Medical record diagnosis missing writing detection method, device and equipment and readable storage medium

Abstract

The application discloses a medical record diagnosis omission detection method and device, equipment and a readable storage medium. In the scheme, after determining the medical record text to be detected, the candidate omission diagnosis corresponding to the medical record text is determined. Then, the target candidate omission diagnosis is determined from the candidate omission diagnosis. The target candidate omission diagnosis is the candidate omission diagnosis consistent with the written diagnosis semantics in the medical record text, and the candidate omission diagnosis determined as the written diagnosis symptom. Finally, the other candidate omission diagnosis in the candidate omission diagnosis except the target candidate omission diagnosis is determined as the omission diagnosis. The scheme detects the medical record diagnosis omission from two aspects of diagnosis semantic similarity and diagnosis symptom judgment. When a candidate omission diagnosis is a written diagnosis symptom, even if the semantics are consistent, the candidate omission diagnosis can still be avoided to be detected as an omission diagnosis, thereby improving the accuracy of the medical record diagnosis omission detection result.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and more specifically, to a method, apparatus, device, and readable storage medium for detecting omissions in medical record diagnosis. Background Technology

[0002] Completing medical record diagnoses is a crucial step in the medical field. Complete and accurate diagnoses help reduce the risk of medical disputes, and are of great significance for analyzing subsequent treatment behaviors, optimizing resource allocation, and improving the quality of medical services. Therefore, detection of omissions in medical record diagnoses is particularly necessary.

[0003] Currently, the detection of omissions in medical record diagnoses is generally done manually. This involves doctors or hospital medical record quality control personnel reviewing the records and marking any diagnoses that may be omitted. However, the actual writing of medical records is highly complex, making manual detection of omissions a massive undertaking with low efficiency. Furthermore, the professional competence of the personnel significantly impacts the accuracy of the results. To address the low efficiency and accuracy issues caused by manual omission detection, some existing technologies utilize deep learning models, rule bases, or knowledge bases. However, these technologies rely solely on semantic similarity to detect omissions. In reality, two semantically dissimilar diagnoses may be the same diagnosis, leading to further low accuracy in the detection results.

[0004] Therefore, how to provide a method for detecting omissions in medical record diagnosis to improve the accuracy of the detection results has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, this application proposes a method, apparatus, device, and readable storage medium for detecting omissions in medical record diagnoses. The specific solution is as follows:

[0006] A method for detecting omissions in medical record diagnoses, the method comprising:

[0007] Identify the medical record text to be tested;

[0008] Identify the candidate missing diagnoses corresponding to the medical record text;

[0009] From the candidate omission diagnoses, a target candidate omission diagnosis is determined, wherein the target candidate omission diagnosis is a candidate omission diagnosis that is semantically consistent with the written diagnosis in the medical record text, and a candidate omission diagnosis that is determined to be a symptom of the written diagnosis.

[0010] The candidate write-missing diagnoses other than the target candidate write-missing diagnoses are identified as write-missing diagnoses.

[0011] Optionally, determining the candidate missed diagnosis corresponding to the medical record text includes:

[0012] Generate prompt information corresponding to the medical record text, the prompt information being used to indicate the candidate missing diagnosis corresponding to the medical record text;

[0013] Invoke the generative model and input the prompt information corresponding to the medical record text into the generative model;

[0014] Based on the response from the generative model, candidate omission diagnoses corresponding to the medical record text are determined.

[0015] Optionally, the step of generating the prompt information corresponding to the medical record text includes:

[0016] The medical record text is segmented to obtain multiple medical record text fragments;

[0017] For each medical record text segment, a prompt message corresponding to the medical record text segment is generated. The prompt message is used to indicate the candidate missing diagnosis corresponding to the medical record text segment. The combination of the prompt messages corresponding to each medical record text segment is the prompt message corresponding to the medical record text.

[0018] Optionally, the generative model is obtained by fine-tuning an existing generative model in the medical field; the fine-tuning of the existing generative model in the medical field includes:

[0019] The training medical record text, the diagnosis corresponding to the training medical record text, and the tags of each diagnosis in the training medical record text are obtained. The tags of each diagnosis in the training medical record text include a first tag, a second tag, and a third tag. The first tag is used to indicate whether the diagnosis is confirmed, the second tag is used to indicate the diagnosis type, and the third tag is used to indicate the recall support fragment of the diagnosis.

[0020] Based on the training medical record text, the diagnosis corresponding to the training medical record text, and the labels of each diagnosis in the training medical record text, multiple training corpora are generated, each training corpus including input text and expected output text.

[0021] The input text is input into an existing generative model in the medical field. The goal is to make the output of the existing generative model in the medical field approach the expected output text corresponding to the input text. The existing generative model in the medical field is then fine-tuned.

[0022] Optionally, determining the target candidate write-missing diagnosis from the candidate write-missing diagnoses includes:

[0023] Based on the candidate missing write diagnosis and the written diagnosis, multiple diagnosis pairs are determined, each diagnosis pair including one candidate missing write diagnosis and one written diagnosis;

[0024] For each diagnostic pair, determine whether the semantics of the two diagnoses in the diagnostic pair are consistent;

[0025] If the two diagnoses in the diagnosis pair have the same semantics, then the candidate missing diagnosis in the diagnosis pair is determined as the target candidate missing diagnosis.

[0026] If the two diagnoses in the diagnosis pair are semantically inconsistent, the diagnosis pair is compared with each diagnosis symptom pair in the preset set of diagnosis symptom pairs.

[0027] If the diagnosis pair matches a diagnosis symptom pair in the set of diagnosis symptom pairs, then the candidate missed diagnosis is determined as the target candidate missed diagnosis.

[0028] Optionally, determining multiple diagnostic pairs based on the candidate missed write diagnoses and the written diagnoses includes:

[0029] The candidate missing write diagnoses are normalized to obtain normalized candidate missing write diagnoses.

[0030] The written diagnosis is normalized to obtain the normalized written diagnosis;

[0031] Based on the normalized candidate missed write diagnoses and the normalized written diagnoses, the multiple diagnosis pairs are determined.

[0032] Optionally, determining multiple diagnostic pairs based on the candidate missed write diagnoses and the written diagnoses includes:

[0033] The candidate missed write diagnosis and the written diagnosis are combined in pairs to form candidate diagnosis pairs; each candidate diagnosis pair includes one candidate missed write diagnosis and one written diagnosis.

[0034] Each of the candidate diagnostic pairs is normalized to obtain normalized candidate diagnostic pairs;

[0035] The candidate diagnostic pairs after normalization are deduplicated to determine the plurality of diagnostic pairs.

