing to obtain the first initial feature vector after convolution processing.

[0116] In this implementation, step S1031 constructs a context graph D∈R corresponding to a typical diagnosis and treatment process query. n×n Furthermore, a heterogeneous graph A∈R corresponding to the target diagnosis and treatment process query and the typical diagnosis and treatment process document was constructed. m×n Then, both can be input together into the first graph convolutional layer of the pre-built graph network quality inspection model for convolution processing, such as... Figure 2 As shown, the first initial feature vector after convolution is obtained.

[0117] The first graph convolutional layer performs multiple convolution operations based on the topological graph. Each convolution aggregates first-order neighbor features, and multiple convolutions can aggregate multi-order neighbor features. First, a is obtained according to the following formula (1). t , representing the heterogeneous graph after the query process has been filtered.

[0118] a t =AD (1)

[0119] Where A represents the heterogeneous graph corresponding to the target diagnosis and treatment process query and the typical diagnosis and treatment process document; D represents the context graph corresponding to the typical diagnosis and treatment process query; a t The result of multiplying the structural vectors of A and D is represented by t; t represents the t-th network layer.

[0120] To characterize the varying importance of features from different neighbors at different orders to the representation of the central node, this embodiment can also incorporate a gating mechanism into the convolutional result, namely, an update gate and a forget gate, corresponding to formulas (2) and (3) below, respectively, where W, U, and b represent trainable weights and bias parameters, respectively; σ represents the activation function. The parameters contained in these two gates can control whether and to what extent neighbor node features are adopted, where z... t r represents the features that need to be updated after convolution. t This represents the features that are forgotten (i.e., do not need to be updated) after convolution.

[0121] z t =σ(W z a t +U z h t-1 +b z (2)

[0122] r t =σ(W r a t +U r h t-1 +b r (3)

[0123] Subsequently, in this embodiment of the application, the forget gate can be used to calculate the features after the previous (i.e., the (t-1)th) convolution to obtain the features after the forget gate. (The calculation formula is shown in formula (4) below), and then, the final information aggregation is performed to represent the result of a query and a document matching (i.e., the first initial feature vector) h. t The specific calculation formula is shown in formula (5) below:

[0124]

[0125]

[0126] Among them, h t This represents the first initial feature vector obtained after convolution processing through the first graph convolutional layer.

[0127] Step B: Input the first initial feature vector into the first new graph construction layer to extract the features of the core words and obtain the first feature vector.

[0128] It should be noted that although two layers of graph convolutional layers can be deployed, for the diagnosis and treatment process (including the typical diagnosis and treatment process and the target diagnosis and treatment process), only some core words truly describe the treatment, and the other parts may just be connecting words. For example, for the diagnosis and treatment process "After the patient was admitted to the hospital, relevant examinations were completed, surgical contraindications were excluded, and a 'left eye combined trabeculectomy' was performed under local anesthesia on February 5, 2018. The operation was successful. After the operation, the anterior chamber of the patient formed well, and the aqueous humor outflow was smooth. The patient's vital signs were stable during and after the operation, and the patient returned to the ward safely. Anti-inflammatory and other symptomatic treatments were given", the parts that describe the treatment are only the core words such as "completed relevant examinations, excluded surgical contraindications, and a 'left eye combined trabeculectomy' was performed under local anesthesia on February 5, 2018. The operation was successful" and "the patient's vital signs were stable during and after the operation, and the patient returned to the ward safely. Anti-inflammatory and other symptomatic treatments were given", and the other parts may just be connecting words.

[0129] Therefore, in order to improve the recognition accuracy and efficiency, after obtaining the first initial feature vector h after convolution processing by the first layer of graph convolutional layer through step A t further, the first initial feature vector h t can be input into the first layer of new graph construction layer to extract the features of core words, such as Figure 2 shown, to obtain the first feature vector h t’ for performing subsequent step C.

[0130] x Specifically, after obtaining the first initial feature vector h after convolution processing by the first layer of graph convolutional layer t the value of h t can be used as the relevance score with each other word, and then ranked (rank operation) according to the scores, and the nodes with lower rankings are removed. That is, h t ∈R m×n —>h t' ∈R m×n' where n' represents the number of nodes retained after ranking, which can be set according to the actual situation and empirical values, and satisfies n' < n. The entire implementation process is as Figure 2 shown in the "new graph construction layer" of the model part in

[0131] Step C: Input the first feature vector into the second layer of graph convolutional layer for convolution processing to obtain the second initial feature vector after convolution processing.

