Artificial intelligence-based anesthesia record and charge sheet error detection system and method

Abstract

The present application provides an artificial intelligence-based anesthesia record and billing sheet error detection system and method, belonging to the technical field of medical information management, which comprises: a quality control model training subsystem for training a quality control artificial intelligence model; an error detection subsystem for detecting errors based on the quality control artificial intelligence model according to input data; a prompt correction subsystem for prompting the target error type and modification suggestions based on the system pop-up window when errors are detected; and an automatic correction subsystem for automatic correction according to the confirmation information of the system pop-up window. The artificial intelligence-based anesthesia record and billing sheet error detection system and method introduce the Deep Seek model to train the quality control artificial intelligence model, the model learns the error-prone points in history, automatically detects errors on the input data, displays the error type and suggestions through the pop-up window, and automatically modifies according to the confirmation information, thereby reducing the labor cost of anesthesia record and billing sheet quality control work and improving the quality control efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical information management, in particular to an anesthesia record and charge slip error detection system and method based on artificial intelligence. BACKGROUND

[0002] Anesthesia record and charge errors can cause serious consequences, involving patient safety, legal disputes, financial losses and medical institution reputation. Medical insurance audit is tedious and strict, involving checking of multiple links such as anesthesia record and charge slip, to ensure the compliance and accuracy of medical behavior. Since medical insurance audit needs to check a large number of items, including whether the choice of anesthesia method is reasonable, whether the dose and type of drug use conform to the standard, whether the operation matches the charge item, and whether there is repetition or error in the charge details, a large amount of manpower is needed for quality control work every day. Such high-repetition and tedious tasks not only increase the workload of medical personnel, but also easily lead to fatigue due to long-term high-intensity work, and then cause human errors.

[0003] Therefore, there is an urgent need for an anesthesia record and charge slip error detection system and method based on artificial intelligence to at least solve the above problems. SUMMARY

[0004] One of the purposes of the present application is to provide an anesthesia record and charge slip error detection system and method based on artificial intelligence, which introduces a Deep Seek model trained instead of artificial anesthesia record and charge slip error detection quality control artificial intelligence model, the model learns the error-prone points in history, automatically detects errors on the input data obtained in real time, and displays the error type and suggestions in a pop-up window. According to the confirmation information, automatic modification is carried out, which greatly reduces the labor cost of anesthesia record and charge slip quality control, and the quality control efficiency is also higher.

[0005] The anesthesia record and charge slip error detection system based on artificial intelligence provided by the embodiment of the present application comprises:

[0006] A quality control model training subsystem for training a quality control artificial intelligence model;

[0007] An error detection subsystem for detecting errors based on the quality control artificial intelligence model according to the input data, wherein the input data includes anesthesia record and charge slip;

[0008] A prompt correction subsystem for prompting the target error type and modification suggestions based on the system pop-up window when an error is detected;

[0009] An automatic correction subsystem for automatically correcting according to the confirmation information of the system pop-up window.

[0010] Preferably, the quality control model training subsystem trains the quality control artificial intelligence model, comprising:

[0011] Obtain the historical annotation records of quality control personnel; the historical annotation records include: error characteristics and primary error type;

[0012] Based on the Deep Seek model, a quality control AI model is trained according to historical annotation records.

[0013] Preferably, the quality control model training subsystem obtains historical annotation records from quality control personnel, including:

[0014] Obtain the initial pre-recorded and labeled records from quality control personnel;

[0015] Obtain the sampling results of the first pre-recorded and labeled records;

[0016] If the sampling inspection result is unqualified, determine the annotator and annotation time of the corresponding first pre-recorded annotation record;

[0017] Based on a pre-defined template for describing non-standard annotation features, describe the primary non-standard annotation features that the annotators target.

[0018] Obtain the first pre-collected annotation records of the annotators within the preset time range of the annotation time, and use them as the second pre-collected annotation records;

[0019] Establish the association between annotation personnel, non-standard annotation features of the first target, and annotation records of the second pre-collected data;

[0020] Based on the first non-standard annotation features of the associated target and the second pre-included annotation records, determine the third pre-included annotation records that need to be removed from the second pre-included annotation records;

[0021] The first and third pre-collection annotation records that failed the sampling inspection were removed, and the remaining first pre-collection annotation records were used as historical annotation records.