[0036] Optionally, the method for determining the preset set of diagnostic symptoms includes:

[0037] Obtain the reference medical record text, the candidate missed diagnoses corresponding to the reference medical record text, and the written diagnoses corresponding to the reference medical record text;

[0038] The written diagnoses corresponding to the reference medical record text are combined in pairs to obtain a first set of diagnosis pairs. The candidate missing diagnoses corresponding to the reference medical record text and the written diagnoses corresponding to the reference medical record text are combined in pairs to obtain a second set of diagnosis pairs.

[0039] Calculate the co-occurrence probability of each of the first diagnostic pairs in the first diagnostic pair set;

[0040] Based on the co-occurrence probability of each of the first diagnostic pairs in the first diagnostic pair set, the set of diagnostic symptom pairs is determined from the first diagnostic pair set and the second diagnostic pair set.

[0041] A medical record diagnostic omission detection device, the device comprising:

[0042] The medical record text determination unit is used to determine the medical record text to be detected.

[0043] A candidate omission diagnosis determination unit is used to determine the candidate omission diagnoses corresponding to the medical record text.

[0044] A target candidate omission diagnosis determination unit is used to determine a target candidate omission diagnosis from the candidate omission diagnoses, wherein the target candidate omission diagnosis is a candidate omission diagnosis that is semantically consistent with the written diagnosis in the medical record text, and a candidate omission diagnosis that is determined to be a symptom of the written diagnosis.

[0045] The missing write diagnosis determination unit is used to determine other candidate missing write diagnoses besides the target candidate missing write diagnosis as missing write diagnoses.

[0046] Optionally, the candidate missing write diagnosis and determination unit includes:

[0047] The prompt information generation subunit is used to generate prompt information corresponding to the medical record text, and the prompt information is used to indicate the candidate missing diagnosis corresponding to the medical record text.

[0048] The generative model calling subunit is used to call the generative model and input the prompt information corresponding to the medical record text into the generative model;

[0049] The candidate omission diagnosis determination subunit is used to determine the candidate omission diagnosis corresponding to the medical record text based on the response of the generative model.

[0050] Optionally, the prompt information generation subunit is specifically used for:

[0051] The medical record text is segmented to obtain multiple medical record text fragments;

[0052] For each medical record text segment, a prompt message corresponding to the medical record text segment is generated. The prompt message is used to indicate the candidate missing diagnosis corresponding to the medical record text segment. The combination of the prompt messages corresponding to each medical record text segment is the prompt message corresponding to the medical record text.

[0053] Optionally, the generative model is obtained by fine-tuning an existing generative model in the medical field; the device includes a fine-tuning unit, which is specifically used for:

[0054] The training medical record text, the diagnosis corresponding to the training medical record text, and the tags of each diagnosis in the training medical record text are obtained. The tags of each diagnosis in the training medical record text include a first tag, a second tag, and a third tag. The first tag is used to indicate whether the diagnosis is confirmed, the second tag is used to indicate the diagnosis type, and the third tag is used to indicate the recall support fragment of the diagnosis.

[0055] Based on the training medical record text, the diagnosis corresponding to the training medical record text, and the labels of each diagnosis in the training medical record text, multiple training corpora are generated, each training corpus including input text and expected output text.

[0056] The input text is input into an existing generative model in the medical field. The goal is to make the output of the existing generative model in the medical field approach the expected output text corresponding to the input text. The existing generative model in the medical field is then fine-tuned.

[0057] Optionally, the target candidate missing write diagnosis and determination unit includes:

[0058] The diagnostic pair determination unit is used to determine multiple diagnostic pairs based on the candidate missing write diagnosis and the written diagnosis, wherein each diagnostic pair includes one candidate missing write diagnosis and one written diagnosis;

[0059] The judgment unit is used to determine, for each diagnosis pair, whether the semantics of the two diagnoses in the diagnosis pair are consistent;

[0060] The first processing unit is configured to determine the candidate missing diagnosis in the diagnosis pair as the target candidate missing diagnosis if the two diagnoses in the diagnosis pair have the same semantics.

[0061] The comparison unit is used to compare the diagnosis pair with each diagnosis symptom pair in a preset set of diagnosis symptom pairs if the semantics of the two diagnoses in the diagnosis pair are inconsistent.

[0062] The second processing unit is configured to determine the candidate missed diagnosis as the target candidate missed diagnosis if the diagnosis pair matches a certain diagnosis symptom pair in the set of diagnosis symptom pairs.

[0063] Optionally, the diagnostic determination unit is specifically used for:

[0064] The candidate missing write diagnoses are normalized to obtain normalized candidate missing write diagnoses.

[0065] The written diagnosis is normalized to obtain the normalized written diagnosis;

[0066] Based on the normalized candidate missed write diagnoses and the normalized written diagnoses, the multiple diagnosis pairs are determined.

[0067] Optionally, the diagnostic determination unit is specifically used for:

[0068] The candidate missed write diagnosis and the written diagnosis are combined in pairs to form candidate diagnosis pairs; each candidate diagnosis pair includes one candidate missed write diagnosis and one written diagnosis.

[0069] Each of the candidate diagnostic pairs is normalized to obtain normalized candidate diagnostic pairs;

[0070] The candidate diagnostic pairs after normalization are deduplicated to determine the plurality of diagnostic pairs.

[0071] Optionally, the device further includes a diagnostic symptom pair determination unit, which is specifically used for:

[0072] Obtain the reference medical record text, the candidate missed diagnoses corresponding to the reference medical record text, and the written diagnoses corresponding to the reference medical record text;

[0073] The written diagnoses corresponding to the reference medical record text are combined in pairs to obtain a first set of diagnosis pairs. The candidate missing diagnoses corresponding to the reference medical record text and the written diagnoses corresponding to the reference medical record text are combined in pairs to obtain a second set of diagnosis pairs.

[0074] Calculate the co-occurrence probability of each of the first diagnostic pairs in the first diagnostic pair set;

[0075] Based on the co-occurrence probability of each of the first diagnostic pairs in the first diagnostic pair set, the set of diagnostic symptom pairs is determined from the first diagnostic pair set and the second diagnostic pair set.

[0076] A medical record diagnostic omission detection device, comprising a memory and a processor;

[0077] The memory is used to store programs;

[0078] The processor is used to execute the program to implement the various steps of the medical record diagnosis omission detection method as described above.

[0079] A readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the various steps of the medical record diagnosis omission detection method as described above.

