Automatic input method, system and device of medical ticket information and storage medium

Abstract

This invention provides a method, system, device, and storage medium for automatically entering medical invoice information, belonging to the field of financial and insurance data processing technology. The method includes: performing optical character recognition (OCR) on a medical invoice image to obtain a set of invoice characters; inputting the invoice character set into a machine learning model to determine invoice category labels and form a set of invoice category labels, wherein the machine learning model is used to determine the invoice category label corresponding to each invoice character in the invoice character set; based on the invoice category label set, determining a subset of invoice characters corresponding to each invoice category label from the invoice character set to constitute structured data; and storing the structured data in an invoice information database so that when the invoice information database is queried, it returns medical invoice information represented by the structured data. This invention can be used to provide medical invoice information.

Description

Technical Field

[0001] This invention relates to the field of financial and insurance data processing technology, specifically to an automatic entry method for medical invoice information, an automatic entry system for medical invoice information, an electronic device, and a machine-readable storage medium. Background Technology

[0002] Insurance companies and medical institutions typically do not share information systems. In insurance claims, insurance company personnel usually require claimants to provide medical invoices issued by medical institutions. The personnel then manually enter this information into the insurance company's system. However, because medical invoices contain a large amount of drug information and medical terminology, manual data entry is tedious, time-consuming, and inconsistent in accuracy. Furthermore, since claims calculation and case review monitoring rely on the entered medical invoice information, claimants often receive their compensation quickly, leading to lengthy claims settlement times. As business volume increases, labor costs also rise significantly.

[0003] Currently, rules for organizing characters are developed by experts, and Optical Character Recognition (OCR) technology is used to reduce manual intervention. However, OCR technology has significant differences in its ability to recognize medical invoices of various styles and from different regions. Furthermore, after errors or omissions occur during recognition, it becomes difficult to automatically organize the medical invoice information using established rules (such as keyword rules). The invoice will then require manual entry and adjustment, making it difficult for insurance companies' information systems to efficiently obtain medical invoice information. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, device, and storage medium for automatically entering medical invoice information. This avoids the difficulty in automatically entering medical invoice information caused by the difference in the recognition capabilities of optical character recognition technology for medical invoice images of various styles and from different regions. In this way, it realizes structured data entry of medical invoice information for common types, has system robustness, and can avoid the impact of erroneous characters and / or missing characters on information entry.

[0005] To achieve the above objectives, this specification adopts the following approach:

[0006] In a first aspect, embodiments of the present invention provide a method for automatically entering medical invoice information, the method comprising:

[0007] Optical character recognition is performed on medical invoice images to obtain the invoice character set;

[0008] The set of bill characters is input into a machine model to determine the bill category label and form a set of bill category labels. The machine model is used to determine the bill category label corresponding to each bill character in the set of bill characters.

[0009] Based on the set of bill category labels, a subset of bill characters corresponding to each bill category label is determined from the set of bill characters to form structured data;

[0010] The structured data is stored in a bill information database so that when the bill information database is queried, it returns medical bill information based on the structured data representation.

[0011] In a second aspect, embodiments of the present invention provide an automatic data entry system for medical invoice information, the automatic data entry system comprising:

[0012] The recognition module is used to perform optical character recognition on medical invoice images to obtain a set of invoice characters;

[0013] The classification module is used to input the set of bill characters into the machine model to determine the bill category label and form a set of bill category labels. The machine model is used to determine the bill category label corresponding to each bill character in the set of bill characters.

[0014] The structuring module is used to determine a subset of bill characters corresponding to each bill category label from the bill character set based on the bill category label set, so as to form structured data;

[0015] A storage module is used to store the structured data in a bill information database, so that when the bill information database is queried, it returns medical bill information based on the structured data representation.

[0016] Thirdly, embodiments of the present invention provide an electronic device, the electronic device comprising:

[0017] At least one processor;

[0018] A memory connected to the at least one processor;

[0019] The memory stores instructions that can be executed by the at least one processor, and the at least one processor implements the aforementioned method by executing the instructions stored in the memory.

[0020] Fourthly, embodiments of the present invention provide a machine-readable storage medium storing machine instructions that, when executed on a machine, cause the machine to perform the aforementioned method.

[0021] In this invention, after recognizing the characters on a medical invoice image, a set of invoice characters is obtained. These characters are arranged in the order of OCR recognition. Each character in the set is input into the machine learning model to obtain a corresponding invoice category label, forming a set of invoice category labels. This set is a label set for characters, not for the entire invoice image. It can identify multiple invoice category labels for the same medical invoice image, rather than simply classifying the invoice. Then, the invoice characters are systematically divided and mapped using the invoice category label set, without relying on the characteristics of the recognized characters themselves. This classification and mapping allows for the creation of a subset of invoice characters, forming structured data. It eliminates the need to compare recognized characters with keywords, avoiding errors in character recognition and preventing omissions that could negatively impact medical invoice information entry. Finally, the structured data is written into an invoice information database, enabling automated medical invoice entry. Furthermore, the structured data in the database is searchable, particularly the amounts on medical invoices, which facilitates amount calculation and monitoring, improving the efficiency of compensation calculation and case review monitoring.

[0022] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used in conjunction with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0024] Figure 1 This is a schematic diagram of the main method steps in an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram illustrating the input-output relationship of an exemplary machine model according to an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of an exemplary application server scenario according to an embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram of a module of an exemplary electronic device according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this specification clearer, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the present invention.

