Automatic construction system of standard diagnosis and treatment scheme of multiple disease types

Abstract

The application discloses a kind of standard diagnosis and treatment scheme automatic construction system of multiple typing of disease, it is characterized in that, comprising: data preprocessing module;ADRG grouping module;Core project automatic identification module;Regression tree grouping module.The present application technology combines clinical medical knowledge and machine learning method, identifies the standard treatment scheme of multiple typing under disease, forms disease knowledge base, to provide reference for the audit of fraud insurance case.The present application uses the medical knowledge of ADRG under the premise, uses some logical judgment to screen out the core project under disease to the medical data, uses regression tree to group important features.The grouping result and the clinical interpretability of the feature of grouping are strong, and the accuracy is also higher.In the process of grouping, both consider the project involved in diagnosis and treatment scheme, and can use model to realize efficient grouping.

Description

Technical Field

[0001] This invention relates to an automatic construction system for standard diagnosis and treatment plans for multiple disease subtypes, which is used to provide a reference for the review of fraud and insurance fraud cases. Background Technology

[0002] Currently, the review of medical records is mainly divided into manual review and intelligent review.

[0003] Manual review involves clinicians using their medical knowledge to manually determine whether a patient's visit is fraudulent. The accuracy of this review depends heavily on the clinicians' knowledge level, and therefore some omissions are possible. The manual review process is tedious and time-consuming, and requires a high level of knowledge from the clinicians.

[0004] Intelligent review relies on a pre-established knowledge base and rule base. Currently, knowledge base construction mainly involves manual extraction of various authoritative materials, which has limitations in efficiency and breadth. Automated knowledge base construction can be divided into knowledge graph construction technology based on book text corpora and knowledge mining construction based on medical big data. Both related technologies are still in the exploratory research stage.

[0005] In the field of data mining, methods such as directly conducting usage frequency statistics or using purely data-driven LDA topic modeling suffer from low interpretability and insufficient accuracy of the model results. Therefore, a necessary knowledge structure and composite data mining methods become essential.

[0006] From the perspective of clinical medical knowledge, common methods for grouping and summarizing medical record data include DRG / DIP grouping. DRG grouping is relatively coarse and only considers the cost similarity of samples, without considering the influence of other factors on the treatment plan, making it difficult to meet the requirements for intra-group treatment similarity. DIP grouping is too detailed, with little difference between each subgroup under the treatment plan, resulting in insufficient knowledge usability value and insufficient interpretability from the perspective of the clinical meaning of the treatment plan.

[0007] Therefore, a new algorithm model needs to be built, which combines widely recognized classification knowledge with a big data model for automatic and efficient grouping. Summary of the Invention

[0008] The purpose of this invention is to construct a new algorithm model for reviewing medical records, while combining widely recognized classification knowledge and a big data model for automatic and efficient grouping.

[0009] To achieve the above objectives, the technical solution of the present invention is to provide an automatic construction system for standard diagnosis and treatment plans for multiple disease subtypes, characterized in that it includes:

[0010] The data preprocessing module is used to acquire input medical record data and preprocess the input medical record data;

[0011] The ADRG packet module further includes MDC packet units and ADRG packet units, wherein:

[0012] The MDC grouping unit matches the disease code of the primary diagnosis in the medical record homepage with the corresponding MDC group table to obtain the MDC group to which each medical record data belongs.

[0013] The ADRG grouping unit further divides cases into three subcategories based on treatment method: "surgical," "non-surgical," and "operational." Cases with the same primary diagnosis and / or primary operation are then grouped into the ADRG group within each subcategory.

[0014] The core item automatic identification module is used to identify core items in the ADRG group, and further includes a pre-detection unit and a core item identification unit, wherein:

[0015] The pre-detection unit is used to determine whether a disease can be represented by a few items as the main treatment plan for the patient. That is, it performs a feasibility test on the core items representing the treatment plan. For diseases that can be represented by a few items as the main treatment plan for the patient, the core item identification unit is activated.

[0016] The core item identification unit is used to select the core items of the current disease as important grouping features based on the screening rules.