[0080] By employing the above technical solution, this application discloses a method, apparatus, device, and readable storage medium for detecting omissions in medical record diagnoses. In this solution, after determining the medical record text to be detected, candidate omission diagnoses corresponding to the medical record text are first identified; then, a target candidate omission diagnose is determined from the candidate omission diagnoses. The target candidate omission diagnose is a candidate omission diagnose semantically consistent with the written diagnosis in the medical record text, as well as a candidate omission diagnose for a symptom that is determined to be a written diagnosis; finally, the other candidate omission diagnoses besides the target candidate omission diagnose are determined as omission diagnoses. This solution detects omissions in medical record diagnoses from two aspects: semantic similarity of diagnoses and symptom determination of diagnoses. When a candidate omission diagnose is a symptom of a written diagnosis, even if the two are semantically consistent, the candidate omission diagnose can still be prevented from being detected as an omission diagnose, thereby improving the accuracy of the medical record omission detection results. Attached Figure Description

[0081] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0082] Figure 1 This is a flowchart illustrating a method for detecting omissions in medical record diagnosis disclosed in an embodiment of this application;

[0083] Figure 2 This is a flowchart illustrating a method for determining candidate missing diagnoses corresponding to medical record text using a generative large model, as disclosed in an embodiment of this application.

[0084] Figure 3 This is a flowchart illustrating a method for fine-tuning an existing generative model in the medical field, as disclosed in an embodiment of this application.

[0085] Figure 4 This is a flowchart illustrating a method for determining a target candidate write-missing diagnosis from the candidate write-missing diagnoses disclosed in an embodiment of this application;

[0086] Figure 5 This is a flowchart illustrating a method for determining a preset set of diagnostic symptom pairs disclosed in an embodiment of this application;

[0087] Figure 6 This is a schematic diagram of a medical record diagnostic omission detection device disclosed in an embodiment of this application;

[0088] Figure 7 This is a hardware structure block diagram of a medical record diagnostic omission detection device disclosed in an embodiment of this application. Detailed Implementation

[0089] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0090] The following embodiments will be used to describe the medical record diagnosis omission detection method provided in this application.

[0091] Reference Figure 1 , Figure 1 This is a flowchart illustrating a method for detecting omissions in medical record diagnosis disclosed in an embodiment of this application. The method may include the following steps:

[0092] Step S101: Determine the medical record text to be detected.

[0093] In this application, the medical record text is the text corresponding to the patient's medical record. It can be an electronic medical record entered based on the patient's self-report and the doctor's inquiry, or it can be the text obtained by optical character recognition of the medical record image obtained by scanning or photographing a handwritten paper medical record. This application does not impose any limitations on this.

[0094] The medical record text to be detected is the medical record text for which a diagnosis omission detection is required. It can be the original medical record text to be detected, or the medical record text obtained after preprocessing the original medical record text to be detected. Preprocessing includes desensitization and information filtering. Specifically, desensitization and information filtering can involve deleting fields such as personal or physician name, address, admission time, bed number, differential diagnosis, etc. from the original medical record text to obtain the medical record text to be detected. This application does not impose any restrictions on this.

[0095] Step S102: Determine the candidate missing diagnosis corresponding to the medical record text.

[0096] Medical records contain extensive descriptions of the patient's condition and information about the examinations and treatments received, as well as conclusive information from the doctor's diagnosis. The descriptions of the condition and the information about the examinations and treatments received can indirectly reflect the possible diseases the patient may have, while the conclusive information from the doctor's diagnosis directly reflects the disease the patient has. In this application, the candidate omitted diagnoses corresponding to the medical record text are the diseases the patient may have. These candidate omitted diagnoses may or may not be directly recorded in the medical record text; this application does not impose any limitations on this.

[0097] In this application, there are multiple ways to determine the candidate omission diagnoses corresponding to medical record text. One possible implementation is to use the disease name as the named body and perform named body recognition on the medical record text. Alternatively, information on disease standards or specific drugs can be pre-collected, and the names of examination items or drugs can be used as the named bodies to perform named body recognition on the medical record text to obtain the names of examination items or drugs contained in the medical record text, and then match them with the information on disease standards or specific drugs, thereby determining the candidate omission diagnoses corresponding to the medical record text. Furthermore, the diseases contained in sample medical record texts can be pre-labeled, and a deep learning model can be trained based on this, enabling the deep learning model to learn the mapping relationship between medical record text and diseases, thus outputting the diseases contained in the input medical record text as candidate omission diagnoses. Additionally, generative models such as ChatGPT, PaLM (Pathways Language Model), Pangu Big Data Model, and Xinghuo Cognitive Big Data Model possess strong natural language understanding capabilities, and in this application, generative big data models can also be used to determine the candidate omission diagnoses corresponding to medical record text. This application does not impose any limitations on these methods.

[0098] It should be noted that, in this application, in order to improve the completeness of the candidate omission diagnosis corresponding to the medical record text, the medical record text can be processed in various ways as described above to obtain multiple candidate omission diagnosis results, and then the multiple candidate omission diagnosis results can be merged to obtain the final candidate omission diagnosis. This application does not limit the scope of the application.

[0099] Step S103: Determine a target candidate omission diagnosis from the candidate omission diagnoses, wherein the target candidate omission diagnosis is a candidate omission diagnosis that is semantically consistent with the written diagnosis in the medical record text, and a candidate omission diagnosis that is determined to be a symptom of the written diagnosis.

[0100] The method to determine the written diagnoses in the medical record text can be to obtain the diagnosis list field in the medical record text, such as the discharge diagnosis list field on the first page of the medical record text or the discharge diagnosis list field in the discharge record text, and determine the written diagnoses in the medical record text based on the diagnosis list field in the medical record text.

[0101] In this application, for each candidate omitted diagnosis, it is necessary to determine whether it is indeed an omitted diagnosis. Generally, if two diagnoses are semantically identical, they can be considered the same diagnosis. Therefore, in this application, when a candidate omitted diagnosis is semantically identical to an already written diagnosis, the candidate omitted diagnosis can be excluded. However, in actual medical record writing and quality evaluation, according to textbooks such as the 8th edition of *Clinical Diagnostics* and medical record writing standards, some diagnoses of non-specific clinical symptoms and signs in the early stages of admission should be excluded if a clear diagnosis is obtained after in-depth examination and comprehensive analysis. The phrase "diagnosis of non-specific clinical symptoms and signs in the early stages of admission" may not be semantically identical to "a clear diagnosis after in-depth examination and comprehensive analysis." In such cases, "diagnosis of non-specific clinical symptoms and signs in the early stages of admission" will be identified as an omitted diagnosis. Therefore, in this application, candidate omitted diagnoses that are not omitted diagnoses are excluded based on both the semantic similarity of the diagnosis and the symptom determination of the diagnosis.

[0102] In this application, the candidate missing diagnosis can be compared with the written diagnosis in the medical record text based on the semantic similarity of the diagnosis and the symptom determination of the diagnosis, and then the target candidate missing diagnosis can be determined from the candidate missing diagnosis. The specific implementation method will be described in detail through the following embodiments, and will not be described here.

[0103] Step S104: Determine the candidate write-missing diagnoses other than the target candidate write-missing diagnoses as write-missing diagnoses.