[0132] After obtaining the first feature vector h t’ through step B, further, the first feature vector h t’ can be input into the second layer of new graph construction layer for convolution processing to obtain the second initial feature vector h t+1This is used to execute the subsequent step D. The specific implementation process is similar to that of step A, and will not be described in detail here.

[0133] Step D: Input the second initial feature vector into the second new graph construction layer, and perform feature extraction of core words again to obtain the second feature vector.

[0134] Step C yields the second initial feature vector h, which has undergone convolution processing by the second graph convolutional layer. t+1 Then, the second initial feature vector h can be further... t+1 The second new graph construction layer is input, and the core word features are extracted again to obtain the second feature vector h. (t+1)’ This is used to execute the subsequent step E. The specific implementation process is similar to step B, and will not be described in detail here.

[0135] Step E: Concatenate the feature vectors, first feature vector, and second feature vector of the heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process to obtain the concatenated feature vector. Input the concatenated feature vector into the feature readout layer for dimensionality reduction to obtain the dimensionality-reduced feature vector.

[0136] The first feature vector h is obtained through steps B and D respectively. t’ The second eigenvector h (t+1)' Then, the two can be further concatenated with the target diagnosis and treatment process obtained through step S1031 and the heterogeneous graph A corresponding to the typical diagnosis and treatment process to obtain the concatenated feature vector, such as... Figure 2 As shown in the "multi-scale concatenation" diagram, the concatenated feature vector is input into the feature readout layer (specifically including, etc.). Figure 2 The two fully connected layers shown are subjected to dimensionality reduction processing to reduce the dimension to 1, resulting in a dimensionality-reduced feature vector, which is then used to perform the subsequent step F.

[0137] Step F: Determine the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process based on the dimensionality-reduced feature vector.

[0138] After obtaining the dimensionality-reduced feature vector through E, it can be further normalized using the sigmoid function to determine the final similarity between the query and the document. Figure 2 The output socre is shown in the figure. The specific calculation formula is as follows:

[0139]

[0140] Where | indicates horizontal concatenation; sum_pooling indicates summation.

[0141] S104: Based on the similarity calculation results, determine the consistency between the target discharge diagnosis and the target treatment process.

[0142] In this embodiment, after calculating the similarity between the final query and the document through step S103 (as in the above formula (6)), it can be further determined whether the target discharge diagnosis is consistent with the target treatment process based on the effective analysis of the calculation results of the similarity.

[0143] Specifically, to improve the accuracy of judging the consistency between the target discharge diagnosis and the target treatment process, a preset threshold for similarity can be set based on the actual situation or experience. After obtaining the similarity between the query and the document, the consistency between the target discharge diagnosis and the target treatment process can be determined by judging whether the similarity exceeds the preset threshold. That is, judging whether the similarity exceeds the preset threshold. If it does, it means that the target treatment process query and the typical treatment process document are consistent, and thus it can be determined that the target treatment process query and the target discharge diagnosis are consistent. If not, it means that the target treatment process query and the typical treatment process document are inconsistent, and thus it can be determined that the target discharge diagnosis and the target treatment process are inconsistent. At the same time, various preset methods (such as voice or text) can be used to provide corresponding error reminders.

[0144] The specific value of the preset threshold can be set according to the actual situation. This application embodiment does not limit this value and can set it to any value between 0 and 1. For example, the preset threshold can be set to 0.45.

[0145] It should be noted that patients often have multiple purposes when they come to the hospital. Not all patients undergo the complete treatment process after hospitalization. Some may only want conservative treatment, some may give up treatment due to financial reasons or their own physical condition, some may choose to be discharged as soon as possible because they want to go to a higher-level hospital for treatment, and some may only want to have a check-up at home and prepare for treatment later. This application has adopted corresponding handling methods for these situations.

[0146] Specifically, firstly, medical records where patients did not receive adequate treatment upon arrival at the hospital can be identified in advance. That is, if the consultation process only includes a diagnosis and no treatment-related descriptions, these records will be categorized and tagged as lacking treatment, awaiting manual processing later. For patients who only received simple or conservative treatment, disease-specific treatment knowledge can be used, breaking down the treatment into smaller granularities such as medication names, surgical procedures, and operational names. In this way, as long as such names are mentioned during the consultation, or even just one medication, it can be considered as treatment being performed, rather than simply judging it as a discrepancy between the discharge diagnosis and the treatment process.