[0017] The regression tree grouping module uses the core items and medical record data features identified by the core item automatic identification module in the ADRG group as input features to generate a regression tree. Based on the generated regression tree, the regression tree grouping module makes the following judgments: the nodes of the regression tree grouping are used as the treatment plan under the current disease; based on the grouping results of the regression tree, the module calculates the features of other dimensions of the samples within the group to obtain the numerical range of different features under each treatment plan, providing a reference for judging whether newly included samples are suspected of fraud.

[0018] Preferably, the input medical record data includes visit information, medical order information, medical record cover page information, and ADRG information.

[0019] Preferably, the preprocessing of the input medical record data includes filtering out data whose age is an integer between 0 and 110, whose gender is male or female, whose medical institution level is level two or three, and whose medical category is inpatient; and: mapping the disease diagnosis in the original data to the ICD10 coding sub-category level, and based on the ICD10 hierarchical directory table, associating ICD10 sub-category, category, and chapter information as a supplementary information input model for disease diagnosis; standardizing and mapping the item names to the three major medical insurance catalogs, and removing items with negative amounts; and removing commonly used items.

[0020] Preferably, the commonly used items include routine admission examination items, items with low diagnostic specificity, and items with unclear meaning.

[0021] Preferably, in the MDC group table, 26 MDC groups are set up based on the main diagnosis on the medical record cover page and with reference to the main diagnostic categories of the National Medical Security Disease Diagnosis Related Groups.

[0022] Preferably, the implementation of the pre-detection unit includes the following steps:

[0023] Select all medical records within a fixed time period, filter the items in the prescription details of each visit, and then sort the items by cost from largest to smallest. Calculate the proportion of the total cost of the TOP_N items to the total cost of the visit, thus obtaining the cumulative item cost percentage for each visit under a certain disease. Set a threshold A for the cumulative item cost percentage and a threshold B for N in TOP_N. If the cumulative item cost percentage under a disease is greater than A and the N value is less than the threshold B, then the disease can be considered to be represented by certain items that constitute the main treatment plan for the visit.

[0024] By calculating the average cumulative cost percentage distribution corresponding to different TOP_N values ​​under a disease, it is found that if the cumulative cost percentage of the sample disease is greater than the set threshold and the N value is less than the set threshold, it indicates that the disease can be represented by the TOP_N item treatment plan.

[0025] Preferably, the filtering rules include:

[0026] The absolute value of the total project amount is greater than C yuan;

[0027] The project is specific: its usage rate in the current ADRG group is more than D% higher than the average usage rate of other ADRG groups.

[0028] The project is characterized by high frequency and high value: the combined ranking of the number of projects and the amount of money involved is within the top N, where N <= 21.

[0029] Preferably, in the regression tree grouping module, after generating the regression tree, the regression tree is first pruned, and then a judgment is made based on the pruned regression tree. The pruning principles include: after pruning, the CV within the group is less than a threshold; after pruning, the overall RIV decreases by the smallest amount.

[0030] This invention combines clinical medical knowledge and machine learning methods to identify standard treatment plans for multiple subtypes of diseases, forming a disease knowledge base, thereby providing a reference for the review of fraud and insurance abuse cases. Based on the medical knowledge of ADRG, this invention uses logical judgments to filter out core items under each disease from medical data and uses regression trees to group important features. This results in grouping results and features with strong clinical interpretability and high accuracy. The grouping process considers both the items involved in the treatment plan and utilizes models to achieve efficient grouping.

[0031] Compared with the prior art, the present invention has the following advantages: 1) It provides automatically generated multi-subtype diagnosis and treatment plans for various diseases; 2) The identification of core items is based on dynamic TOPN, which is more intelligent and reasonable; 3) The generation efficiency of diagnosis and treatment plans is high; 4) The clinical interpretability of diagnosis and treatment plans is high. Attached Figure Description

[0032] Figure 1 This illustrates the main grouping process of ADRG groups;

[0033] Figure 2 This illustrates Task 1, which involves grouping regression trees.

[0034] Figure 3 This illustrates Task 2, which involves grouping regression trees.