[0029] As mentioned earlier, when manually entering medical invoice information, it is often necessary to fill in corresponding medical invoice information forms for each invoice category. For example, if a medical invoice image contains the information "examination fee 118.00 yuan," then the invoice category "examination fee" (which is the category of the medical procedure) and the amount of 118.00 yuan should be entered in the form. When faced with a large amount of information to be entered, the data entry is very time-consuming. Errors are prone to occur in entering monetary (numerical) information, and the correspondence between monetary amounts and invoice categories is also prone to errors. While OCR technology and expert-defined character processing rules can match keyword rules when OCR recognition is accurate, the OCR recognition results inevitably contain erroneous or missing characters when dealing with a large number of invoices of various styles and from different regions. Manual input and adjustments are still required, making it difficult to apply to the automatic data entry of medical invoice information in insurance claims scenarios.

[0030] Therefore, this manual provides a solution for the automatic entry of medical invoice information. It enables the structured data entry of common types of medical invoices, including those of various styles and from different regions. This eliminates the need for character matching rules such as keyword rules and regular expressions corresponding to invoice image classification, and also eliminates the need to retrain the OCR recognition model. In the entry process described in this manual, optical character recognition (OCR) is performed on the medical invoice image to obtain a set of invoice characters. The accuracy of character recognition is more stable than manual recognition, especially for Arabic numerals, such as monetary amounts. Then, a machine learning model is used to classify the invoice characters, obtaining a category label and a set of labels for each character. This allows for the determination of the correspondence between invoice characters and category labels. The corresponding invoice characters are then identified based on the category labels, thereby finding key character groups (subsets) based on the correspondence to form structured data. Subsequently, the structured data is stored in an invoice information database to complete the automated entry, providing medical invoice information based on structured data representation when the database is queried.

[0031] In a first aspect, embodiments of the present invention provide an automatic method for entering medical invoice information, which can be applied to a server. This server can communicate with user devices to exchange data through a configured interface. The server in these embodiments can be one or more servers, which can be physical servers, server instances, or container instances with communication capabilities. The hardware of the instance can be a resource instance in a server cluster, composed of processor and memory resources, possessing computing, instruction processing, and communication functions; such as a cloud server. The user device can be the user device of business personnel and / or users applying for claims. The user device can include mobile electronic devices and computers, etc., devices with computing, instruction processing, and communication capabilities. The server in these embodiments can be configured with scripts, programs, and / or model files, such as model files for machine models used for label classification, OCR recognition services, etc., and can also be configured with a database or communicate with a database to store structured data and / or query medical invoice information.

[0032] In the embodiments of the present invention, please refer to Figure 1 The aforementioned methods for automatically entering medical invoice information may include:

[0033] S1) The server can perform optical character recognition on medical invoice images to obtain a set of invoice characters.

[0034] In some possible implementations, the server can receive medical invoice images uploaded by user devices. These images may correspond to a user identifier, which could include the identity of the claimant and / or the identity of the agent entrusted to upload the image. The medical invoice image may include scanned images and / or photographs of the medical invoice, which can be data from an electronic document or separate image format data. The server can obtain the invoice characters from the medical invoice image using an OCR recognition service. The aforementioned step S1) may include:

[0035] S101) Arrange the boxes to be recognized on the medical invoice image in the recognition order;

[0036] S102) Determine the ticket characters within the frame to be identified;

[0037] S103) According to the identification order, the determined ticket characters are formed into a ticket character set.

[0038] In some possible implementation manners, the OCR recognition service may have a recognition order. For example, relative to the reference point / origin in the medical bill image, the recognition order may be from top to bottom in the vertical axis direction and from left to right in the horizontal axis direction. The recognition frames to be recognized may be arranged on the medical bill image according to the recognition order. The OCR recognition service may convert the image blocks in the recognition frames into characters (including words, symbols, etc.) in the computer by detecting the dark and bright patterns of the image blocks in the recognition frames and using character recognition methods. For example, the OCR recognition service may recognize the image blocks in the recognition frames to be recognized and determine the corresponding bill characters. The OCR recognition service or the server may combine the determined bill characters into a bill character set according to the recognition order. The bill character set may be composed of a combination of correct characters, a combination of correct characters and incorrect characters (caused by unclear images, stains, mutilation, etc.), and / or a combination of missing semantic characters (that is, there are missing correct characters between the correct characters in the recognition result and there may be mixed incorrect characters; caused by unclear images, stains, mutilation, etc.).

[0039] For example, the bill character set composed of a combination of correct characters (correct bill character set) may include "Examination fee: 118.00 yuan", "Western medicine fee: 999.00 yuan", "Chinese herbal medicine fee: 60.00 yuan", "Xiao'er Chaigui: 1 box of 12 bags, 25 yuan", "Examination fee: 118.00 yuan, Laboratory fee: 443.00 yuan", etc.; the bill character set composed of a combination of correct characters and incorrect characters may include "Inspection fee: 118.00 yuan" (incorrect character: 杳), "Western medicine fee: 999.00 yuan" (incorrect characters: 酉 and 弗), "Xiao'er Chaigui: 1 box of 12 bags, 25 yuan" (incorrect characters: 枉 and 贷), etc.; the bill character set composed of a combination of missing semantic characters may include "Inspection fee: 118.00 yuan" (missing character: 查), "Xiao'er 丨乚 Chaigui: 112 bags, 25 yuan" (missing character: 盒, incorrect characters: 丨 and 乚), "Chinese herbal medicine fee: 60.00 yuan" (missing character: 中), etc. The bill character sets composed of a combination of correct characters and incorrect characters and composed of a combination of missing semantic characters may be referred to as incorrect and missing bill character sets. It can be observed that the relative position sequence between the correct characters (i.e., between the overlapping characters) in the incorrect and missing bill character sets is the same as that in the correct bill character set. Therefore, a machine model can be adopted, and this machine model can utilize the position information between the upper and lower characters of the bill characters in the bill character set to obtain the bill category labels corresponding to the bill characters at each input position.