[0104] This embodiment discloses a method for detecting omissions in medical record diagnoses. In this scheme, after determining the medical record text to be detected, candidate omission diagnoses corresponding to the medical record text are first identified. Then, a target candidate omission diagnosis is determined from the candidate omission diagnoses. The target candidate omission diagnosis is one that is semantically consistent with a written diagnosis in the medical record text, as well as a candidate omission diagnosis for a symptom that is determined to be a written diagnosis. Finally, the other candidate omission diagnoses besides the target candidate omission diagnosis are determined as omission diagnoses. This scheme detects omissions in medical record diagnoses from two aspects: semantic similarity of diagnoses and symptom determination. When a candidate omission diagnosis is a symptom of a written diagnosis, even if the two are semantically consistent, the candidate omission diagnosis can still be avoided from being detected as an omission diagnosis, thereby improving the accuracy of the medical record omission detection results.

[0105] As mentioned in the above embodiments, generative models such as ChatGPT, PaLM (Pathways Language Model), Pangu Big Data Model, and Xinghuo Cognitive Big Data Model possess strong natural language understanding capabilities. In this application, generative big data models can also be used to determine candidate omission diagnoses corresponding to medical record texts. Next, in another embodiment of this application, the specific implementation method of using generative big data models to determine candidate omission diagnoses corresponding to medical record texts will be described, referring to... Figure 2 , Figure 2 This is a flowchart illustrating a method for determining candidate missing diagnoses corresponding to medical record text using a generative large model, as disclosed in an embodiment of this application. The method may include the following steps:

[0106] Step S201: Generate prompt information corresponding to the medical record text, the prompt information being used to indicate the candidate missing diagnosis corresponding to the medical record text.

[0107] In this application, specific questions and preset prompt information templates can be preset. The preset specific questions can be, for example, "What diagnoses are present in this medical record text?" or "What diagnoses are contained in the above medical record text?" The preset prompt information templates can include slots for filling preset questions and slots for filling medical record text.

[0108] As one possible implementation, the medical record text and a preset specific question can be filled into the corresponding slot of a prompt message template to generate a prompt message corresponding to the medical record text.

[0109] For ease of understanding, let's assume the preset prompt template is "Given medical record text: {input}\{prompt}", where the "{input}" slot is used to fill the medical record text, and the "{prompt}" slot is used to fill the preset question, "What diagnoses are included in the above medical record text?". The medical record text is: "Dr. Teng**, during rounds, diagnosed the patient with intestinal adhesions, gallstones with atrophic cholecystitis, and common bile duct stones. The patient underwent laparoscopic adhesiolysis, cholecystectomy, common bile duct incision and stone removal, and T-tube drainage. The surgery was successful, and the patient's vital signs are currently stable. There is no bile-like or turbid fluid in the abdominal drainage tube, and no signs of peritonitis in the abdomen, indicating no bile leakage or intestinal fistula." The corresponding prompt message for this medical record text could be:

[0110] "Given medical record text: Chief Physician Teng**, during rounds, diagnosed the patient with intestinal adhesions, gallstones with atrophic cholecystitis, and common bile duct stones. The patient underwent laparoscopic adhesiolysis, cholecystectomy, common bile duct incision and stone removal, and T-tube drainage. The surgery was successful, and the patient's vital signs are currently stable. There is no bile-like or turbid fluid in the abdominal drainage tube, and there are no signs of peritonitis in the abdomen, indicating that there is no bile leakage or intestinal fistula."

[0111] What diagnoses are included in the above medical record text?

[0112] Considering that the length of medical record text may exceed the text length threshold that generative models can process, as an alternative implementation, the medical record text can be segmented into multiple medical record text fragments. For each medical record text fragment, a corresponding prompt message is generated. This prompt message is used to indicate the candidate missing diagnosis corresponding to the medical record text fragment. The combination of the prompt messages corresponding to each medical record text fragment constitutes the prompt message corresponding to the entire medical record text.

[0113] In this application, there can be a variety of ways to segment the medical record text, such as segmenting by length (e.g., segmenting by every 256 characters), segmenting by field (e.g., extracting fields such as diagnosis process and ward round records), segmenting by specific punctuation marks (e.g., segmenting the medical record text by “.”), and one or more combinations thereof. This application does not impose any limitations on these methods.

[0114] In this application, for each medical record text fragment, the medical record text fragment and a preset specific question can be filled into the corresponding slot of a prompt information template to generate a prompt information corresponding to the medical record text fragment.

[0115] For ease of understanding, let's assume the medical record text is 'text'. We can then segment it into multiple medical record text fragments: text1, text2, ..., text2. n Where n is an integer greater than or equal to 1, the preset prompt template is "Given medical record text: {input}\{prompt}", where the "{input}" slot is used to fill in the medical record text fragment, and the "{prompt}" slot is used to fill in the preset question, which is "What diagnoses are contained in the above medical record text?". The specific prompt information corresponding to each of the medical record text fragments is as follows:

[0116] Given medical record text: text1

[0117] What diagnoses are included in the above medical record?

[0118] Given medical record text: text2

[0119] What diagnoses are included in the above medical record?

[0120] ...

[0121] Given medical record text: text n

[0122] What diagnoses are included in the above medical record?

[0123] Step S202: Invoke the generative model and input the prompt information corresponding to the medical record text into the generative model.

[0124] In this application, if the prompt information corresponding to the medical record text includes prompt information corresponding to multiple medical record text fragments, when inputting the prompt information corresponding to each of the medical record text fragments into the generative model, one prompt information corresponding to one medical record text fragment can be input at a time, or multiple prompt information corresponding to multiple medical record text fragments can be input at a time. This application does not impose any limitations on this.

[0125] It should be noted that the generative model in this application can be an existing generative model in the medical field. Considering that existing large-scale generative models in the medical field have poor understanding of medical record text, this application can also use training medical record text to fine-tune the existing generative model in the medical field to obtain a fine-tuned generative model for this step. The fine-tuning method for the existing generative model in the medical field will be described in detail in the following embodiments, and will not be described here.

[0126] Step S203: Based on the response of the generative model, determine the candidate omission diagnosis corresponding to the medical record text.

[0127] In this application, the candidate omission diagnoses corresponding to the medical record text are structured information, such as a list containing multiple elements, each representing a candidate omission diagnosis. The response of the generative model is generally natural language description information, from which the candidate omission diagnoses corresponding to the medical record text can be extracted using methods such as regular expression matching.

[0128] For example, the response from a generative model would be "The medical record text contains diagnosis1, diagnosis2, ..., diagnosis". n Then, the candidate missing diagnoses corresponding to the medical record text can be [diagnosis1, diagnosis2, ..., diagnosis2]. n ].