[0035] Figure 4 This illustrates the preliminary results of the regression tree generation using cataracts as an example.

[0036] Figure 5 This illustrates the sample data of the final results obtained using cataracts as an example. Detailed Implementation

[0037] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0038] This embodiment discloses an automatic construction system for standard diagnosis and treatment plans for multiple subtypes of diseases, which includes:

[0039] 1) Data Preprocessing Module

[0040] This is used to preprocess the input medical record data after it has been acquired.

[0041] In this embodiment, the input data includes:

[0042] (1) Medical information, including the name of the medical institution, the level of the medical institution, the time of admission / discharge, the diagnosis of the disease, the number of days of hospitalization, and the total medical expenses;

[0043] (2) Medical order information, which further includes the item category, item name, date of issuance, quantity used, unit price, and amount used;

[0044] (3) The medical record cover page information further includes the discharge primary diagnosis code, other diagnosis codes, surgery and operation codes, age, age under one year old, newborn birth weight, and gender;

[0045] (4) ADRG information, further including:

[0046] The CHS-DRG Major Diagnostic Categories (MDCs) include the primary diagnostic disease codes and their corresponding MDC codes.

[0047] "ADRG-Inclusion Criteria Configuration-Diagnosis": Includes diagnostic codes and corresponding ADRG groups;

[0048] "ADRG-Inclusion Criteria Configuration-Surgery": Includes the surgery code and the corresponding ADRG group.

[0049] In this embodiment, the data preprocessing module preprocesses the input medical record data by filtering out data that meets the following conditions and performing relevant processing on the data:

[0050] (1) Age: An integer between 0 and 110;

[0051] (2) Gender: Male, Female;

[0052] (3) Medical institution level: Level II and Level III;

[0053] (4) Medical Category: Inpatient;

[0054] (5) Disease diagnosis: Map the disease diagnosis in the raw data to the ICD10 coding sub-level, and based on the ICD10 level directory table, associate the ICD10 sub-category, category and chapter information as supplementary information input model for disease diagnosis;

[0055] (6) Project Name: Standardized mapping to the three major medical insurance catalogs, excluding projects with negative amounts;

[0056] (7) Commonly Used Items: Items that are frequently used (20%–35%) are excluded. In this embodiment:

[0057] Excluding routine admission examinations, such as complete blood count, urinalysis, liver function tests, and kidney function tests;

[0058] Excluding items with low diagnostic specificity (which are commonly used in all diagnoses);

[0059] Excluding items with unclear meanings, such as consultation fees, examination fees, and 5% reimbursement for Western medicine.

[0060] (ii) ADRG packet module, further including MDC packet unit and ADRG packet unit. Figure 1 This illustrates the main grouping process of ADRG groups.

[0061] Within the MDC grouping unit, based on the primary diagnosis on the medical record's front page and referencing the main diagnostic categories of the National Medical Insurance Disease Diagnosis Related Groups, there are 26 main diagnostic categories. Obtaining the MDC group to which each medical record data belongs through the MDC grouping unit involves the following steps:

[0062] a) Regional medical record data coding processing: Based on the national medical insurance version of "Medical Insurance Disease Diagnosis Classification and Code (ICD-10)" and "Medical Insurance Surgical Operation Classification and Code (ICD-9CM-3)", the diagnosis and operation of diseases in regional data are coded to achieve unified disease diagnosis coding and surgical operation coding in the region.

[0063] b) Regional medical record data filtering: Filter out medical records with missing diagnoses, and filter out institutions and their corresponding medical records that are not at level two or three.

[0064] c) MDC Pre-grouping: Before classifying MDC (Major Diagnostic Category), based on the data on the medical record front page, organ transplant recipients, those who have used a ventilator for more than 96 hours, those who are less than 29 days old, those whose primary diagnosis or other diagnosis is HIV, or those who have suffered severe trauma are pre-grouped to form MDCA, MDCP, MDCY, and MDCZ.