[0022] Preferably, the quality control model training subsystem determines, based on the associated first target non-standard annotation features and the second pre-collected annotation records, the third pre-collected annotation records that need to be removed from the second pre-collected annotation records, including:

[0023] Determine the first non-standard annotation type label based on the non-standard annotation characteristics of the first target;

[0024] Based on the first non-standard annotation type label, retrieve the index label in the non-standard annotation feature description template, determine that there is a consistent index label, and use it as the target label;

[0025] Connect to the sub-description template corresponding to the target tag;

[0026] Based on the sub-description template, obtain the second target non-standard annotation features in the second pre-collected annotation record associated with the first target non-standard annotation features corresponding to the sub-description template;

[0027] If successful, the corresponding second pre-collection annotation record will be used as the third pre-collection annotation record.

[0028] Preferably, the automatic correction subsystem performs automatic correction based on the confirmation information in the system pop-up window, including:

[0029] If the confirmation message indicates acceptance of the suggestion, proceed with the actual correction process for the input marker data.

[0030] The standard correction process for obtaining modification suggestions includes: the node type, node task description, ideal execution time of the node task, and node task verification strategy for each process node.

[0031] Based on the actual revised process and the standard revised process, inquire about the process nodes.

[0032] Preferably, based on the actual correction process and the standard correction process, process node inquiries are conducted, including:

[0033] Analyze the actual correction process. If the actual execution time of a node task in a process node is longer than the ideal execution time of the node task, the corresponding process node will be taken as the target process node.

[0034] Based on the query robot model, queries are initiated to the target process node according to the node task description;

[0035] Get a response to your inquiry;

[0036] If the query response fails to be obtained, the management node of the target node is determined based on the node type of the target process node and the preset organizational structure diagram, and the query failure information is sent to the management node.

[0037] If the query response is successfully retrieved, parse the query response to obtain its semantics;

[0038] Based on the semantics of the query response, determine whether adjustments or modifications are needed.

[0039] If adjustments or modifications are needed, adjust the modification suggestions for the target error type corresponding to the input marked data based on the semantics of the query response;

[0040] If no adjustments or modifications are required, adjust the node rating of the target process node based on the semantics of the query response.

[0041] Preferably, based on the semantics of the query response, the modification suggestions for adjusting the input labeled data corresponding to the target error type include:

[0042] The semantic items in the query response semantics are matched with the preset correction trigger words. If the semantic matching degree is greater than the preset semantic matching degree threshold, the corresponding correction trigger word is used as the target correction trigger word.

[0043] Obtain the historical trigger records of the target correction trigger words, and determine the semantic sequence of the historical trigger records based on the historical trigger records;

[0044] Based on the semantic sequence of historical trigger records, the word order features are determined; the word order features include: the relative positions of the semantic terms describing the steps to be corrected and the target correction trigger words, and the relative positions of the semantic terms describing the correction steps and the target correction trigger words.

[0045] Perform feature clustering on word meaning features to obtain the word order features represented by the clusters;

[0046] Based on the word order features represented by clustering, construct a sentence structure deconstruction template;

[0047] Deconstruct the semantics of the query response based on the sentence structure deconstruction template to obtain the deconstruction results;

[0048] Based on the historical trigger records extracted from the cluster-representative word order features, a comprehension model corresponding to the sentence deconstruction template is trained.

[0049] Based on the deconstruction results and the understanding of the model, determine the description of the steps to be corrected and the description of the corrected steps.

[0050] Based on the modification suggestions, the description of the steps to be corrected, and the description of the correction steps, determine the modification suggestions for the target error type corresponding to the adjusted input marker data.

[0051] The artificial intelligence-based error detection method for anesthesia records and billing slips provided in this embodiment of the invention includes:

[0052] Step 1: Train the quality control AI model;

[0053] Step 2: Based on the quality control AI model, perform error detection according to the input data, which includes: anesthesia record sheet and billing slip;

[0054] Step 3: When an error is detected, a system pop-up window will display the target error type and suggested modifications;

[0055] Step 4: Perform automatic correction based on the confirmation information in the system pop-up window.

[0056] The beneficial effects of this invention are as follows:

[0057] This invention introduces a Deep Seek model to train an artificial intelligence quality control model that replaces manual error checking of anesthesia records and billing slips. The model learns from historical error-prone points, automatically checks for errors in real-time input data, displays error types and suggestions in pop-up windows, and automatically modifies based on confirmation information. This greatly reduces the labor costs of quality control work on anesthesia records and billing slips, and also improves quality control efficiency.