[0147] For example: Figure 3 As shown, for the target discharge diagnosis of "mitral valve prolapse," after retrieving eight typical treatment records from the typical treatment record database and extracting these eight typical treatment processes, the similarity between the target treatment process and these eight typical treatment processes can be calculated, yielding eight similarity values: 0.1, 0.02, 0.01, 0.05, 0.05, 0.02, 0.06, and 0.01. Assuming a preset threshold of 0.45, it can be seen that none of these eight similarity values ​​exceed the preset threshold of 0.45, and the target medical record does not belong to the category of no treatment or lack of treatment. Therefore, it can be determined that the discharge diagnosis and treatment process of this medical record are inconsistent. Furthermore, it can be determined that the case underwent urethral cystolithotomy, and the corresponding discharge diagnosis is more likely a diagnosis of multiple bladder stones.

[0148] Or, such as Figure 4 As shown, for the target discharge diagnosis "rotator cuff injury", after obtaining the corresponding 6 typical treatment records from the typical treatment record database and extracting these 6 typical treatment processes, the similarity between the target treatment process and these 6 typical treatment processes can be calculated, resulting in 6 similarity values ​​of 0.95, 0.05, 0.01, 0.71, 0.21, and 0.75. Assuming the preset threshold is 0.45, it can be seen that 0.95, 0.71, and 0.75 are all higher than the preset threshold of 0.45. Therefore, among these 6 typical treatment records, 3 have a similarity higher than the threshold with the target treatment process in the target medical record to be tested. Thus, it can be determined that the discharge diagnosis and treatment process of this medical record are consistent.

[0149] Next, this application will introduce the construction process of the graph network quality inspection model mentioned in step S1032 above. The graph network quality inspection model is obtained by comparative training using a loss function based on positive sample medical record data and negative sample medical record data. The specific construction process may include the following steps (1)-(3):

[0150] Step (1): Construct training sample medical record data, which includes positive sample medical record data and negative sample medical record data.

[0151] It should be noted that the graph network quality inspection model is trained by comparing positive and negative sample medical record data using a loss function. Through the constraints of this model and the loss function, the distance between the test data and the positive sample representation can be continuously narrowed, while the distance between the test data and the negative sample representation can be widened, ultimately obtaining a model with a more discriminative ability to characterize the test data. Therefore, the construction of positive and negative sample data is crucial for training a model with higher computational accuracy.

[0152] However, the treatment process corresponding to a discharge diagnosis is not unique; it varies depending on the patient's actual condition, their own wishes, the doctor's skill level, and the hospital's resources. For example, for the same disease, some people may choose surgical treatment, while others may only choose drug treatment for other reasons. In special cases, some people may have drug allergies and can only recover on their own. For the same disease, hospitals with abundant resources or highly skilled doctors will have more professional treatment methods, such as using more sophisticated instruments and performing more complex surgeries, while in resource-limited areas, only conservative and basic treatment methods can be chosen.

[0153] Therefore, in this embodiment, to construct a model with higher computational accuracy, this application constructs training sample medical record data from the perspective of a pre-built typical treatment medical record database, rather than from the perspective of clinical knowledge and books. The training sample medical record data includes positive sample medical record data and negative sample medical record data.

[0154] Specifically, when constructing positive sample medical record data, for a discharge diagnosis, existing treatment medical records are first selected from a pre-built typical treatment medical record database, and then the treatment process fields are filtered out. After data cleaning and filtering, data discharged without specific treatment (such as no treatment or simple processing) and other types of low-quality data are removed. Doctors then annotate the data, and during the annotation process, if doctors find that the quality of cases corresponding to the disease is generally poor, they will construct several typical treatment medical records as support based on past experience. After confirmation by doctors, the final training discharge diagnosis and its corresponding typical treatment data can be saved as positive sample data.

[0155] For example, when constructing a positive sample of data for a patient discharged with a diagnosis of "glaucoma," the corresponding treatment data could be: "After admission, relevant examinations were completed, surgical contraindications were ruled out, and on February 5, 2018, a 'compound trabeculectomy of the left eye' was performed under local anesthesia. The surgery was successful, and postoperatively, the anterior chamber formed well, and aqueous humor flowed smoothly. The patient's vital signs were stable during and after the surgery, and the patient returned to the ward safely, receiving anti-inflammatory and other symptomatic treatments." In this case, trabeculectomy was performed, which is a typical treatment for glaucoma.