[0065] (4) MDC grouping: The standardized diagnostic code on the first page of the medical record is matched with the diagnostic code of the "National Medical Security Disease Diagnosis Related Groups (CHS-DRG) Subgrouping Scheme (Version 1.0)" (which includes MDC codes and disease diagnosis codes of the medical insurance version of ICD-10) to obtain the MDC group to which each medical record data belongs.

[0066] The ADRG grouping unit performs ADRG grouping of medical record data according to the detailed rules in each MDC category, specifically including the following:

[0067] Within each subdivided MDC group, participants are assigned to the ADRG group (Diagnosis Related Group) according to certain matching priorities and conditions. In this embodiment, the matching conditions and rules can be adjusted according to the specific needs of the regional medical insurance bureau. The ADRG grouping unit is ultimately divided into 376 ADRG groups, including 167 surgical groups, 22 non-surgical groups, and 187 internal medicine groups.

[0068] (iii) The core item automatic identification module further includes a pre-detection unit and a core item identification unit.

[0069] Pre-detection unit:

[0070] All medical records within a fixed time period are selected. The items in the prescription details for each visit are filtered, and then the items are sorted from highest to lowest cost. The proportion of the total cost of the top N items to the total cost of each visit is calculated, thus obtaining the cumulative cost percentage of the top 1, top 2, ..., top N items for each visit under a specific disease. Thresholds A for the cumulative cost percentage and B for N in the top N are set. If the cumulative cost percentage for a disease is greater than A and N is less than B, then the disease is considered to be represented by a few items that constitute the main treatment plan for the visit.

[0071] By calculating the average cumulative cost percentage distribution corresponding to different TOP_N values ​​under the disease (ICD3), it was found that if the cumulative item cost percentage of the sample disease is greater than the set threshold and the N value is less than the set threshold, it indicates that the treatment plan for this disease can be represented by the TOP_N items.

[0072] Core Project Identification Unit:

[0073] After the pre-detection unit obtains the TOP_N items for a disease, the core item identification unit further refines the core items for that disease based on screening rules. The main principles for setting the screening rules include that core items need to be specific, have high values ​​and high frequency, and exclude commonly used items. In this embodiment, the core item identification unit identifies core items by including the following steps:

[0074] a) Under the current ADRG group, filter the usage frequency percentage R1 and amount percentage R2 of items with a value greater than 500 yuan, and calculate 0.5*R1+0.5*R2 to obtain the importance index R of amount and frequency.

[0075] b) Under the current ADRG group, sort the items in step a) according to the importance index R of amount and frequency.

[0076] c) For each project, calculate the average value of the R1 index in step a) to obtain the average utilization rate AVG_R1 of the project across all ADRG groups; under the current ADRG group, calculate the projects whose utilization rate R1 is greater than 30% of the average utilization rate AVG_R1 and are ranked in the top 20 in step 2, and designate them as the core projects in the current ADRG group.

[0077] For example, the core items of the Cataract ADRG group are "intraocular lens implantation" and "intraocular lens". These two items have a usage rate in the Cataract ADRG group that is more than 30% higher than the average usage rate of items in other ADRG groups, and the importance index R of the amount and frequency of these two items are ranked 1 and 2 respectively.

[0078] (iv) Regression Tree Grouping Module. The regression tree grouping task includes: Task 1 – Selecting features for regression tree grouping; Task 2 – Selecting indicators for determining tree branches based on features, including the following:

[0079] 1) Determining the input features of the regression tree

[0080] Calculate the Pearson correlation coefficients of age, material cost ratio, surgical cost ratio, drug cost ratio, length of hospital stay, and total cost. Combine this with business understanding to determine the characteristics of the input regression tree.

[0081] The input metrics for the final model include: total amount of core items, number of core items, percentage of surgical costs, percentage of drug costs, age, length of hospital stay, and institution level.

[0082] Since the core items are obtained based on the dynamic TOP_N items for different diseases and a series of rule logics, the number of core items may vary for different diseases.