[0058] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0059] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a schematic diagram of an anesthesia record and billing slip error detection system based on artificial intelligence in an embodiment of the present invention;

[0062] Figure 2 This is a schematic diagram of an artificial intelligence-based error detection method for anesthesia record sheets and billing slips in an embodiment of the present invention. Detailed Implementation

[0063] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0064] This invention provides an artificial intelligence-based error detection system for anesthesia records and billing slips, such as... Figure 1 As shown, it includes:

[0065] Quality control model training subsystem 1, used to train quality control artificial intelligence models;

[0066] The quality control model training subsystem trains the quality control artificial intelligence model, including:

[0067] Obtain the historical annotation records of quality control personnel; the historical annotation records include: error characteristics and primary error type; quality control personnel are the personnel who check whether there are errors in the anesthesia record sheet and billing slip; error characteristics are: the manifestation of document errors of different primary error types (error classification labels, such as: missing data, logical contradiction), for example: the primary error type is logical contradiction, and the error characteristic is that the billing items do not match the surgical content, or the primary error type is missing data, and the error characteristic is that the corresponding amount content of the billing items is blank;

[0068] Based on the Deep Seek model, a quality control AI model is trained according to historical annotation records. During training, error features are used as model input and the first error type is used as model output. The quality control AI model is an AI model that automatically performs document error detection based on real-time input anesthesia records and billing slips.

[0069] Error detection subsystem 2 is used to perform error detection based on the quality control artificial intelligence model and the input data, which includes anesthesia record sheets and billing slips. The input data is generated in real time.

[0070] The prompt correction subsystem 3 is used to prompt the target error type and modification suggestions based on the system pop-up window when an error is detected; the target error type is: the type of error in the input data detected by the quality control artificial intelligence model; the modification suggestions are generated based on the modification experience of historical records of the same error.

[0071] Automatic Correction Subsystem 4 is used to automatically correct errors based on the confirmation information in the system pop-up window. The confirmation information is: after viewing the system pop-up window, the medical information administrator enters information regarding whether the system will implement the modification suggestions. For example, if the medical information administrator views the system pop-up window and clicks the "Accept Suggestion" icon, the system will automatically implement the modification suggestions; otherwise, it will not implement the modification suggestions.

[0072] The working principle and beneficial effects of the above technical solution are as follows:

[0073] This invention introduces a Deep Seek model to train an artificial intelligence quality control model that replaces manual error checking of anesthesia records and billing slips. The model learns from historical error-prone points, automatically checks for errors in real-time input data, displays error types and suggestions in pop-up windows, and automatically modifies based on confirmation information. This greatly reduces the labor costs of quality control work on anesthesia records and billing slips, and also improves quality control efficiency.

[0074] In one embodiment, the quality control model training subsystem obtains historical annotation records from quality control personnel, including:

[0075] Obtain the first pre-recorded annotation record from the quality control personnel; the first pre-recorded annotation record is: all anesthesia records and billing slips with errors marked by the quality control personnel in the past.

[0076] Obtain the sampling inspection results of the first pre-included annotation records; Sampling inspection results: The quality control management personnel conduct random sampling inspections of the first pre-included annotation records to verify the quality control results;

[0077] If the sampling inspection result is that the sampling inspection is unqualified, determine the labeling personnel and labeling time of the corresponding first pre-collection labeling record; the labeling personnel are: the quality control personnel corresponding to the first pre-collection labeling record for which the sampling inspection result is unqualified;

[0078] Based on a pre-defined template for describing non-standard annotation features, the first target non-standard annotation feature for the annotator is described. The pre-defined template for describing non-standard annotation features pre-sets multiple non-standard annotation types (e.g., inspection time not meeting the standard) and their non-standard annotation behaviors (e.g., inspection time of 3 seconds) and their description rules (e.g., when the inspection time does not meet the standard, record the specific inspection time), which are extracted and defined by the quality control management personnel from historical quality control management records. The first target non-standard annotation feature is: the non-standard annotation feature extracted from the first pre-recorded annotation record of the non-conforming sample, such as: inspection time of 3 seconds.

[0079] The system retrieves the first pre-recorded annotation records from annotators within a preset time range at the annotation time, and uses these records as the second pre-recorded annotation records. The preset time range is manually set, for example, half an hour before and half an hour after the annotation time.