[0156] After constructing the positive sample medical record data, when constructing the negative sample medical record data—specifically, for cases where the discharge diagnosis and treatment process are inconsistent—random sampling cannot be used. This is because for many diseases, the treatment methods inherently overlap. For example, regardless of the location of the inflammation, the treatment methods will include anti-inflammatory therapy, typically using anti-inflammatory drugs; similarly, most malignant tumor treatments include chemotherapy and radiotherapy, only targeting different tumor locations. Therefore, when constructing the negative sample data, we first need to identify other diseases whose treatment methods do not overlap with the current disease, and then randomly select typical treatment data from the typical treatment records corresponding to these diseases as negative sample data for storage.

[0157] The following is an example of a negative sample: a patient with gastric cancer undergoes a heart-related surgery.

[0158] For example, when constructing a negative sample of a patient discharged with a diagnosis of "malignant gastric tumor," the typical treatment data corresponding to this diagnosis can be used as the corresponding treatment process. For instance, it could be: "After admission, relevant examinations were completed, surgical contraindications were ruled out, and after active preoperative preparation, atrial septal defect repair and tricuspid valve repair were performed under general anesthesia, cardiopulmonary bypass, and video-assisted thoracoscopic surgery. The surgery was successful, and postoperative care included cardiotonics, diuretics, and nutritional support. The patient recovered satisfactorily, vital signs were stable, and follow-up examinations showed no abnormalities. The patient was discharged."

[0159] Step (2): Construct the training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in the medical record data of the training samples.

[0160] After constructing a large amount of training sample medical record data including positive sample medical record data and negative sample medical record data through step (1), the training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in each positive sample medical record data and negative sample medical record data in the training sample medical record data can be constructed in a similar manner to step S1031 above, so as to obtain the graph network quality inspection model by executing the subsequent step (3).

[0161] Step (3): Input the training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in the training sample medical record data into the initial graph network quality inspection model, and train the graph network quality inspection model by adjusting the loss function.

[0162] In this embodiment, the initial graph network quality inspection model includes a graph convolutional layer, a new graph construction layer, and a feature output layer. The number of graph convolutional layers and new graph construction layers is not limited; for example, they can be set to two layers, as described in the model example above.

[0163] Based on this, after constructing the training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in each positive sample medical record data and negative sample medical record data in step (2), the training context graph and training heterogeneous graph corresponding to each sample diagnosis and treatment process pair (including positive sample pair and negative sample pair) can be input into the initial graph network quality inspection model for training. The model outputs the similarity between the two diagnosis and treatment processes in each sample diagnosis and treatment process pair (including positive sample pair and negative sample pair). Through multiple rounds of model training, the training termination condition is met. At this point, the graph network quality inspection model is obtained.

[0164] In each round of model training, a pre-defined loss function is used. This loss function is designed to narrow the distance between positive sample medical record data representations and widen the distance between negative sample medical record data representations. By adjusting this loss function, the correlation between each pair of treatment processes (including positive and negative sample pairs) can be constrained, resulting in increasing similarity between the two treatment processes in a positive sample pair and decreasing similarity between the two treatment processes in a negative sample pair.

[0165] Specifically, in order to train a graph network quality inspection model with better recognition performance, this embodiment uses the hinge loss function during training, and the specific calculation formula is as follows:

[0166] loss = max(0, 1- <q,d + >+ <q,d - >) (7)

[0167] in, <q,d + > indicates the similarity between the input diagnostic process to be tested and another diagnostic process in its corresponding positive sample pair; <q,d - > represents the similarity between the input diagnostic process and the other diagnostic process in its negative sample pair. It can be seen that only when the positive sample pairs become more and more similar (i.e., the similarity between the two diagnostic processes in the positive sample pair is higher) and the negative sample pairs become more and more dissimilar (i.e., the similarity between the two diagnostic processes in the negative sample pair is lower) will the value of the loss function decrease, and thus the constraint target of the loss function be achieved.

[0168] In this way, through multiple rounds of training, the value of the loss function is continuously adjusted until the value of the loss function converges, for example, until the value of the loss function reaches its minimum and remains basically unchanged. At this point, the training can be stopped, and the graph network quality inspection model obtained at this time can be used as the trained graph network quality inspection model.

[0169] Through the above steps (1)-(3), a graph network quality inspection model can be trained and generated based on the sample diagnosis and treatment process (including positive sample pairs and negative sample pairs). Furthermore, the generated graph network quality inspection model can be tested and verified using test medical records. The specific verification process may include the following steps (1)-5):

[0170] Step 1): Obtain the test medical record and extract the test discharge diagnosis and test treatment process from the test medical record.