[0040] In the embodiment of the present invention, the foregoing automatic entry method for medical bill information may further include:

[0041] S2) The server may input the bill character set into the machine model to determine the bill category labels and form a bill category label set, and the machine model is used to determine the bill category labels corresponding to the bill characters in the bill character set.

[0042] In some possible implementations, the server can be configured with a trained machine model, which can be applied in step S2). The trained machine model can include a trained BERT model and a single-layer neural network model, such as a BERT model and a linear model (BERT+linear model). The linear model can be a multi-class neural network model. The foregoing step S2) can include:

[0043] S201) Input the set of bill characters into the trained BERT model to extract embedding vectors from the BERT model;

[0044] S202) Input the embedding vectors into the trained single-layer neural network model to determine the bill category labels corresponding to each bill character that is the input of the BERT model.

[0045] In some possible implementations, the BERT model can be a pre-trained model trained by fine-tuning with training data. It can convert the bill characters in the set of bill characters, which are used as inputs, into embedding vectors in the semantic space. The single-layer neural network model can convert the embedding vectors in the semantic space into multi-dimensional vectors in the label space. The predicted values under each dimension in the multi-dimensional vectors can respectively represent the bill category labels, so that the bill category labels corresponding to each bill character in the set of bill characters can be obtained and a set of bill category labels can be formed, rather than the classification labels of medical bill images.

[0046] In the first example, please refer to Figure 2 , the determined bill category labels can include an item name label A and an amount label B (and a quantity label C). Input the "examination fee 118.00 yuan" recognized by the OCR recognition service into the machine model. For example, in the multi-dimensional vector corresponding to the character "examination", the predicted value corresponding to the item name label A will be greater than the predicted value corresponding to the amount label B, and the bill category label of the character "examination" is determined to be the item name label A; in the multi-dimensional vector corresponding to the character "8", the predicted value corresponding to the amount label B will be greater than the predicted value corresponding to the item name label A, and the bill category label of the character "8" is determined to be the amount label B. Determined in this way, the set of bill category labels output by the machine model can include 3 item name labels A and 7 amount labels B, that is, AAABBBBBBB (including the amount label corresponding to the decimal point. At this time, no character belongs to the category of the quantity label C), and the recognition order of the bill characters can be recorded accordingly (or the corresponding relationship can be recorded separately using index values / numbers) to record the corresponding relationship between each bill character and the bill category label, so as to obtain the bill category labels between the bill characters, rather than the category labels of medical bill images.

[0047] In the second example, the determined bill category labels may include an item name label A, an amount label B, and a (drug) quantity label C. The "Xiaol Chai Gui 112 bags 25 yuan" recognized by the OCR recognition service is input into the machine model. For example, among the multi-dimensional vectors corresponding to the character "乚", the predicted value corresponding to the item name label A is the maximum value in the vector. The bill category label of the character "乚" is determined to be the item name label A. Among the multi-dimensional vectors corresponding to the character "2" before the character "袋", the predicted value corresponding to the quantity label C is the maximum value in the vector. The bill category label of the character "2" is determined to be the quantity label C. Determined in this way, the set of bill category labels output by the machine model may include 5 item name labels A, 4 quantity labels C, and 3 amount labels B, that is, AAAAACCCCBBB. At the same time, the corresponding relationship between each bill character and the bill category label is recorded, so as to obtain the bill category labels between the bill characters, rather than the category labels of the medical bill image. Although there are character omissions and errors in the OCR recognition service recognition result, there are correct characters in the recognition result and the relative position order between the correct characters in the recognition result is the same as that of the recognition result containing only correct characters, which will not affect the bill category labels determined by the machine model.

[0048] In the embodiment of the present invention, it should be noted that, in order to implement the machine model for the foregoing classification function, a BERT model pre-trained in Chinese and a single-layer neural network model can be used for training.

[0049] In some possible implementation manners, in terms of training, the machine model includes a (Chinese) pre-trained BERT model and a single-layer neural network model; the single-layer neural network model is used to convert the embedding vector extracted from the BERT model into a predicted output vector in the bill category label space; any predicted value in the predicted output vector corresponds to a bill category label; the bill category label corresponding to the maximum predicted value in the predicted output vector is used as the bill category label corresponding to the bill character input into the machine model. For example, it can be defined that the bill category labels include labels such as invoice title, invoice code, item name, drug quantity, medical unit, amount, or inspection category name, etc. At this time, any bill category label determined during use includes: invoice title label, invoice code label, item name label, drug quantity label, medical unit label, amount label, or inspection category name label, etc. The training method of the foregoing machine model may include:

[0050] T1) The server can respond to user operations or user device instructions by marking the bill category labels of bill characters in ordinary medical bill information and abnormal medical bill information respectively to obtain training data, the training data including the marked bill characters and bill category labels;

[0051] The T2 server can train the BERT model and the single-layer neural network model based on the training data.