[0129] Existing technologies for detecting missing medical record diagnoses using deep learning models require multiple models to recall candidate missing diagnoses in stages, which is prone to error propagation and necessitates repeated analysis of the same medical record text, resulting in low detection efficiency. This embodiment solves the problem of error propagation caused by staged recall by using a generative large-scale model with strong natural language understanding capabilities to analyze the medical record text. The generative model can consider the semantic relationships between different diagnoses in the medical record text, making the recalled candidate missing diagnoses more accurate and comprehensive, and also avoiding repeated analysis of the same medical record text, thus improving detection efficiency.

[0130] Next, in another embodiment of this application, a detailed description will be given of how to fine-tune existing generative models in the medical field, referring to... Figure 3 , Figure 3 This is a flowchart illustrating a method for fine-tuning an existing generative model in the medical field, as disclosed in an embodiment of this application. The method may include the following steps:

[0131] Step S301: Obtain the training medical record text, the diagnosis corresponding to the training medical record text, and the tags of each diagnosis in the training medical record text. The tags of each diagnosis in the training medical record text include a first tag, a second tag, and a third tag. The first tag is used to indicate whether the diagnosis is confirmed, the second tag is used to indicate the diagnosis type, and the third tag is used to indicate the recall support fragment of the diagnosis.

[0132] In this application, the method for determining the training medical record text can be to first collect original medical record texts from different sources and departments, preprocess the collected original medical record texts to obtain preprocessed medical record texts, and then segment the preprocessed long medical record texts to obtain the training medical record texts. The preprocessing includes desensitization and information filtering. Specifically, desensitization and information filtering can involve deleting fields such as personal or physician names, addresses, admission times, bed numbers, and differential diagnoses from the original medical record texts. There can be various ways to segment the medical record texts, such as segmenting by length (e.g., segmenting by every 256 characters), segmenting by fields (e.g., extracting fields such as diagnosis process and ward round records), segmenting by specific punctuation marks (e.g., segmenting the medical record text by "."), and one or more combinations thereof. This application does not impose any limitations on these methods.

[0133] As one possible implementation, the diagnosis corresponding to the training medical record text can be determined manually or by using a deep learning model, rule base, or knowledge base.

[0134] For example, the training medical record text reads: "Dr. Teng**, during ward rounds, diagnosed the patient with intestinal adhesions, gallstones with atrophic cholecystitis, and common bile duct stones. The patient underwent laparoscopic adhesiolysis, cholecystectomy, common bile duct incision and stone removal, and T-tube drainage. The surgery was successful, and the patient's vital signs are currently stable. There is no bile-like or turbid fluid in the abdominal drainage tube, and no signs of peritonitis indicate no bile leakage or intestinal fistula." Entity extraction can be used to extract diagnostic entities from this training medical record text, yielding the corresponding diagnosis: "Intestinal adhesions, intestinal fistula, gallstones with atrophic cholecystitis."

[0135] The labels for each diagnosis in the training medical record text can be manually labeled or rule-based. For example, for the diagnosis of "enterocutaneous fistula," the regular expression "(no|deny)[^。,]*occurrence" is matched to "no bile leakage, no intestinal fistula occurrence," thus determining that "enterocutaneous fistula" is not a confirmed diagnosis. Furthermore, by classifying the rule, such as if it is a "denial" rule, it can be determined that "enterocutaneous fistula" is not a confirmed diagnosis because it is denied. The supporting text, manually labeled as "The patient's vital signs are currently stable, the abdominal drainage tube shows no bile-like or turbid fluid, and there are no signs of peritonitis in the abdomen, indicating no bile leakage or intestinal fistula occurrence," is used. Therefore, the diagnosis of "enterocutaneous fistula" has the following labels: "Confirmed diagnosis: Not confirmed; Diagnosis type: Denial; Supporting text: "The patient's vital signs are currently stable, the abdominal drainage tube shows no bile-like or turbid fluid, and there are no signs of peritonitis in the abdomen, indicating no bile leakage or intestinal fistula occurrence."

[0136] Step S302: Based on the training medical record text, the diagnosis corresponding to the training medical record text, and the labels of each diagnosis in the training medical record text, generate multiple training corpora, each of which includes input text and expected output text.

[0137] In this application, multiple training corpora can be generated based on the training medical record text, the diagnosis corresponding to the training medical record text, and the tags of each diagnosis in the training medical record text, combined with a preset template. The training corpora can involve different task types, such as single negation tasks, explanatory tasks, extraction tasks, etc. The expression form of the training corpora needs to be as rich as possible, for example, the method of manually annotating thought chains can be used.

[0138] Preset templates include: "{input}\nWhat diagnoses are present in this medical record text?", "What diagnoses are present in the following medical record text\n{input}?", "Is the diagnosis {input}\n{diagnosis} a confirmed diagnosis in this text?", "{input}\nWhy is the diagnosis {diagnosis} not a confirmed diagnosis in the above medical record text?", etc., where input is the medical record text and diagnosis is a specific diagnosis.

[0139] To facilitate understanding, the following examples of training corpora are provided:

[0140] Example 1:

[0141]

[0142] Example 2:

[0143]

[0144] Example 3:

[0145]

[0146]

[0147] Step S303: Input the input text into an existing generative model in the medical field, and fine-tune the existing generative model in the medical field so that the output of the existing generative model in the medical field approaches the expected output text corresponding to the input text.

[0148] In this application, the expected output text can be used as the correct answer. The output of an existing generative model in the medical field is compared with the expected output text. A loss function is designed to measure the difference between the two outputs. The model parameters are continuously updated through a feedback mechanism until the model converges, thus completing the fine-tuning.

[0149] By fine-tuning an existing generative model in the medical field using the embodiments of this application, the generative model will have good diagnostic recall capabilities, thereby laying the foundation for the accuracy of subsequent test results.

[0150] In another embodiment of this application, the specific implementation of step S103, which determines the target candidate write-missing diagnosis from the candidate write-missing diagnoses, will be described, referring to... Figure 4 , Figure 4 This is a flowchart illustrating a method for determining a target candidate write-missing diagnosis from the candidate write-missing diagnoses disclosed in an embodiment of this application. The method may include the following steps:

[0151] Step S401: Based on the candidate missing write diagnosis and the written diagnosis, determine multiple diagnosis pairs, each diagnosis pair including one candidate missing write diagnosis and one written diagnosis.

[0152] As one possible implementation, the candidate missed write diagnosis and the written diagnosis can be combined in pairs to form a diagnosis pair.

[0153] Considering that the same diagnosis may be described in different ways, resulting in multiple diagnoses that actually represent the same diagnosis, in order to avoid repetitive judgments on the same diagnosis, as an implementation method, this application can first perform normalization processing on the candidate missing diagnoses to obtain normalized candidate missing diagnoses, and then perform normalization processing on the written diagnoses to obtain normalized written diagnoses. Based on the normalized candidate missing diagnoses and the normalized written diagnoses, multiple diagnosis pairs are then determined.