[0080] Establish the association between annotation personnel, non-standard annotation features of the first target, and annotation records of the second pre-collected data;

[0081] Based on the first non-standard annotation features of the associated target and the second pre-included annotation records, determine the third pre-included annotation records that need to be removed from the second pre-included annotation records;

[0082] The first and third pre-collection annotation records that failed the sampling inspection were removed, and the remaining first pre-collection annotation records were used as historical annotation records.

[0083] The working principle and beneficial effects of the above technical solution are as follows:

[0084] The quality of historical annotation records greatly affects the accuracy of the quality control AI model trained on them. The quality of historical annotation records depends on the experience of quality control personnel, their state during quality inspection, etc. Therefore, it is necessary to screen the first pre-collected annotation records (all anesthesia records and billing slips with errors marked by quality control personnel in the past).

[0085] Checking all the first pre-collected annotation records is inefficient. Therefore, the quality control management personnel first conduct random checks to determine the annotators and annotation times of the unqualified first pre-collected annotation records. At the same time, a non-standard annotation feature description template is introduced, and the first target non-standard annotation features of the annotators are described against the descriptions of the unqualified first pre-collected annotation records.

[0086] Generally, the work status of annotation personnel is relatively stable within a certain period of time. Therefore, when non-standard quality inspection behavior occurs at the annotation time, the probability of the second pre-collected annotation records being of unqualified quality within the preset time range of the annotation time will greatly increase. Therefore, based on the non-standard annotation characteristics of the first target and the second pre-collected annotation records, the third pre-collected annotation records that need to be removed from the second pre-collected annotation records are determined. The first and third pre-collected annotation records that fail the sampling inspection are removed, and the remaining first pre-collected annotation records are used as historical annotation records, which improves the error detection accuracy of the quality control artificial intelligence model.

[0087] In one embodiment, the quality control model training subsystem determines, based on the associated first target non-standard annotation features and the second pre-collected annotation records, the third pre-collected annotation records that need to be removed from the second pre-collected annotation records, including:

[0088] The first non-standard annotation type label is determined based on the non-standard annotation features of the first target; the first non-standard annotation type label is the ID of the non-standard annotation type corresponding to the non-standard annotation features of the first target.

[0089] Based on the first non-standard annotation type label, the index label in the non-standard annotation feature description template is retrieved, and it is determined that there is a consistent index label for the label, which is then used as the target label; the index label is: the ID of the non-standard annotation type in the non-standard annotation feature description template;

[0090] Connect to the sub-description template corresponding to the target label; the sub-description template is: the description rules of the non-standard annotation behavior of the non-standard annotation type corresponding to the target label in the non-standard annotation feature description template;

[0091] Based on the sub-description template, obtain the second target non-standard annotation features in the second pre-collected annotation records associated with the first target non-standard annotation features corresponding to the sub-description template; the first target non-standard annotation corresponding to the sub-description template is: the first target non-standard annotation used by the sub-description template, for example: the sub-description template is: when the inspection time does not meet the standard, record the specific inspection time, and the first target non-standard annotation feature is: inspection time 3 seconds; the second target non-standard annotation features are: non-standard annotation features extracted by comparing and extracting the second pre-collected annotation records (the first pre-collected annotation records within the half-hour time range before and after the annotation time of the 3-second inspection time) using the sub-description template, for example: inspection time 2 seconds;

[0092] If successful, the corresponding second pre-collection annotation record will be used as the third pre-collection annotation record.

[0093] The working principle and beneficial effects of the above technical solution are as follows:

[0094] Generally, random sampling uses a non-standard annotation feature description template for comprehensive description. However, this method has low sampling efficiency. This invention effectively solves this problem by leveraging the similarity of non-standard behaviors by the same annotator within the same timeframe. Based on the first non-standard annotation type label of the first target non-standard annotation feature, the ID of the non-standard annotation type in the non-standard annotation feature description template is retrieved. If the types match, the description rule of the corresponding non-standard annotation behavior is used as a sub-description template. The sub-description template is then invoked to attempt to obtain the second target non-standard annotation feature of the second pre-collected annotation record. If the attempt is successful, it indicates that the same type of non-standard annotation behavior exists in the corresponding second pre-collected annotation record, and the corresponding second pre-collected annotation record is used as the third pre-collected annotation record. This allows for adaptive dynamic matching of the sub-description template and determination of the description range (second pre-collected annotation record) based on the non-standard annotation type and annotation time of the sampled first target non-standard annotation feature, improving matching efficiency and further enhancing the efficiency of training data screening.