[0083] Taking cataracts as an example, the feature inputs for the regression tree are shown in the table below:

[0084] Indicator Chinese Name Medical institution level Age groups Hospital stay duration in segments Drug cost breakdown by segment Segmented by Surgical Fee Number of intraocular lens implantation surgeries Number of artificial lenses Cost of intraocular lens implantation The cost of artificial lenses

[0085] 2) Data cleaning for regression trees

[0086] For continuous independent variables, reasonable intervals can be divided from a clinical perspective. For example, for age, it can be divided into 0-17, 18-60, and 60 and above, which makes it easier for the regression tree to segment the data.

[0087] Splitting training and validation data: To verify the accuracy of the grouped model, the historical data needs to be randomly split in a 4:1 ratio, with 80% used as training data for the regression tree and 20% used as validation data for pruning the regression tree.

[0088] 3) Generation of regression tree

[0089] After obtaining the cleaned input features, the regression tree generation process begins. The specific steps are as follows:

[0090] a) For the cleaned training data, enumerate all branch nodes and split values. Each split will divide the data into two parts.

[0091] For example, the cost of a high-value project is greater than 1,000 yuan, and the cost of a high-value project is less than 1,000 yuan.

[0092] b) For all branching methods, calculate the corresponding RIV, and select the branching method with the highest RIV as the final branching method for this iteration. This branching method will minimize the overall cost variability of the data. The RIV calculation formula is as follows:

[0093] RIV=(SSQ-∑SSQ n ) / SSQ

[0094]

[0095] In the formula, SSQ represents the sum of squared deviations; SSQ n x represents the sum of squared deviations of the nth subgroup; i This represents the cost of the i-th item within a single subgroup; This represents the average cost of all items within a single subgroup.

[0096] c) Calculate the CV within the two subgroups in the branching method of step b). If the CV value within any subgroup after grouping is less than the average CV value before grouping, and the number of people in the group is greater than the threshold, then a subgroup can be formed; if the CV value within any subgroup after grouping is greater than the average CV before grouping, and the number of people in the group is greater than the threshold, then repeat steps a) and b) for that subgroup.

[0097] d) End: If the CV of all subgroups has decreased compared to the ungrouped state in the previous step, and the number of participants in each subgroup is greater than the threshold, or the subgroups cannot be further subdivided, then a regression tree model is generated.

[0098] Taking cataracts as an example, the preliminary results of the regression tree generation are as follows: Figure 4 As shown.

[0099] 4) Manually review and use test data to prune the regression tree.

[0100] Because regression trees are prone to overfitting, pruning is necessary. First, groups with a CV value greater than a threshold need to be manually reviewed to determine their eligibility to form their own groups.

[0101] Identify subgroups within a group that are not saturated with data, and determine whether they should be merged into a higher-level group or discarded.

[0102] The obtained regression tree is used to group the validation dataset, and the grouping effect of the model on the validation dataset is evaluated, thereby realizing bottom-up decision tree pruning. The pruning principles are as follows: after pruning, the CV within the group is less than the threshold; after pruning, the overall RIV decrease is minimized.

[0103] 5) Model Evaluation

[0104] A unique, random sample of 80% of the original data was drawn. The sampled data was then grouped using the regression tree model described above. The percentage of groups with RIV and CV greater than the threshold was then calculated.

[0105] At the same time, the node indicators of the group and the group values ​​corresponding to the node indicators are extracted, and the grouping logic is clinically judged to determine whether it belongs to a relatively typical diagnosis and treatment plan.

[0106] 6) Output Results

[0107] The final grouping results of the regression tree under the ADRG grouping rule were obtained. The node features and values ​​of the regression tree were extracted, processed, and the features with higher importance were combined to serve as the multi-subtype diagnostic and treatment plan within the ADRG group. Simultaneously, features from other dimensions within the subgroups were summarized to provide data reference for judging the fraud probability of newly included case samples in anti-fraud scenarios.

[0108] Taking cataracts as an example, the final sample data is as follows: Figure 5 As shown.