[0154] As another possible implementation, the candidate missing diagnosis and the written diagnosis can be combined in pairs to form candidate diagnosis pairs. Each candidate diagnosis pair includes one candidate missing diagnosis and one written diagnosis. Then, each candidate diagnosis pair is normalized to obtain normalized candidate diagnosis pairs. Finally, the normalized candidate diagnosis pairs are deduplicated to determine multiple diagnosis pairs.

[0155] It should be noted that the standardization process can be vectorization of the diagnostic text, or standardization can be performed by obtaining key elements such as the dominant words, location, cause, and nature of the diagnosis through a knowledge graph.

[0156] For ease of understanding, assume that the candidate omitted diagnoses include [diagnosis1, diagnosis2, ..., diagnosis] n The diagnoses already written include [diagnosisOrigin1, diagnosisOrigin2, ..., diagnosisOrigin] n If the diagnosis pair is [diagnosis1, diagnosisOrigin1], then the diagnosis pair is an example of [diagnosis1, diagnosisOrigin1].

[0157] Step S402: For each diagnostic pair, determine whether the semantics of the two diagnoses in the diagnostic pair are consistent. If the semantics of the two diagnoses in the diagnostic pair are consistent, proceed to step S403. If the semantics of the two diagnoses in the diagnostic pair are inconsistent, proceed to step S404.

[0158] In this application, the semantic similarity between the candidate missing diagnosis and the written diagnosis in the diagnosis pair can be calculated, thereby determining whether the two diagnoses in the diagnosis pair are semantically consistent.

[0159] Step S403: Determine the candidate missing write diagnosis in the diagnosis pair as the target candidate missing write diagnosis.

[0160] Step S404: Compare the diagnosis pair with each diagnosis symptom pair in the preset set of diagnosis symptom pairs. If the diagnosis pair matches a certain diagnosis symptom pair in the set of diagnosis symptom pairs, then determine the candidate missing diagnosis as the target candidate missing diagnosis. If the diagnosis pair does not match any of the diagnosis symptom pairs in the set of diagnosis symptom pairs, then determine the candidate missing diagnosis as a non-target candidate missing diagnosis.

[0161] In this application, the predefined set of diagnostic symptom pairs includes multiple diagnostic symptom pairs. Each diagnostic symptom pair contains two diagnoses, where one diagnosis is a symptom corresponding to the other diagnosis. For ease of understanding, let's assume a diagnostic symptom pair is [diagnosis]. a ,diagnosis b ], then diagnosis a For diagnosis b Symptoms, or, diagnosis b For diagnosis a Symptoms.

[0162] The preset set of diagnostic symptoms is determined based on a large amount of medical record data, and the specific determination method will be explained in detail in the following embodiments.

[0163] In another embodiment of this application, the method for determining the preset set of diagnostic symptoms is described in detail. (Refer to...) Figure 5 , Figure 5 This is a flowchart illustrating a method for determining a preset set of diagnostic symptom pairs disclosed in an embodiment of this application. The method may include the following steps:

[0164] Step S501: Obtain the reference medical record text, the candidate missing diagnosis corresponding to the reference medical record text, and the written diagnosis corresponding to the reference medical record text.

[0165] The method for determining the reference medical record text can be to first collect original medical record texts from different sources and departments, preprocess the collected original medical record texts to obtain preprocessed medical record texts, which will be used as reference medical record texts. The preprocessing method can be referred to the relevant descriptions in the foregoing embodiments, and will not be repeated here. The candidate omitted diagnoses corresponding to the reference medical record texts can be determined manually or using existing technology; this application does not impose any limitations on this. The method for determining the written diagnoses corresponding to the reference medical record texts can be referred to the relevant descriptions in the foregoing embodiments, and will not be repeated here.

[0166] Step S502: Combine the written diagnoses corresponding to the reference medical record text in pairs to obtain a first set of diagnosis pairs; combine the candidate missing diagnoses corresponding to the reference medical record text and the written diagnoses corresponding to the reference medical record text in pairs; and combine the candidate missing diagnoses corresponding to the reference medical record text in pairs to obtain a second set of diagnosis pairs.

[0167] It should be noted that the first diagnostic pair set includes multiple first diagnostic pairs, and the second diagnostic pair set includes multiple second diagnostic pairs. Each first diagnostic pair includes two written diagnostics, and each second diagnostic pair includes one candidate missed diagnostic and one written diagnostic, or includes two candidate missed diagnostics.

[0168] Step S503: Calculate the co-occurrence probability of each first diagnostic pair in the first diagnostic pair set.

[0169] In this application, the ratio between the number of times the first diagnostic pair appears in all reference medical record texts and the total number of times each diagnostic pair in the first diagnostic pair set appears in all reference medical record texts can be calculated as the co-occurrence probability of the first diagnostic pair.

[0170] Step S504: Based on the co-occurrence probability of each first diagnostic pair in the first diagnostic pair set, determine the diagnostic symptom pair set from the first diagnostic pair set and the second diagnostic pair set.

[0171] The higher the co-occurrence ratio of a diagnostic pair, the greater the co-occurrence probability, the more frequent the co-occurrence, and the more prevalent the existence of the diagnostic pair. This suggests that the pair likely represents two diseases commonly occurring in the same patient, or a common disease and its complication. In this case, one disease is less likely to be a symptom of the other. Conversely, a lower co-occurrence ratio indicates a lower probability and fewer co-occurrences, meaning that doctors would not write both diagnoses simultaneously in their clinical practice. One disease is more likely to be a symptom of the other. Therefore, in this application, it is necessary to exclude the first diagnostic pair with a higher co-occurrence probability, as well as the second diagnostic pair that is semantically consistent with the first diagnostic pair with a higher co-occurrence probability.

[0172] As one possible implementation, a first threshold and a second threshold can be preset, with the first threshold being greater than the second threshold. The specific values ​​of the first and second thresholds can be set according to the scenario requirements, and this application does not impose any limitations. Then, a target second diagnostic pair can be determined from the second diagnostic pair set. The target second diagnostic pair is a second diagnostic pair semantically consistent with the first diagnostic pair whose co-occurrence probability is greater than the first preset threshold. The first diagnostic pairs in the first diagnostic pair set whose co-occurrence probability is less than the second preset threshold, and the second diagnostic pairs in the second diagnostic pair set other than the target second diagnostic pair, are combined to obtain a candidate diagnostic symptom pair set.

[0173] As one possible implementation, the set of candidate diagnostic symptom pairs can be used as the diagnostic symptom pair set. Considering the lack of standardization in the descriptions of each diagnosis within the candidate diagnostic symptom pair set, as another possible implementation, the candidate diagnostic symptom pair set can be labeled to determine the diagnostic symptom pair set.