[0171] Step 2): Obtain the test typical treatment medical record corresponding to the test discharge diagnosis from the pre-built typical treatment medical record library, and obtain the test typical diagnosis and treatment process from the test typical treatment medical record.

[0172] In this embodiment, in order to test and verify the graph network quality inspection model, it is first necessary to obtain test medical records. For example, a method similar to constructing sample treatment process pairs (including positive sample pairs and negative sample pairs) can be used. After obtaining the test medical records, not only can the test discharge diagnosis and test treatment process be obtained from them, but also the typical treatment medical records corresponding to the test discharge diagnosis need to be obtained from the pre-constructed typical treatment medical record library as test typical treatment medical records, and the test typical treatment process needs to be obtained from the test typical treatment medical records.

[0173] Step 3): Construct a context graph corresponding to the typical diagnostic and treatment process; and construct a heterogeneous graph corresponding to the diagnostic and treatment process and the typical diagnostic and treatment process.

[0174] After obtaining the test diagnosis and treatment process and the typical test diagnosis and treatment process through steps 1) and 2), the context graph corresponding to the typical test diagnosis and treatment process and the heterogeneous graph corresponding to the test diagnosis and treatment process and the typical test diagnosis and treatment process can be constructed in a manner similar to step S1031 above, so as to execute the subsequent step 4).

[0175] Step 4): Input the context graph corresponding to the typical diagnostic and treatment process, the heterogeneous graph corresponding to the diagnostic and treatment process and the typical diagnostic and treatment process into the graph network quality inspection model to identify the test similarity between the diagnostic and treatment process and the typical diagnostic and treatment process.

[0176] After constructing the context graph corresponding to the typical diagnostic and treatment process in step (3), and the heterogeneous graph corresponding to the diagnostic and treatment process and the typical diagnostic and treatment process, the two can be input into the graph network quality inspection model to identify the test similarity between the diagnostic and treatment process and the typical diagnostic and treatment process, so as to execute the subsequent step (5).

[0177] Step 5): When the test similarity does not exceed the preset threshold, the test medical records are used again as training sample medical records to update the graph network quality inspection model.

[0178] After obtaining the test similarity between the test treatment process and the typical test treatment process in step 4), if the test similarity does not exceed the preset threshold set manually, the test medical record can be used again as a training sample medical record. That is, the test treatment process and the typical test treatment process are used again as positive sample pairs to update the graph network quality inspection model. For the testing of negative sample pairs, a similar testing and model update can be performed by referring to the testing process of positive sample pairs. The specific testing and update process will not be elaborated here.

[0179] Through the above embodiments, the graph network quality inspection model can be effectively verified using test medical records. When the test similarity does not meet the preset threshold, the graph network quality inspection model can be adjusted and updated in a timely manner, thereby helping to improve the model's recognition accuracy and precision.

[0180] In summary, this embodiment provides a quality inspection method for ensuring the consistency of discharge diagnosis and treatment process. First, the target medical record to be inspected is identified, and the target discharge diagnosis and treatment process are obtained from it. Then, typical treatment records corresponding to the target discharge diagnosis are obtained from a pre-built database of typical treatment records, and typical treatment processes are obtained from these records. Next, the similarity between the target treatment process and the typical treatment processes is calculated. Based on the similarity calculation result, the consistency between the target discharge diagnosis and the target treatment process can be determined. Therefore, this embodiment utilizes a pre-built database of typical treatment records to calculate the similarity between the target diagnosis and treatment process of the target medical record and the typical treatment processes of all typical treatment records corresponding to the target discharge diagnosis in the database. The consistency between the target discharge diagnosis and the target treatment process is then determined based on the similarity calculation result, thereby improving the accuracy of the quality inspection results and providing technical support for the effective implementation of DRG payment.

[0181] Second Embodiment

[0182] This embodiment will introduce a quality inspection device for ensuring consistency between discharge diagnosis and treatment process; please refer to the above method embodiment for related content.

[0183] See Figure 5 This is a schematic diagram of the composition of a quality inspection device for consistency between discharge diagnosis and treatment process provided in this embodiment. The device 500 includes:

[0184] The first acquisition unit 501 is used to determine the target medical record to be inspected and to acquire the target discharge diagnosis and target treatment process from the target medical record.