[0095] In one embodiment, the automatic correction subsystem performs automatic correction based on the confirmation information in the system pop-up window, including:

[0096] If the confirmation message indicates acceptance of the suggestion, proceed with the actual correction process by obtaining the input marker data; the input marker data is the data obtained after the system marks the target error type and correction suggestion on the input data when an error is detected.

[0097] The standard correction process for obtaining modification suggestions includes: the node type, node task description, ideal execution time of the node task, and node task verification strategy for each process node; process nodes are the network communication nodes of the departments (e.g., anesthesiology department, cashier, etc.) involved in each step of the modification process; the node task description is the step description of the execution steps involved in the process node in the modification process; the ideal execution time of the node task is determined based on the complexity of the step description and the number of tasks queued at the process node; the node task verification strategy is the strategy for verifying whether the node task has been completed (e.g., determining whether the signature was successful based on digital signature technology);

[0098] Analyze the actual correction process. If the actual execution time of a node task in a process node is longer than the ideal execution time of the node task, the corresponding process node will be taken as the target process node.

[0099] Based on the query robot model, queries are initiated to the target process node according to the node task description; the query robot model is: an AI model that initiates queries to the target process node that does not execute the process node task in a timely manner.

[0100] Obtain query responses; query responses are the responses from the target process node to the query robot model's queries.

[0101] If the query response fails to be obtained, the management node of the target node is determined based on the node type of the target process node and the preset organizational structure diagram, and the query failure information is sent to the management node; the preset organizational structure diagram is the organizational structure of the management relationship of various departments in the hospital.

[0102] If the query response is successfully retrieved, parse the query response to obtain its semantics;

[0103] Based on the semantics of the query response, determine whether adjustments or modifications are needed. During the determination, the semantics of the query response are input into the query robot model, and the query robot model outputs whether there are any suggestions for adjustments or modifications to the query response.

[0104] If adjustments or modifications are needed, adjust the modification suggestions for the target error type corresponding to the input marked data based on the semantics of the query response;

[0105] If no adjustments or modifications are required, adjust the node rating of the target process node based on the semantics of the query response.

[0106] The working principle and beneficial effects of the above technical solution are as follows:

[0107] The automatic correction process requires automatically connecting to the process nodes of each correction step according to the standard correction procedure and the order of the correction steps. After successful connection, the node task is issued to the process node based on the node task description. At the same time, the ideal execution time of the node task is determined according to the task complexity and the queued tasks of the process node. The actual execution time of the node task is compared with the ideal execution time in real time. When the actual execution time of the node task is longer than the ideal execution time, the corresponding process node (target process node) may not have been processed in time, or an incorrect correction process may have occurred (e.g., an unreasonable process due to inappropriate modification suggestions). Therefore, it is necessary to query the corresponding process node (target process node) in a timely manner. The specific query process is as follows:

[0108] An inquiry robot model is introduced to initiate inquiries about the task descriptions of target process nodes. After the inquiry, the target process node may respond to the inquiry robot model or not. If it does not respond, the non-response situation is reported to the management node according to the node type and the preset organizational structure diagram. If it responds, it is determined whether the target process node has provided adjustment suggestions. If it has provided adjustment suggestions, the process is adjusted and corrected based on the adjustment suggestions. Otherwise, the target process node has caused a delay in the correction process and has not advanced the next step of the process in a timely manner. The adjustment point score is downgraded, which improves the efficiency of automatic correction.

[0109] In one embodiment, based on the query response semantics, the modification suggestion for adjusting the input labeled data corresponding to the target error type includes:

[0110] The semantic items in the query response semantics are matched with preset correction trigger words. If the semantic match is greater than the preset semantic match threshold, the corresponding correction trigger word is used as the target correction trigger word. The response semantic items are specific semantic units or keywords. For example, if the query response semantics are: "This record form should not be approved by me, it should be copied to a in department A", the response semantic items would be: "this record form", "should not", "to", "me", "approval", "should", "copy to", and "a in department A". The preset correction trigger words are set manually, such as "should not", "wrong", and "not".