[0052] In some possible implementations, the server may be configured with training scripts and / or programs (and dependent environments) to respond to user operations or user device instructions, label bill category tags for bill characters in medical bill information, and load training data to train a machine learning model. Ordinary medical bill information may consist of a set of bill characters containing only correct characters in semantic order, while abnormal medical bill information may consist of a set of incomplete and / or erroneous bill characters. Ordinary medical bill information, belonging to a subset of bill characters with the same bill category tag, may include correct characters and their semantic order from the National Drug Database, Medical Insurance Item Table, and / or custom annotation data (e.g., drugs and medical insurance items covered by insurance policies). Abnormal medical bill information can be identified by comparing it to the National Drug Database, Medical Insurance Item Table, and / or custom annotation data to obtain semantically complete and / or correctly labeled corresponding ordinary medical bill information. Therefore, the bill category tags of the bill characters in this corresponding ordinary medical bill information can be used to mark the abnormal medical bill information.

[0053] In some possible examples, the aforementioned step T1) may include:

[0054] T101) A bill category label that marks the bill characters in each (collected) ordinary medical bill information relative to the national drug database, medical insurance item list and / or custom annotation data.

[0055] For example, "inspection fee of 118.00 yuan" is sequentially labeled with 3 item name tags (the same item name tag A) and 7 amount tags (the same amount tag B). The labeled data is then [inspection-A, inspection-A, fee-A, 1-B, 1-B, 8-B, .-B, 0-B, 0-B, yuan-B].

[0056] T102) Mark the first subset of bill characters with incomplete semantics or incorrect characters in the abnormal medical bill information as the same bill category label; the same bill category label is the bill category label of the second subset of bill characters in the normal medical bill information corresponding to the abnormal medical bill information, where the difference characters between the second subset of bill characters and the first subset of bill characters, after being placed into the first subset of bill characters in the order of the bill characters in the second subset of bill characters, the semantics of the first subset of bill characters are complete; the difference characters between the second subset of bill characters and the first subset of bill characters, after being replaced into the first subset of bill characters in the order of the bill characters in the second subset of bill characters, there are no incorrect characters in the first subset of bill characters.

[0057] For example, the normal medical bill information corresponding to "Examination fee 118.00 yuan" (difference character: 查) is "Check fee 118.00 yuan". Therefore, mark "检", "杳", "费" as the same bill category label (item name label), and mark "Examination fee 118.00 yuan" as 3 item name labels and 7 amount labels in sequence; another example, the normal medical bill information corresponding to "Small 丨乚 Cinnabar Twig 112 bags 25 yuan" is "Pediatric Cinnabar Twig 1 box 12 bags 25 yuan". Therefore, mark "Small 丨乚 Cinnabar Twig 112 bags 25 yuan" (difference characters: 儿, and the missing "盒") as 5 item name labels A, 4 quantity labels C and 3 amount labels B in sequence. At this time, the marked data is [小 - A, 丨 - A, 乚 - A, 柴 - A, 桂 - A, 1 - C, 1 - C, 2 - C, 袋 - C, 2 - B, 5 - B, 元 - B], and each bill character has a corresponding bill category label. It can be understood that the foregoing marking operation can be executed in response to the foregoing instructions, and the foregoing instructions can include click instructions, input text instructions, swipe / select instructions, etc.

[0058] In some possible examples, the foregoing normal medical bill information (national drug library catalog, medical insurance medical item table, and self - marked data) can be used as the main training corpus data. The normal medical bill information belonging to the same normal medical bill can be used as a sample. To provide robustness against misspelled and missing characters, abnormal medical bill information (for example, selected from OCR recognition results, system tests, and usage history records) can be added to the training corpus data to obtain training data, which can perform data augmentation operations such as misspelled characters, missing characters, and label ratio alignment (by placing specified symbols between the bill characters used as input to align character position markers and / or mark the starting position characters) for the actual text and character distribution, rather than using the abnormal medical bill information as a negative sample for medical bill image classification. The foregoing step T2) can include:

[0059] T201) can use training data as input to a pre-trained language model BERT, and input the BERT output into a single-layer neural network to transform the semantic space vector into a label space. By comparing the predicted values ​​in the label space, the predicted label (i.e., the ticket category label) corresponding to each input ticket character is obtained. Cross-entropy loss is calculated between this label and the actual label corresponding to the character. Training stops when the loss value no longer decreases. The loss value can be calculated using the following formula:

[0060]

[0061] In this formula, N is the total number of samples, M is the number of ticket category labels, and L i It is the loss value of the i-th (positive integer and less than or equal to N) sample, y ic p is the true value when the current ticket character in the i-th sample belongs to the c-th ticket category label category. ic Let be the predicted value when the current bill character in the i-th sample belongs to the c-th bill category label category. Each label can be encoded as a label vector, which can be a multi-dimensional vector. For example, if there are 3 bill category labels, namely item name label A, amount label B, and quantity label C, the label vector can include [1,0,0] for item name label A, [0,1,0] for amount label B, and [0,0,1] for quantity label C.

[0062] In this embodiment of the invention, after training, the set of ticket characters determined by the OCR recognition service can be used as input to the trained machine model. In use, the subset of ticket characters corresponding to each ticket category label in the set of ticket category labels output by the machine model may include a first subset of ticket characters containing semantically incomplete or erroneous characters, or a second subset of ticket characters containing semantically complete or erroneous characters. Specifically, the ticket characters that distinguish the second subset from the first subset, after being placed into the first subset according to the ticket character order in the second subset, ensure the semantic completeness of the first subset; and the ticket characters that distinguish the second subset from the first subset, after being replaced into the first subset according to the ticket character order in the second subset, ensure that the first subset does not contain erroneous characters.