[0185] The second acquisition unit 502 is used to acquire typical treatment records corresponding to the target discharge diagnosis from a pre-built typical treatment record database, and to acquire typical diagnosis and treatment processes from the typical treatment records.

[0186] The calculation unit 503 is used to calculate the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process;

[0187] The quality inspection unit 504 is used to determine the consistency between the target discharge diagnosis and the target treatment process based on the similarity calculation results.

[0188] In one implementation of this embodiment, the calculation unit 503 includes:

[0189] Construct sub-units for constructing a context graph corresponding to the typical diagnosis and treatment process; and constructing a heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process;

[0190] The identification subunit is used to input the context graph corresponding to the typical diagnosis and treatment process, the heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process into a pre-constructed graph network quality inspection model to identify the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process; the graph network quality inspection model is obtained by comparative training using a loss function based on positive sample medical record data and negative sample medical record data.

[0191] In one implementation of this embodiment, the apparatus further includes:

[0192] The first construction unit is used to construct training sample medical record data, which includes positive sample medical record data and negative sample medical record data.

[0193] The second construction unit is used to construct the training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in the training sample medical record data;

[0194] The training unit is used to input the training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in the training sample medical record data into the initial graph network quality inspection model, and to train the graph network quality inspection model by adjusting the loss function.

[0195] The initial graph network quality inspection model includes a graph convolutional layer, a new graph construction layer, and a feature output layer.

[0196] In one implementation of this embodiment, the loss function is a hinge loss function; the loss function is used to shorten the distance between positive sample medical record data representations and to widen the distance between negative sample medical record data representations.

[0197] In one implementation of this embodiment, the construction subunit includes:

[0198] The first word segmentation subunit is used to perform word segmentation processing on the typical diagnosis and treatment process to obtain the various word segments contained in the typical diagnosis and treatment process;

[0199] The first calculation subunit is used to traverse each segmented word contained in the typical diagnosis and treatment process through a sliding window, and to calculate the dependency relationship between each segmented word using the mutual information PMI formula.

[0200] A subunit is generated to generate a context graph corresponding to the typical diagnosis and treatment process based on the dependency relationships between the segmented words.

[0201] In one implementation of this embodiment, the construction subunit includes:

[0202] The second word segmentation subunit is used to segment the target diagnosis and treatment process and the typical diagnosis and treatment process into words, respectively, to obtain the word segments contained in the target diagnosis and treatment process and the typical diagnosis and treatment process.

[0203] The second calculation subunit is used to calculate the word vectors corresponding to each segmented word contained in the target diagnosis and treatment process and the typical diagnosis and treatment process.

[0204] The third calculation subunit is used to perform pairwise similarity calculations between the word vectors corresponding to each segmented word in the target diagnosis and treatment process and the word vectors corresponding to each segmented word in the typical diagnosis and treatment process, so as to obtain the adjacency matrix of the heterogeneous graphs corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process.

[0205] In one implementation of this embodiment, the apparatus includes:

[0206] The third acquisition unit is used to acquire test medical records and to acquire test discharge diagnoses and test treatment processes from the test medical records.

[0207] The fourth acquisition unit is used to acquire the test typical treatment medical record corresponding to the test discharge diagnosis from the pre-built typical treatment medical record library, and to acquire the test typical diagnosis and treatment process from the test typical treatment medical record;

[0208] The third construction unit is used to construct a context graph corresponding to the typical diagnostic and treatment process of the test; and to construct a heterogeneous graph corresponding to the diagnostic and treatment process of the test and the typical diagnostic and treatment process of the test.

[0209] The identification unit is used to input the context graph corresponding to the typical diagnostic and treatment process and the heterogeneous graph corresponding to the diagnostic and treatment process and the typical diagnostic and treatment process into the graph network quality inspection model, and identify the test similarity between the diagnostic and treatment process and the typical diagnostic and treatment process.

[0210] The update unit is used to update the graph network quality inspection model by reusing the test medical records as training sample medical records when the test similarity does not exceed the preset threshold.

[0211] In one implementation of this embodiment, the graph network quality inspection model includes a first graph convolutional layer, a first new graph construction layer, a second graph convolutional layer, a second new graph construction layer, and a feature readout layer; the identification subunit includes:

[0212] The first convolutional subunit is used to input the context graph corresponding to the typical diagnosis and treatment process, the target diagnosis and treatment process and the heterogeneous graph corresponding to the typical diagnosis and treatment process into the first graph convolutional layer of the graph network quality inspection model for convolution processing to obtain the first initial feature vector after convolution processing.