[0111] Obtain historical trigger records for the target correction trigger words, and determine the semantic sequence of historical trigger records based on the historical trigger records; the historical trigger records are: the reasons for the process error reported by the historical process nodes when the record modification process failed in the past; the semantic sequence of historical trigger records is the semantic unit extracted from the historical trigger records, and the position of the semantic unit in the sequence is relatively consistent with the position of the corresponding extracted record text in the record;

[0112] Based on the semantic sequence of historical trigger records, determine the word order features; the word order features include: the relative position of the semantic item describing the step to be corrected and the target correction trigger word, and the relative position of the semantic item describing the correction step and the target correction trigger word; the relative position includes: the positional relationship between the target correction trigger word and the target correction trigger word (including whether it is before or after the target correction trigger word, and the number of semantic units between them);

[0113] The semantic features are clustered to obtain the cluster representative word order features. The cluster representative word order features are: the feature representation of the declarative sentence structure of the historically commonly used process error reasons, for example: representing the semantic items of the step to be corrected at the 1st, 2nd, and 3rd semantic positions after the target correction trigger word, and the semantic items of the correction step description at the 5th, 6th, and 7th semantic positions after the target correction trigger word.

[0114] Based on the cluster representative word order features, a sentence structure deconstruction template is constructed. The sentence structure deconstruction template is as follows: the query response semantics are divided according to the statement sentence structure of the process error reason corresponding to the cluster representative word order features. For example, the query response semantics at the 1st, 2nd, and 3rd word positions after the target correction trigger word are taken as the semantics of the step to be corrected, and the query response semantics at the 5th, 6th, and 7th word positions after the target correction trigger word are taken as the semantics of the correction step description.

[0115] Deconstruct the semantics of the query response based on the sentence structure deconstruction template to obtain the deconstruction results; the deconstruction results are: query response semantic items marked with the type of deconstructed content (such as: target correction trigger words, semantic items describing the steps to be corrected, and semantic items describing the correction steps, etc.);

[0116] Based on the historical trigger records extracted according to the cluster representative word order features, the understanding model corresponding to the sentence structure deconstruction template is trained. The understanding model includes: a semantic model that understands the target correction trigger content, the description content of the steps to be corrected, and the description content of the correction steps. It is obtained by feeding the content descriptions corresponding to the above content types in the historical trigger records for training.

[0117] Based on the deconstruction results and the understanding of the model, determine the description of the steps to be corrected and the description of the corrected steps.

[0118] The working principle and beneficial effects of the above technical solution are as follows:

[0119] This invention introduces target correction trigger words. Based on the historical trigger records containing the target correction trigger words, the semantic sequence of the historical trigger records is determined. Word order features are extracted and feature clustering is performed to obtain a characteristic representation of the declarative sentence structure of historically commonly used process error reasons (cluster representative word order features). Based on the cluster representative word order features, a sentence structure deconstruction template is constructed. The query response semantics are deconstructed according to the sentence structure deconstruction template to obtain the deconstruction result. The semantic decomposition of the query response semantics is more reasonable. At the same time, the corresponding understanding model of the sentence structure deconstruction template is trained according to the historical trigger records extracted according to the cluster representative word order features. The understanding model performs targeted semantic content understanding based on historical input data of different deconstruction content types, resulting in more accurate understanding and more accurate identification of the correction step description content. Based on the modification suggestions, the description content of the step to be corrected, and the description content of the correction step, the modification suggestions corresponding to the target error type of the adjusted input labeled data are determined, making the suggested adjustment process more reasonable.

[0120] This invention provides an artificial intelligence-based method for detecting errors in anesthesia records and billing slips, such as... Figure 2 As shown, it includes:

[0121] Step 1: Train the quality control AI model;

[0122] Step 2: Based on the quality control AI model, perform error detection according to the input data, which includes: anesthesia record sheet and billing slip;

[0123] Step 3: When an error is detected, a system pop-up window will display the target error type and suggested modifications;

[0124] Step 4: Perform automatic correction based on the confirmation information in the system pop-up window.