Claims

1. An automatic construction system for standard diagnosis and treatment plans for multiple subtypes of a disease, characterized in that, include: The data preprocessing module is used to acquire input medical record data and preprocess the input medical record data; The ADRG packet module further includes MDC packet units and ADRG packet units, wherein: The MDC grouping unit matches the disease code of the primary diagnosis in the medical record homepage with the corresponding MDC group table to obtain the MDC group to which each medical record data belongs. The ADRG grouping unit further divides cases into three subcategories based on treatment method: "surgical," "non-surgical," and "operational." Cases with the same primary diagnosis and / or primary operation are then grouped into the ADRG group within each subcategory. The core item automatic identification module is used to identify core items in the ADRG group, and further includes a pre-detection unit and a core item identification unit, wherein: The pre-detection unit is used to determine whether a disease can be represented by a few items as the main treatment plan for the patient. That is, it performs a feasibility test on the core items representing the treatment plan. For diseases that can be represented by a few items as the main treatment plan for the patient, the core item identification unit is activated. The implementation of the pre-detection unit includes the following steps: Select all medical records within a fixed time period, filter the items in the prescription details of each visit, and then sort the items by cost from largest to smallest. Calculate the proportion of the total cost of the TOP_N items to the total cost of the visit, thus obtaining the cumulative item cost percentage for each visit under a certain disease. Set a threshold A for the cumulative item cost percentage and a threshold B for N in TOP_N. If the cumulative item cost percentage under a disease is greater than A and the N value is less than the threshold B, then the disease can be considered to be represented by certain items that constitute the main treatment plan for the visit. By calculating the average cumulative cost percentage distribution corresponding to different TOP_N values ​​under a disease, it is found that if the cumulative item cost percentage of the sample disease is greater than the set threshold and the N value is less than the set threshold, it indicates that the disease can be represented by the TOP_N item treatment plan. The core item identification unit is used to select the core items of the current disease as important grouping features based on the screening rules. The filtering rules include: The absolute value of the total project amount is greater than C yuan; The project is specific: its usage rate in the current ADRG group is more than D% higher than the average usage rate of other ADRG groups. The project is characterized by high frequency and high value: the combined ranking of the number of projects and the amount of money involved is within the top N, where N <= 21; The regression tree grouping module uses the core items and medical record data features identified by the core item automatic identification module in the ADRG group as input features to generate a regression tree. Based on the generated regression tree, the regression tree grouping module makes the following judgments: the nodes of the regression tree grouping are used as the treatment plan under the current disease; based on the grouping results of the regression tree, the module calculates the features of other dimensions of the samples within the group to obtain the numerical range of different features under each treatment plan, providing a reference for judging whether newly included samples are suspected of fraud.

2. The automatic construction system for standard diagnosis and treatment plans for multiple subtypes of diseases as described in claim 1, characterized in that, The input medical record data includes visit information, doctor's orders, medical record cover page information, and ADRG information.

3. The automatic construction system for standard diagnosis and treatment plans for multiple subtypes of diseases as described in claim 1, characterized in that, The preprocessing of the input medical record data includes filtering out data with ages between 0 and 110, genders of male or female, medical institution levels of level two or three, and medical categories of inpatient. It also involves: mapping disease diagnoses in the original data to ICD10 coding sub-levels; and, based on the ICD10 hierarchical directory table, associating ICD10 sub-categories, categories, and chapter information as supplementary information input models for disease diagnoses; standardizing and mapping item names to the three major medical insurance catalogs, removing items with negative amounts; and removing commonly used items.

4. The automatic construction system for standard diagnosis and treatment plans for multiple subtypes of diseases as described in claim 3, characterized in that, The commonly used items include routine admission examinations, items with low diagnostic specificity, and items with unclear meaning.

5. The automatic construction system for standard diagnosis and treatment plans for multiple subtypes of diseases as described in claim 1, characterized in that, In the MDC group table, based on the main diagnosis on the medical record cover page and referring to the main diagnostic categories of the National Medical Security Disease Diagnosis Related Groups, 26 MDC groups are set up.

6. The automatic construction system for standard diagnosis and treatment plans for multiple subtypes of diseases as described in claim 1, characterized in that, In the regression tree grouping module, after the regression tree is generated, the regression tree is pruned first, and then a judgment is made based on the pruned regression tree. The pruning principles include: after pruning, the CV within the group is less than the threshold; after pruning, the overall RIV decreases by the smallest amount.

Images