[0213] The first extraction subunit is used to input the first initial feature vector into the first new graph construction layer to extract the features of core words and obtain the first feature vector.

[0214] The second convolutional subunit is used to input the first feature vector into the second graph convolutional layer for convolution processing to obtain the second initial feature vector after convolution processing.

[0215] The second extraction subunit is used to input the second initial feature vector into the second new graph construction layer, and perform feature extraction of core words again to obtain the second feature vector.

[0216] The dimensionality reduction subunit is used to concatenate the feature vectors corresponding to the heterogeneous graphs of the target diagnosis and treatment process and the typical diagnosis and treatment process, the first feature vector, and the second feature vector to obtain the concatenated feature vector, and input the concatenated feature vector into the feature readout layer for dimensionality reduction processing to obtain the dimensionality-reduced feature vector.

[0217] A sub-unit is defined to determine the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process based on the dimensionality-reduced feature vector.

[0218] Furthermore, this application embodiment also provides a quality inspection device for the consistency of discharge diagnosis and treatment process, including: a processor, a memory, and a system bus;

[0219] The processor and the memory are connected via the system bus;

[0220] The memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform any of the above-described implementations of the quality control method for consistency of discharge diagnosis and treatment processes.

[0221] Furthermore, embodiments of this application also provide a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any of the above-described methods for quality control of consistency between discharge diagnosis and treatment processes.

[0222] Furthermore, this application also provides a computer program product, which, when run on a terminal device, enables the terminal device to execute any of the above-described methods for quality inspection of the consistency of discharge diagnosis and treatment processes.

[0223] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0224] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0225] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0226] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A quality control method for ensuring consistency between discharge diagnosis and treatment process, characterized in that, include: Identify the target medical records to be inspected, and obtain the target discharge diagnosis and target treatment process from the target medical records; From a pre-built database of typical treatment medical records, obtain the typical treatment medical record corresponding to the target discharge diagnosis, and obtain the typical diagnosis and treatment process from the typical treatment medical record; Calculate the similarity between the target treatment process and the typical treatment process; Based on the similarity calculation results, the consistency between the target discharge diagnosis and the target treatment process is determined; The calculation of the similarity between the target treatment process and the typical treatment process includes: The typical diagnosis and treatment process is segmented into words to obtain the segmented words contained in the typical diagnosis and treatment process; The typical diagnostic and treatment process is traversed through each segmented word by a sliding window, and the dependency relationship between each segmented word is calculated using the mutual information PMI formula. Based on the dependencies between the segmented words, a context diagram corresponding to the typical diagnosis and treatment process is generated; The target treatment process and the typical treatment process are segmented into words to obtain the segmented words contained in each of the target treatment process and the typical treatment process. Calculate the word vectors corresponding to each segmented word contained in the target diagnosis and treatment process and the typical diagnosis and treatment process; The word vectors corresponding to each segmented word in the target diagnosis and treatment process are compared with the word vectors corresponding to each segmented word in the typical diagnosis and treatment process. The pairwise similarity is calculated to obtain the adjacency matrix of the heterogeneous graphs corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process. The context graph corresponding to the typical diagnosis and treatment process, the heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process are jointly input into a pre-constructed graph network quality inspection model to identify the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process; the graph network quality inspection model is obtained by comparative training using a loss function based on positive sample medical record data and negative sample medical record data.

2. The method according to claim 1, characterized in that, The graph network quality inspection model is constructed as follows: Construct training sample medical record data, which includes positive sample medical record data and negative sample medical record data; Construct a training context graph and a training heterogeneous graph corresponding to the diagnosis and treatment process in the medical record data of the training samples; The training context graph and training heterogeneous graph corresponding to the diagnosis and treatment process in the training sample medical record data are input into the initial graph network quality inspection model, and the graph network quality inspection model is trained by adjusting the loss function. The initial graph network quality inspection model includes a graph convolutional layer, a new graph construction layer, and a feature output layer.

3. The method according to claim 1 or 2, characterized in that, The loss function is a hinge loss function; the loss function is used to shorten the distance between positive sample medical record data representations and to widen the distance between negative sample medical record data representations.