[0063] In this embodiment of the invention, the aforementioned automated data entry operation can continue, and the aforementioned method for automatically entering medical invoice information may further include:

[0064] S3) Based on the set of bill category labels, determine the subset of bill characters corresponding to each bill category label from the set of bill characters to form structured data.

[0065] In some possible implementations, the specific data structure and data encapsulation format of the structured data can be selected and set based on the invoice information database and the testing and usage effects. For example, key-value non-relational structured data and / or table-based relational structured data can be used, but there is no need to use the character characteristics of the invoice characters themselves.

[0066] In some possible examples, the aforementioned step S3) may include:

[0067] S301) Use the bill category labels in the bill category label set as fields of the specified data structure;

[0068] S302) By using the ticket characters corresponding to each ticket category label, the ticket characters are used as the values ​​corresponding to the field in the specified data structure. This value can be all ticket characters in the ticket character subset.

[0069] S303) Using the fields and corresponding values ​​in the specified data structure, structured data is constructed according to the specified data format.

[0070] The data structure can include data nodes with fields and values, as well as the number of the data node. In some possible applications, the bill character set "inspection fee 118.00 yuan" can be divided into two subsets of bill characters by the bill category label: one subset "inspection fee" and another subset "118.00 yuan". In some other possible applications, each time a label corresponding to a name appears, a data node can be initialized, which can be divided into different fee items. For example, "Examination fee 118.00 yuan, medicine fee 25 yuan", the corresponding invoice category label set consists of item name label A and amount label B, i.e., AAABBBBBBBAAABBB. In this case, after the first A appears, a data node is initialized to record the label "AAABBBBBBB" and the corresponding invoice character "Examination fee 118.00 yuan". The invoice characters "examination fee" and "118.00 yuan" are two subsets of the invoice characters in the first data node. The first A after the seventh B is the second appearance of item name label A, so a new data node is created to record the label "AAABBB" and the corresponding invoice character "medicine fee 25 yuan". The invoice characters "medicine fee" and "25 yuan" are two subsets of the invoice characters in the second data node.

[0071] In some possible applications, for the first aforementioned example, the fields are the item (specifically a medical item) name label A and the amount label B. The value corresponding to field A is "examination fee", and the value corresponding to field B is "118.00 yuan", which can be used as the data node numbered 1. If the set of ticket characters determined by the OCR recognition service for the corresponding medical bill image in the first example also includes "drug fee 25 yuan", and the machine model determines that the corresponding set of ticket category labels also includes 3 item name labels A and 3 amount labels B. The value corresponding to field A is "drug fee", and the value corresponding to field B is "25 yuan", which can be used as the data node numbered 2.

[0072] In some possible applications, for the second aforementioned example, the medical bill image in this example can be different from that in the first example. For example, the region and the type of the bill are different. At this time, the fields are the item (specifically a drug) name label A, the amount label B, and the quantity label C. The value corresponding to field A is "Xiaoyao Chaihu", the value corresponding to field C is "112 bags", and the value corresponding to field B is "25 yuan", which can be used as the data node numbered 3. In addition, the data structure can also include a user identification field corresponding to the data node and a timestamp (the current moment or the moment when the image is uploaded, year-month-day-hour-minute) field. The user identification can include the identity identification of the user applying for claim settlement / authorized business personnel and / or the claim settlement case identification, etc. The specified data format can include text format, array, table, etc., which can be used to determine the format of the combined data nodes or the format between different fields in each data node. The text format can include Json text format, number table format, etc. For example, when using the Json text format, the structured data can include:

[0073] {"Item Name Label A": "Examination Fee"; "Amount Label B": "118.00 Yuan"; "Number": "1"; "User Identification": "1"; "Timestamp": "202203101322"};

[0074] {"Item Name Label A": "Drug Fee"; "Amount Label B": "25 Yuan"; "Number": "2"; "User Identification": "1"; "Timestamp": "202206161551"};

[0075] {"Item Name Label A": "Xiaoyao Chaihu"; "Quantity Label C": "112 Bags"; "Amount Label B": "25 Yuan"; "Number": "3"; "User Identification": "2"; "Timestamp": "202305100901"}.

[0076] It is worth noting that although the structured data mentioned above may contain data nodes with incomplete semantics and / or erroneous characters, the division and corresponding records based on the ticket category labels between each ticket character, without using the rule-matching characteristics of the ticket characters themselves determined by the OCR recognition service, will not cause further interruption of database entry.

[0077] In this embodiment of the invention, the aforementioned method for automatically entering medical bill information may further include:

[0078] S4) The server can store the structured data in the invoice information database so that when the invoice information database is queried, it returns medical invoice information based on the structured data representation.

[0079] In some possible implementations, the server can save structured data to a designated invoice information database. Responding to query commands sent by user devices, the server can return structured data presented visually (via forms and style sheets) as medical invoice information. This automates the input of characters from medical invoice images, eliminating the need for manual intervention. Input speed is rapid and costs are significantly reduced. Even if erroneous and / or missing characters exist in the structured data, users will not need to re-upload medical images due to characters not conforming to keyword or character matching rules. For cases with erroneous and / or missing characters in the structured data, claims settlement and case review monitoring can still produce normal results because the correct structured characters in the data, along with the erroneous or missing characters corresponding to the field to which the correct characters belong, combined with the relative positional relationships between the correct characters, still represent the key semantics.