[0125] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An AI-based error detection system for anesthesia records and billing slips, characterized in that: include: The quality control model training subsystem is used to train the quality control artificial intelligence model. The error detection subsystem is used to detect errors based on the quality control artificial intelligence model and the input data, which includes: anesthesia record sheet and billing slip; The error correction subsystem is used to prompt the target error type and modification suggestions based on a system pop-up window when an error is detected. The automatic correction subsystem is used to perform automatic corrections based on the confirmation information in the system pop-up window; The quality control model training subsystem trains the quality control artificial intelligence model, including: Obtain the historical annotation records of quality control personnel; the historical annotation records include: error characteristics and primary error type; Based on the Deep Seek model, a quality control AI model is trained according to historical annotation records; The automatic correction subsystem performs automatic corrections based on the confirmation information in the system pop-up window, including: If the confirmation message indicates acceptance of the suggestion, proceed with the actual correction process for the input marker data. The standard correction process for obtaining modification suggestions includes: the node type, node task description, ideal execution time of the node task, and node task verification strategy for each process node. Analyze the actual correction process. If the actual execution time of a node task in a process node is longer than the ideal execution time of the node task, the corresponding process node will be taken as the target process node. Based on the query robot model, queries are initiated to the target process node according to the node task description; Get a response to your inquiry; If the query response fails to be obtained, the management node of the target node is determined based on the node type of the target process node and the preset organizational structure diagram, and the query failure information is sent to the management node. If the query response is successfully retrieved, parse the query response to obtain its semantics; Based on the semantics of the query response, determine whether adjustments or modifications are needed. If adjustments or modifications are needed, adjust the modification suggestions for the target error type corresponding to the input marked data based on the semantics of the query response; If no adjustments or modifications are needed, adjust the node rating of the target process node based on the semantics of the query response; Among them, the modification suggestions for adjusting the target error type corresponding to the input labeled data based on the semantics of the query response include: The semantic items in the query response semantics are matched with the preset correction trigger words. If the semantic matching degree is greater than the preset semantic matching degree threshold, the corresponding correction trigger word is used as the target correction trigger word. Obtain the historical trigger records of the target correction trigger words, and determine the semantic sequence of the historical trigger records based on the historical trigger records; Based on the semantic sequence of historical trigger records, the word order features are determined; the word order features include: the relative positions of the semantic terms describing the steps to be corrected and the target correction trigger words, and the relative positions of the semantic terms describing the correction steps and the target correction trigger words. Perform feature clustering on word meaning features to obtain the word order features represented by the clusters; Based on the word order features represented by clustering, construct a sentence structure deconstruction template; Deconstruct the semantics of the query response based on the sentence structure deconstruction template to obtain the deconstruction results; Based on the historical trigger records extracted from the cluster-representative word order features, a comprehension model corresponding to the sentence deconstruction template is trained. Based on the deconstruction results and the understanding of the model, determine the description of the steps to be corrected and the description of the corrected steps. Based on the modification suggestions, the description of the steps to be corrected, and the description of the correction steps, determine the modification suggestions for the target error type corresponding to the adjusted input marker data.

2. The artificial intelligence-based anesthesia record and billing system as described in claim 1, characterized in that, The quality control model training subsystem obtains historical annotation records from quality control personnel, including: Obtain the initial pre-recorded and labeled records from quality control personnel; Obtain the sampling results of the first pre-recorded and labeled records; If the sampling inspection result is unqualified, determine the annotator and annotation time of the corresponding first pre-recorded annotation record; Based on a pre-defined template for describing non-standard annotation features, describe the primary non-standard annotation features that the annotators target. Obtain the first pre-collected annotation records of the annotators within the preset time range of the annotation time, and use them as the second pre-collected annotation records; Establish the association between annotation personnel, non-standard annotation features of the first target, and annotation records of the second pre-collected data; Based on the first non-standard annotation features of the associated target and the second pre-included annotation records, determine the third pre-included annotation records that need to be removed from the second pre-included annotation records; The first and third pre-collection annotation records that failed the sampling inspection were removed, and the remaining first pre-collection annotation records were used as historical annotation records. Specifically, based on the associated first target non-standard annotation characteristics and the second pre-included annotation records, the third pre-included annotation records that need to be removed from the second pre-included annotation records are determined, including: Determine the first non-standard annotation type label based on the non-standard annotation characteristics of the first target; Based on the first non-standard annotation type label, retrieve the index label in the non-standard annotation feature description template, determine that there is a consistent index label, and use it as the target label; Connect to the sub-description template corresponding to the target tag; Based on the sub-description template, obtain the second target non-standard annotation features in the second pre-collected annotation record associated with the first target non-standard annotation features corresponding to the sub-description template; If successful, the corresponding second pre-collection annotation record will be used as the third pre-collection annotation record.