[0174] In this application, manual full-scale annotation can be used, or manual sampling and annotation can be performed first, and then diagnostic pairs with similar types can be classified by diagnostic pair similarity or other clustering methods to achieve batch annotation of diagnostic pairs (for example, diagnostic pairs composed of renal function damage with hypocalcemia and hypochloremia can be clustered into one class, and only one pair needs to be annotated), thereby reducing the workload of annotation.

[0175] The medical record diagnosis omission detection device disclosed in the embodiments of this application is described below. The medical record diagnosis omission detection device described below can be referred to in correspondence with the medical record diagnosis omission detection method described above.

[0176] Reference Figure 6 , Figure 6 This is a schematic diagram of a medical record diagnostic omission detection device disclosed in an embodiment of this application. Figure 6 As shown, the medical record diagnosis omission detection device may include:

[0177] The medical record text determination unit 11 is used to determine the medical record text to be detected;

[0178] Candidate omission diagnosis determination unit 12 is used to determine the candidate omission diagnosis corresponding to the medical record text;

[0179] The target candidate omission diagnosis determination unit 13 is used to determine a target candidate omission diagnosis from the candidate omission diagnoses, wherein the target candidate omission diagnosis is a candidate omission diagnosis that is semantically consistent with the written diagnosis in the medical record text, and a candidate omission diagnosis that is determined to be a symptom of the written diagnosis.

[0180] The missing write diagnosis determination unit 14 is used to determine other candidate missing write diagnoses besides the target candidate missing write diagnosis as missing write diagnoses.

[0181] As one possible implementation, the candidate missing write diagnosis and determination unit includes:

[0182] The prompt information generation subunit is used to generate prompt information corresponding to the medical record text, and the prompt information is used to indicate the candidate missing diagnosis corresponding to the medical record text.

[0183] The generative model calling subunit is used to call the generative model and input the prompt information corresponding to the medical record text into the generative model;

[0184] The candidate omission diagnosis determination subunit is used to determine the candidate omission diagnosis corresponding to the medical record text based on the response of the generative model.

[0185] As one possible implementation, the prompt information generation subunit is specifically used for:

[0186] The medical record text is segmented to obtain multiple medical record text fragments;

[0187] For each medical record text segment, a prompt message corresponding to the medical record text segment is generated. The prompt message is used to indicate the candidate missing diagnosis corresponding to the medical record text segment. The combination of the prompt messages corresponding to each medical record text segment is the prompt message corresponding to the medical record text.

[0188] As one possible implementation, the generative model is obtained by fine-tuning an existing generative model in the medical field; the device includes a fine-tuning unit, which is specifically used for:

[0189] The training medical record text, the diagnosis corresponding to the training medical record text, and the tags of each diagnosis in the training medical record text are obtained. The tags of each diagnosis in the training medical record text include a first tag, a second tag, and a third tag. The first tag is used to indicate whether the diagnosis is confirmed, the second tag is used to indicate the diagnosis type, and the third tag is used to indicate the recall support fragment of the diagnosis.

[0190] Based on the training medical record text, the diagnosis corresponding to the training medical record text, and the labels of each diagnosis in the training medical record text, multiple training corpora are generated, each training corpus including input text and expected output text.

[0191] The input text is input into an existing generative model in the medical field. The goal is to make the output of the existing generative model in the medical field approach the expected output text corresponding to the input text. The existing generative model in the medical field is then fine-tuned.

[0192] As one possible implementation, the target candidate missing write diagnosis and determination unit includes:

[0193] The diagnostic pair determination unit is used to determine multiple diagnostic pairs based on the candidate missing write diagnosis and the written diagnosis, wherein each diagnostic pair includes one candidate missing write diagnosis and one written diagnosis;

[0194] The judgment unit is used to determine, for each diagnosis pair, whether the semantics of the two diagnoses in the diagnosis pair are consistent;

[0195] The first processing unit is configured to determine the candidate missing diagnosis in the diagnosis pair as the target candidate missing diagnosis if the two diagnoses in the diagnosis pair have the same semantics.

[0196] The comparison unit is used to compare the diagnosis pair with each diagnosis symptom pair in a preset set of diagnosis symptom pairs if the semantics of the two diagnoses in the diagnosis pair are inconsistent.

[0197] The second processing unit is configured to determine the candidate missed diagnosis as the target candidate missed diagnosis if the diagnosis pair matches a certain diagnosis symptom pair in the set of diagnosis symptom pairs.

[0198] As one possible implementation, the diagnostic determination unit is specifically used for:

[0199] The candidate missing write diagnoses are normalized to obtain normalized candidate missing write diagnoses.

[0200] The written diagnosis is normalized to obtain the normalized written diagnosis;

[0201] Based on the normalized candidate missed write diagnoses and the normalized written diagnoses, the multiple diagnosis pairs are determined.

[0202] As one possible implementation, the diagnostic determination unit is specifically used for:

[0203] The candidate missed write diagnosis and the written diagnosis are combined in pairs to form candidate diagnosis pairs; each candidate diagnosis pair includes one candidate missed write diagnosis and one written diagnosis.

[0204] Each of the candidate diagnostic pairs is normalized to obtain normalized candidate diagnostic pairs;

[0205] The candidate diagnostic pairs after normalization are deduplicated to determine the plurality of diagnostic pairs.

[0206] As one possible implementation, the device further includes a diagnostic symptom pair determination unit, which is specifically used for:

[0207] Obtain the reference medical record text, the candidate missed diagnoses corresponding to the reference medical record text, and the written diagnoses corresponding to the reference medical record text;

[0208] The written diagnoses corresponding to the reference medical record text are combined in pairs to obtain a first set of diagnosis pairs. The candidate missing diagnoses corresponding to the reference medical record text and the written diagnoses corresponding to the reference medical record text are combined in pairs to obtain a second set of diagnosis pairs.

[0209] Calculate the co-occurrence probability of each of the first diagnostic pairs in the first diagnostic pair set;

[0210] Based on the co-occurrence probability of each of the first diagnostic pairs in the first diagnostic pair set, the set of diagnostic symptom pairs is determined from the first diagnostic pair set and the second diagnostic pair set.

[0211] Reference Figure 7 , Figure 7 A hardware structure block diagram of a medical record diagnostic omission detection device provided in this application embodiment is shown below. Figure 7 The hardware structure of the medical record diagnosis omission detection device may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4;

[0212] In this embodiment, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4.

[0213] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0214] Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0215] The memory stores a program, which the processor can call. The program is used for:

[0216] Identify the medical record text to be tested;

[0217] Identify the candidate missing diagnoses corresponding to the medical record text;

[0218] From the candidate omission diagnoses, a target candidate omission diagnosis is determined, wherein the target candidate omission diagnosis is a candidate omission diagnosis that is semantically consistent with the written diagnosis in the medical record text, and a candidate omission diagnosis that is determined to be a symptom of the written diagnosis.