4. The method according to claim 2, characterized in that, The method further includes: Obtain the test medical records, and extract the test discharge diagnosis and test treatment process from the test medical records; From the pre-built typical treatment medical record library, obtain the typical treatment medical record corresponding to the test discharge diagnosis, and obtain the typical diagnosis and treatment process from the typical treatment medical record; Construct a context graph corresponding to the typical diagnostic and treatment process of the test; and construct a heterogeneous graph corresponding to the diagnostic and treatment process of the test and the typical diagnostic and treatment process of the test. The context graph corresponding to the typical diagnostic and treatment process, the heterogeneous graph corresponding to the diagnostic and treatment process and the typical diagnostic and treatment process are input into the graph network quality inspection model to identify the test similarity between the diagnostic and treatment process and the typical diagnostic and treatment process. When the test similarity does not exceed the preset threshold, the test medical records are used again as training sample medical records to update the graph network quality inspection model.

5. The method according to claim 1, characterized in that, The graph network quality inspection model includes a first graph convolutional layer, a first new graph construction layer, a second graph convolutional layer, a second new graph construction layer, and a feature readout layer. The step of inputting the context graph corresponding to the typical treatment process, the heterogeneous graph corresponding to the target treatment process and the typical treatment process into the pre-constructed graph network quality inspection model to identify the similarity between the target treatment process and the typical treatment process includes: The context graph corresponding to the typical diagnosis and treatment process, the target diagnosis and treatment process, and the heterogeneous graph corresponding to the typical diagnosis and treatment process are jointly input into the first graph convolutional layer of the graph network quality inspection model for convolution processing to obtain the first initial feature vector after convolution processing. The first initial feature vector is input into the first new graph construction layer to extract the features of the core words and obtain the first feature vector. The first feature vector is input into the second graph convolutional layer for convolution processing to obtain the second initial feature vector after convolution processing. The second initial feature vector is input into the second new graph construction layer, and the core words are extracted again to obtain the second feature vector; The feature vectors corresponding to the heterogeneous graphs of the target diagnosis and treatment process and the typical diagnosis and treatment process, the first feature vector, and the second feature vector are concatenated to obtain a concatenated feature vector. The concatenated feature vector is then input into the feature readout layer for dimensionality reduction to obtain a dimensionality-reduced feature vector. Based on the reduced-dimensional feature vector, the similarity between the target treatment process and the typical treatment process is determined.

6. A quality control device for ensuring consistency between discharge diagnosis and treatment process, characterized in that, include: The first acquisition unit is used to determine the target medical record to be inspected and to acquire the target discharge diagnosis and target treatment process from the target medical record. The second acquisition unit is used to acquire typical treatment records corresponding to the target discharge diagnosis from a pre-built typical treatment record database, and to acquire typical diagnosis and treatment processes from the typical treatment records. A calculation unit is used to calculate the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process; The quality control unit is used to determine the consistency between the target discharge diagnosis and the target treatment process based on the similarity calculation results. The calculation unit includes: The first word segmentation subunit is used to perform word segmentation processing on the typical diagnosis and treatment process to obtain the various word segments contained in the typical diagnosis and treatment process; The first calculation subunit is used to traverse each segmented word contained in the typical diagnosis and treatment process through a sliding window, and to calculate the dependency relationship between each segmented word using the mutual information PMI formula. A subunit is generated to generate a context graph corresponding to the typical diagnosis and treatment process based on the dependency relationships between the segmented words. The second word segmentation subunit is used to segment the target diagnosis and treatment process and the typical diagnosis and treatment process into words, respectively, to obtain the word segments contained in the target diagnosis and treatment process and the typical diagnosis and treatment process. The second calculation subunit is used to calculate the word vectors corresponding to each segmented word contained in the target diagnosis and treatment process and the typical diagnosis and treatment process. The third calculation subunit is used to perform pairwise similarity calculation between the word vectors corresponding to each segmented word in the target diagnosis and treatment process and the word vectors corresponding to each segmented word in the typical diagnosis and treatment process, so as to obtain the adjacency matrix of the heterogeneous graphs corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process. The identification subunit is used to input the context graph corresponding to the typical diagnosis and treatment process, the heterogeneous graph corresponding to the target diagnosis and treatment process and the typical diagnosis and treatment process into a pre-constructed graph network quality inspection model to identify the similarity between the target diagnosis and treatment process and the typical diagnosis and treatment process; the graph network quality inspection model is obtained by comparative training using a loss function based on positive sample medical record data and negative sample medical record data.

7. A quality inspection device for ensuring consistency between discharge diagnosis and treatment process, characterized in that, include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, the one or more programs including instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the method described in any one of claims 1-5.

Images