3. An artificial intelligence-based method for detecting errors in anesthesia records and billing slips, characterized in that: include: Step 1: Train the quality control AI model; Step 2: Based on the quality control AI model, perform error detection according to the input data, which includes: anesthesia record sheet and billing slip; Step 3: When an error is detected, a system pop-up window will display the target error type and suggested modifications; Step 4: Perform automatic correction based on the confirmation information in the system pop-up window; Step 1: Training the quality control AI model, including: Obtain the historical annotation records of quality control personnel; the historical annotation records include: error characteristics and primary error type; Based on the Deep Seek model, a quality control AI model is trained according to historical annotation records; The automatic correction subsystem performs automatic corrections based on the confirmation information in the system pop-up window, including: If the confirmation message indicates acceptance of the suggestion, proceed with the actual correction process for the input marker data. The standard correction process for obtaining modification suggestions includes: the node type, node task description, ideal execution time of the node task, and node task verification strategy for each process node. Analyze the actual correction process. If the actual execution time of a node task in a process node is longer than the ideal execution time of the node task, the corresponding process node will be taken as the target process node. Based on the query robot model, queries are initiated to the target process node according to the node task description; Get a response to your inquiry; If the query response fails to be obtained, the management node of the target node is determined based on the node type of the target process node and the preset organizational structure diagram, and the query failure information is sent to the management node. If the query response is successfully retrieved, parse the query response to obtain its semantics; Based on the semantics of the query response, determine whether adjustments or modifications are needed. If adjustments or modifications are needed, adjust the modification suggestions for the target error type corresponding to the input marked data based on the semantics of the query response; If no adjustments or modifications are needed, adjust the node rating of the target process node based on the semantics of the query response; Among them, the modification suggestions for adjusting the target error type corresponding to the input labeled data based on the semantics of the query response include: The semantic items in the query response semantics are matched with the preset correction trigger words. If the semantic matching degree is greater than the preset semantic matching degree threshold, the corresponding correction trigger word is used as the target correction trigger word. Obtain the historical trigger records of the target correction trigger words, and determine the semantic sequence of the historical trigger records based on the historical trigger records; Based on the semantic sequence of historical trigger records, the word order features are determined; the word order features include: the relative positions of the semantic terms describing the steps to be corrected and the target correction trigger words, and the relative positions of the semantic terms describing the correction steps and the target correction trigger words. Perform feature clustering on word meaning features to obtain the word order features represented by the clusters; Based on the word order features represented by clustering, construct a sentence structure deconstruction template; Deconstruct the semantics of the query response based on the sentence structure deconstruction template to obtain the deconstruction results; Based on the historical trigger records extracted from the cluster-representative word order features, a comprehension model corresponding to the sentence deconstruction template is trained. Based on the deconstruction results and the understanding of the model, determine the description of the steps to be corrected and the description of the corrected steps. Based on the modification suggestions, the description of the steps to be corrected, and the description of the correction steps, determine the modification suggestions for the target error type corresponding to the adjusted input marker data.

4. The method for detecting errors in anesthesia records and billing slips based on artificial intelligence as described in claim 3, characterized in that, Obtain the historical annotation records of quality control personnel, including: Obtain the initial pre-recorded and labeled records from quality control personnel; Obtain the sampling results of the first pre-recorded and labeled records; If the sampling inspection result is unqualified, determine the annotator and annotation time of the corresponding first pre-recorded annotation record; Based on a pre-defined template for describing non-standard annotation features, describe the primary non-standard annotation features that the annotators target. Obtain the first pre-collected annotation records of the annotators within the preset time range of the annotation time, and use them as the second pre-collected annotation records; Establish the association between annotation personnel, non-standard annotation features of the first target, and annotation records of the second pre-collected data; Based on the first non-standard annotation features of the associated target and the second pre-included annotation records, determine the third pre-included annotation records that need to be removed from the second pre-included annotation records; The first and third pre-collection annotation records that failed the sampling inspection were removed, and the remaining first pre-collection annotation records were used as historical annotation records. Specifically, based on the associated first target non-standard annotation characteristics and the second pre-included annotation records, the third pre-included annotation records that need to be removed from the second pre-included annotation records are determined, including: Determine the first non-standard annotation type label based on the non-standard annotation characteristics of the first target; Based on the first non-standard annotation type label, retrieve the index label in the non-standard annotation feature description template, determine that there is a consistent index label, and use it as the target label; Connect to the sub-description template corresponding to the target tag; Based on the sub-description template, obtain the second target non-standard annotation features in the second pre-collected annotation record associated with the first target non-standard annotation features corresponding to the sub-description template; If successful, the corresponding second pre-collection annotation record will be used as the third pre-collection annotation record.

Images