[0219] The candidate write-missing diagnoses other than the target candidate write-missing diagnoses are identified as write-missing diagnoses.

[0220] Optionally, the refined and extended functions of the program can be found in the description above.

[0221] This application embodiment also provides a readable storage medium that can store a program suitable for execution by a processor, the program being used for:

[0222] Identify the medical record text to be tested;

[0223] Identify the candidate missing diagnoses corresponding to the medical record text;

[0224] From the candidate omission diagnoses, a target candidate omission diagnosis is determined, wherein the target candidate omission diagnosis is a candidate omission diagnosis that is semantically consistent with the written diagnosis in the medical record text, and a candidate omission diagnosis that is determined to be a symptom of the written diagnosis.

[0225] The candidate write-missing diagnoses other than the target candidate write-missing diagnoses are identified as write-missing diagnoses.

[0226] Optionally, the refined and extended functions of the program can be found in the description above.

[0227] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0228] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0229] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting omissions in medical record diagnoses, characterized in that, The method includes: Identify the medical record text to be tested; Identify the candidate missing diagnoses corresponding to the medical record text; From the candidate omission diagnoses, a target candidate omission diagnosis is determined, wherein the target candidate omission diagnosis is a candidate omission diagnosis that is semantically consistent with the written diagnosis in the medical record text, and a candidate omission diagnosis that is determined to be a symptom of the written diagnosis. The candidate write-missing diagnoses other than the target candidate write-missing diagnoses are identified as write-missing diagnoses.

2. The method according to claim 1, characterized in that, The process of determining the candidate missing diagnosis corresponding to the medical record text includes: Generate prompt information corresponding to the medical record text, the prompt information being used to indicate the candidate missing diagnosis corresponding to the medical record text; Invoke the generative model and input the prompt information corresponding to the medical record text into the generative model; Based on the response from the generative model, candidate omission diagnoses corresponding to the medical record text are determined.

3. The method according to claim 2, characterized in that, The prompt information corresponding to the generated medical record text includes: The medical record text is segmented to obtain multiple medical record text fragments; For each medical record text segment, a prompt message corresponding to the medical record text segment is generated. The prompt message is used to indicate the candidate missing diagnosis corresponding to the medical record text segment. The combination of the prompt messages corresponding to each medical record text segment is the prompt message corresponding to the medical record text.

4. The method according to claim 2, characterized in that, The generative model is obtained by fine-tuning an existing generative model in the medical field; the fine-tuning of the existing generative model in the medical field includes: The training medical record text, the diagnosis corresponding to the training medical record text, and the tags of each diagnosis in the training medical record text are obtained. The tags of each diagnosis in the training medical record text include a first tag, a second tag, and a third tag. The first tag is used to indicate whether the diagnosis is confirmed, the second tag is used to indicate the diagnosis type, and the third tag is used to indicate the recall support fragment of the diagnosis. Based on the training medical record text, the diagnosis corresponding to the training medical record text, and the labels of each diagnosis in the training medical record text, multiple training corpora are generated, each training corpus including input text and expected output text. The input text is input into an existing generative model in the medical field. The goal is to make the output of the existing generative model in the medical field approach the expected output text corresponding to the input text, and then fine-tune the existing generative model in the medical field.

5. The method according to claim 1, characterized in that, The step of determining the target candidate write-missing diagnosis from the candidate write-missing diagnoses includes: Based on the candidate missing write diagnosis and the written diagnosis, multiple diagnosis pairs are determined, each diagnosis pair including one candidate missing write diagnosis and one written diagnosis; For each diagnostic pair, determine whether the semantics of the two diagnoses in the diagnostic pair are consistent; If the two diagnoses in the diagnosis pair have the same semantics, then the candidate missing diagnosis in the diagnosis pair is determined as the target candidate missing diagnosis. If the two diagnoses in the diagnosis pair are semantically inconsistent, the diagnosis pair is compared with each diagnosis symptom pair in the preset set of diagnosis symptom pairs. If the diagnosis pair matches a diagnosis symptom pair in the set of diagnosis symptom pairs, then the candidate missed diagnosis is determined as the target candidate missed diagnosis.

6. The method according to claim 5, characterized in that, The process of determining multiple diagnostic pairs based on the candidate missed write diagnoses and the written diagnoses includes: The candidate missing write diagnoses are normalized to obtain normalized candidate missing write diagnoses. The written diagnosis is normalized to obtain the normalized written diagnosis; Based on the normalized candidate missed write diagnoses and the normalized written diagnoses, the multiple diagnosis pairs are determined.

7. The method according to claim 5, characterized in that, The process of determining multiple diagnostic pairs based on the candidate missed write diagnoses and the written diagnoses includes: The candidate missed write diagnosis and the written diagnosis are combined in pairs to form candidate diagnosis pairs; each candidate diagnosis pair includes one candidate missed write diagnosis and one written diagnosis. Each of the candidate diagnostic pairs is normalized to obtain normalized candidate diagnostic pairs; The candidate diagnostic pairs after normalization are deduplicated to determine the plurality of diagnostic pairs.

8. The method according to claim 5, characterized in that, The methods for determining the preset diagnostic symptom set include: Obtain the reference medical record text, the candidate missed diagnoses corresponding to the reference medical record text, and the written diagnoses corresponding to the reference medical record text; The written diagnoses corresponding to the reference medical record text are combined in pairs to obtain a first set of diagnosis pairs. The candidate missing diagnoses corresponding to the reference medical record text and the written diagnoses corresponding to the reference medical record text are combined in pairs to obtain a second set of diagnosis pairs. Calculate the co-occurrence probability of each of the first diagnostic pairs in the first diagnostic pair set; Based on the co-occurrence probability of each of the first diagnostic pairs in the first diagnostic pair set, the set of diagnostic symptom pairs is determined from the first diagnostic pair set and the second diagnostic pair set.

9. A medical record diagnostic omission detection device, characterized in that, The device includes: The medical record text determination unit is used to determine the medical record text to be detected. A candidate omission diagnosis determination unit is used to determine the candidate omission diagnoses corresponding to the medical record text. A target candidate omission diagnosis determination unit is used to determine a target candidate omission diagnosis from the candidate omission diagnoses, wherein the target candidate omission diagnosis is a candidate omission diagnosis that is semantically consistent with the written diagnosis in the medical record text, and a candidate omission diagnosis that is determined to be a symptom of the written diagnosis. The missing write diagnosis determination unit is used to determine other candidate missing write diagnoses besides the target candidate missing write diagnosis as missing write diagnoses.

10. A medical record diagnostic omission detection device, characterized in that, Including memory and processor; The memory is used to store programs; The processor is used to execute the program to implement each step of the medical record diagnosis omission detection method as described in any one of claims 1 to 8.

11. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step of the medical record diagnosis omission detection method as described in any one of claims 1 to 8.

Images