[0080] For example, in data node 3, "Project Name Tag A: "Small | Chai Gui", "Quantity Tag C: "112 bags", "Amount Tag B: "25 yuan"", the business personnel performing the review or settlement can determine that it is a (Chai Gui) type of drug name and the reimbursed drug cost is 25 yuan. At this time, the invoice characters corresponding to the amount tag and the invoice characters corresponding to the quantity tag are divided into different subsets of invoice characters, which is sufficient to determine the settlement / review result (mainly determined by the correlation between the project name and identity identifier, the claim case identifier, the project name, quantity, and amount, etc.), which reduces the settlement time of claim cases, improves the efficiency of the user's claim process, and improves the claim service efficiency of insurance companies. In some applications that are conducive to settlement / review, the access address field of medical invoice images can be recorded in the structured data. After receiving the medical invoice image, the server can generate and store the access address of the medical invoice image as the value of the access address field, so that the business personnel performing the settlement and review can view it without the need for the user to re-upload the image.

[0081] In this embodiment of the invention, automated settlement and / or auditing monitoring can also be performed. The aforementioned server can be used as a monitoring server, and the monitoring method of the monitoring server can include:

[0082] M1) sends a query instruction to the invoice information database, the query instruction carrying user identification and query time range information;

[0083] M2) Based on the medical invoice information returned by the invoice information database, calculate the total amount corresponding to the user identifier within the query time interval, and when the total amount is equal to or lower than a specified threshold, generate a notification message based on the user identifier and the total amount, and send the notification message to the business personnel's device to provide medical invoice information, which is beneficial for business personnel performing settlement business to complete the settlement of compensation.

[0084] M3) When the total amount exceeds a specified threshold, generate a list of drug information corresponding to the user identifier. The specified threshold can be selected based on the average of claim cases, the average of reimbursement item names, and / or the reference value indicated by the region to which the invoice belongs.

[0085] M4) Based on the user identifier, the total amount, and the drug information list, a notification message is generated and sent to the business personnel's device to provide medical invoice information. This is beneficial for business personnel performing audit and monitoring to make anti-fraud judgments. For example, if the total amount deviates significantly from a specified threshold (the difference is 10 times, 100 times, etc. of the specified threshold), a specified identifier can be configured in the notification message to remind business personnel performing audit and monitoring to make anti-fraud judgments.

[0086] In an exemplary scenario of communication between an application server and a user equipment disclosed in an embodiment of the present invention, please refer to... Figure 3Users applying for claims can take pictures of medical receipts using user device 1 and upload the images to the server in the aforementioned embodiment via the communication network. The server can determine the set of receipt characters using OCR recognition services, and then determine the corresponding receipt category tags for each receipt character using a machine learning model, forming a set of receipt category tags. The server then divides the receipt character set into subsets based on the receipt category tags, and uses the correspondence between each receipt character and its corresponding receipt category tag to construct structured data. The server stores the structured data in the database. The server can periodically / timely query the receipt information database and calculate the total amount within the query time interval for each user identifier. If the total amount exceeds a specified threshold, a list of drug information corresponding to the user identifier is generated. Then, based on the user identifier, the total amount, and the list of drug information (which can be used for comparison with a reference list of drug information), a notification message is generated and sent to user device 2 of business personnel 2 via the communication network. User device 2 can present the message to business personnel 2, who can then upload the review results to the server via the communication network. If the total amount is lower than or equal to a specified threshold, a notification message is generated based on the user identifier and the total amount, and the notification message is sent to the user device 3 of the salesperson 1. The user device 3 can present the message to the salesperson 1, and the salesperson 1 can upload the settlement result to the server via the communication network through the user device 3. The server can forward the notification message carrying the business result information to the user device 1, and the user device 1 can present the message to the user who applied for the claim. The business result information includes the settlement result or the review result.

[0087] This invention uses a machine model to obtain the ticket category labels and label sets corresponding to the ticket characters determined by the OCR recognition service. It then uses these label sets to divide and determine the correspondence between the ticket category labels and subsets of ticket characters, forming structured data, which is stored in a ticket information database. This results in low cost, short system construction cycle, and portability of the server (which can form a mirror file). It does not rely on predefined keyword and character rules, has high input efficiency, and can universally cover various regions and types of tickets. It is robust and does not require domain experts to manually define character rules or templates, nor does it require considering the characteristics of the characters themselves, such as keywords, special symbols, or statistical information, for recognition matching patterns or keyword matching.

[0088] In a second aspect, embodiments of the present invention also provide an automatic data entry system for medical invoice information under the same inventive concept as the foregoing embodiments. This automatic data entry system may include:

[0089] The recognition module is used to perform optical character recognition on medical invoice images to obtain a set of invoice characters;

[0090] The classification module is used to input the set of bill characters into the machine model to determine the bill category label and form a set of bill category labels. The machine model is used to determine the bill category label corresponding to each bill character in the set of bill characters.

[0091] The structuring module is used to determine a subset of bill characters corresponding to each bill category label from the bill character set based on the bill category label set, so as to form structured data;

[0092] A storage module is used to store the structured data in a bill information database, so that when the bill information database is queried, it returns medical bill information based on the structured data representation.

[0093] Specifically, optical character recognition is performed on the medical invoice image to obtain the invoice character set, including:

[0094] Arrange the boxes to be recognized on the medical invoice image in the recognition order;

[0095] Identify the ticket characters within the frame to be recognized;

[0096] According to the identification order, the determined ticket characters are formed into a ticket character set.

[0097] Specifically, the set of bill characters is input into the machine learning model to determine the bill category labels and form a set of bill category labels, including:

[0098] The set of ticket characters is input into the trained BERT model to extract embedding vectors from the BERT model;

[0099] The embedding vector is input into the trained single-layer neural network model to determine the ticket category label corresponding to each ticket character that is input to the BERT model.

[0100] Specifically, based on the set of bill category labels, a subset of bill characters corresponding to each bill category label is determined from the set of bill characters to form structured data, including:

[0101] Use the bill category labels from the bill category label set as fields of the specified data structure;

[0102] The ticket character corresponding to each ticket category label is used as the value corresponding to the field in the specified data structure;

[0103] Structured data is constructed using the fields and their corresponding values ​​in the specified data structure, according to the specified data format.

[0104] Specifically, among them,

[0105] The machine model includes a pre-trained BERT model and a single-layer neural network model;

[0106] The single-layer neural network model is used to convert the embedding vectors extracted from the BERT model into predicted output vectors within the ticket category label space;

[0107] Each predicted value in the predicted output vector corresponds to a ticket category label;

[0108] The ticket category label corresponding to the maximum predicted value in the predicted output vector is used as the ticket category label corresponding to the ticket character input to the machine model.

[0109] Specifically, the automatic data entry system may further include: a training module for training the machine model, the training module being used for:

[0110] In response to user operation or user device instructions, ticket category labels are marked on ticket characters in ordinary medical ticket information and abnormal medical ticket information respectively to obtain training data, wherein the training data includes marked ticket characters and ticket category labels;

[0111] The BERT model and the single-layer neural network model are trained based on the training data.

[0112] Specifically, in response to user operations or user device commands, bill category labels are used to mark bill characters in ordinary medical bill information and abnormal medical bill information, including:

[0113] In abnormal medical invoice information, the first subset of incomplete or erroneous invoice characters is marked as the same invoice category label;

[0114] The same type of invoice category label is the invoice category label of the second subset of invoice characters in the ordinary medical invoice information corresponding to the abnormal medical invoice information.

[0115] Specifically, among them,

[0116] The difference between the second subset of bill characters and the first subset of bill characters is that, after being placed into the first subset of bill characters according to the order of bill characters in the second subset of bill characters, the semantics of the first subset of bill characters remain complete.

[0117] Specifically, among them,

[0118] The difference between the second subset of ticket characters and the first subset of ticket characters is that, after being replaced by the first subset of ticket characters in the order of the ticket characters in the second subset of ticket characters, the first subset of ticket characters does not contain any erroneous characters.

[0119] Specifically, any one of the ticket category labels includes:

[0120] Invoice title label, invoice code label, item name label, drug quantity label, medical unit label, amount label, or examination category name label.

[0121] Specifically, the automatic data entry system may further include: a monitoring module, which is used for:

[0122] Send a query instruction to the invoice information database, the query instruction carrying user identifier and query time range information;

[0123] Based on the medical invoice information returned by the invoice information database, calculate the total amount corresponding to the user identifier within the query time interval;

[0124] When the total amount exceeds a specified threshold, a list of drug information corresponding to the user identifier is generated;

[0125] Based on the user identifier, the total amount, and the drug information list, a notification message is generated and sent to the business personnel's device.

[0126] In a third aspect, embodiments of the present invention also provide an electronic device under the same inventive concept as the foregoing embodiments. This electronic device includes: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the at least one processor implements the methods in the foregoing embodiments by executing the instructions stored in the memory. Please refer to... Figure 4 An exemplary electronic device is provided, the internal structure of which can be shown in the following diagram. Figure 4As shown, the device can be a server, industrial control computer, terminal equipment, microcontroller, etc. This electronic device includes a processor A01, a network interface A02, and a memory connected via a bus. The processor A01 provides computing, instruction processing, and control capabilities. The memory includes a main memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01 and a computer program B02. The main memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A04. The network interface A02 is used for communication with a network. When the computer program B02 is executed by the processor A01, it implements the methods described in the foregoing embodiments.

[0127] In a fourth aspect, embodiments of the present invention also provide a machine-readable storage medium under the same inventive concept as the foregoing embodiments, storing machine instructions that, when the machine instructions are run on a machine, cause the machine to execute the methods in the foregoing embodiments.

[0128] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0129] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0130] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium may be non-transient and may include various media capable of storing program code, such as USB flash drives, hard disks, read-only memory (ROM), random access memory (RAM), flash memory, magnetic storage, and optical storage.

[0131] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A method for automatic entry of medical ticket information, characterized by, The automatic data entry method includes: Optical character recognition is performed on medical invoice images to obtain the invoice character set; The set of bill characters is input into a machine model to determine the bill category label and form a set of bill category labels. The machine model is used to determine the bill category label corresponding to each bill character in the set of bill characters. Based on the set of bill category labels, a subset of bill characters corresponding to each bill category label is determined from the set of bill characters to form structured data; The structured data is stored in a bill information database so that when the bill information database is queried, it returns medical bill information based on the structured data representation. The machine model includes a pre-trained BERT model and a single-layer neural network model, and the training method of the machine model includes: In response to user operation or user device instructions, ticket category labels are marked on ticket characters in ordinary medical ticket information and abnormal medical ticket information respectively to obtain training data, wherein the training data includes marked ticket characters and ticket category labels; Based on the training data, the BERT model and the single-layer neural network model are trained; The document category tags that, in response to user operations or user device commands, respectively mark document characters in ordinary medical document information and abnormal medical document information, include: Compared to the national drug database, medical insurance item list, and / or custom-labeled data, mark the invoice category labels for the characters in each collected ordinary medical invoice information; In abnormal medical invoice information, the first subset of incomplete or erroneous invoice characters is marked as the same invoice category label; The same type of invoice category label is the invoice category label of the second subset of invoice characters in the ordinary medical invoice information corresponding to the abnormal medical invoice information.

2. The method of automatic entry of medical billing information according to claim 1, wherein, The optical character recognition (OCR) of the medical invoice image to obtain the invoice character set includes: Arrange the boxes to be recognized on the medical invoice image in the recognition order; Identify the ticket characters within the frame to be recognized; According to the identification order, the determined ticket characters are formed into a ticket character set.

3. The method of automatic entry of medical billing information according to claim 2, wherein, The step of inputting the bill character set into the machine model to determine the bill category label and form a bill category label set includes: The set of ticket characters is input into the trained BERT model to extract embedding vectors from the BERT model; The embedding vector is input into the trained single-layer neural network model to determine the ticket category label corresponding to each ticket character that is input to the BERT model.

4. The method of automatic entry of medical billing information of claim 1, wherein, The step of determining a subset of bill characters corresponding to each bill category label from the bill character set, based on the bill category label set, to form structured data, includes: Use the bill category labels from the bill category label set as fields of the specified data structure; The ticket character corresponding to each ticket category label is used as the value corresponding to the field in the specified data structure; Structured data is constructed using the fields and their corresponding values ​​in the specified data structure, according to the specified data format.

5. The method for automatically entering medical invoice information according to claim 1, characterized in that, The single-layer neural network model is used to convert the embedding vectors extracted from the BERT model into predicted output vectors within the ticket category label space; Each predicted value in the predicted output vector corresponds to a ticket category label; The ticket category label corresponding to the maximum predicted value in the predicted output vector is used as the ticket category label corresponding to the ticket character input to the machine model.

6. The method of automatic entry of medical billing information of claim 1, wherein, in, The difference between the second subset of bill characters and the first subset of bill characters is that, after being placed into the first subset of bill characters according to the order of bill characters in the second subset of bill characters, the semantics of the first subset of bill characters remain complete.

7. The method of automatic entry of medical billing information of claim 1, wherein, in, The difference between the second subset of ticket characters and the first subset of ticket characters is that, after being replaced by the first subset of ticket characters in the order of the ticket characters in the second subset of ticket characters, the first subset of ticket characters does not contain any erroneous characters.

8. The method of automatic entry of medical ticket information according to any one of claims 1 to 7, characterized in that, in, Any ticket category label includes: Invoice title label, invoice code label, item name label, drug quantity label, medical unit label, amount label, or examination category name label.

9. The method of automatic entry of medical ticket information according to any one of claims 1 to 7, characterized in that, Applied to a monitoring server, the monitoring method of the monitoring server includes: Send a query instruction to the invoice information database, the query instruction carrying user identifier and query time range information; Based on the medical invoice information returned by the invoice information database, calculate the total amount corresponding to the user identifier within the query time interval; When the total amount exceeds a specified threshold, a list of drug information corresponding to the user identifier is generated; Based on the user identifier, the total amount, and the drug information list, a notification message is generated and sent to the business personnel's device.

10. An automated entry system for medical billing information, characterized by, The automatic data entry system includes: The recognition module is used to perform optical character recognition on medical invoice images to obtain a set of invoice characters; The classification module is used to input the set of bill characters into the machine model to determine the bill category label and form a set of bill category labels. The machine model is used to determine the bill category label corresponding to each bill character in the set of bill characters. The structuring module is used to determine a subset of bill characters corresponding to each bill category label from the bill character set based on the bill category label set, so as to form structured data; A storage module is used to store the structured data in a bill information database, so that when the bill information database is queried, it returns medical bill information based on the structured data representation. The machine model includes a pre-trained BERT model and a single-layer neural network model. The training module is used to respond to user operations or user device instructions to label the bill category tags of bill characters in ordinary medical bill information and abnormal medical bill information respectively, so as to obtain training data, the training data including labeled bill characters and bill category tags; Based on the training data, the BERT model and the single-layer neural network model are trained; The document category tags that, in response to user operations or user device commands, respectively mark document characters in ordinary medical document information and abnormal medical document information, include: Compared to the national drug database, medical insurance item list, and / or custom-labeled data, mark the invoice category labels for the characters in each collected ordinary medical invoice information; In abnormal medical invoice information, the first subset of incomplete or erroneous invoice characters is marked as the same invoice category label; The same type of invoice category label is the invoice category label of the second subset of invoice characters in the ordinary medical invoice information corresponding to the abnormal medical invoice information.

11. An electronic device, comprising: The electronic device includes: At least one processor; A memory connected to the at least one processor; The memory stores instructions executable by the at least one processor, which implements the method described in any one of claims 1 to 9 by executing the instructions stored in the memory.

12. A machine-readable storage medium storing machine instructions that, when executed on a machine, cause the machine to perform the method described in any one of claims 1 to 9.

Images