Medical data structured processing method and system
By building a speech feature model and a transliteration error check model, combined with the medical knowledge base, the entry error and modular problems in traditional medical record recording methods are solved, efficient and accurate structured medical record data is achieved, and the availability of medical information is improved and archive research efficiency.
Patent Information
- Application Number
- CN202411126122.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-08-16
AI Technical Summary
Traditional medical record recording methods rely on manual writing or keyboard input, resulting in time consumption and entry errors. In addition, intelligent voice transcription technology is prone to snippeting and sentence breaking when processing proprietary vocabulary, and cannot form systematic and modular recording files.
Design a structured medical data processing method, obtain doctor's voice recording data, eliminate interference sounds, build a speech feature model and a transliteration error check model, combine the medical knowledge base to match speech transcription checksum terms, and generate accurate medical records. Text information.
It realizes rapid and accurate screening of medical records and entering voice data, improves the recognition and accuracy of medical record transcription, forms systematic and modular medical record records, and improves the availability of medical information and the efficiency of archive research.
Smart Images

Figure CN119028505A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical information processing, and in particular to a method and system for structured processing of medical data. Background Art
[0002] The traditional way of recording medical records relies on doctors to write or type, which takes a lot of time and energy. At the same time, manual writing or keyboard input is prone to input errors, affecting the accuracy of medical record information.
[0003] In the context of the rapid development of medical information technology today, intelligent speech transcription technology is gradually attracting the attention and application of the medical community. This technology greatly simplifies the process of medical record recording and improves the accuracy and availability of medical information by converting the doctor's oral statements into text form.
[0004] In the prior art, the doctor's oral information is generally transcribed directly into text through a voice transcription engine. When some proprietary vocabulary is involved, it is impossible to intelligently identify voice transcription errors, which easily leads to word and sentence breaks. Moreover, although intelligent transcription improves recording efficiency, the text records formed are colloquial and cannot be systematic and modularized, which is not conducive to hospital archiving and research.
[0005] Therefore, there is an urgent need to design a medical data structured processing method and system to solve the above technical problems. Summary of the invention
[0006] In order to solve the above-mentioned deficiencies in the prior art, the purpose of the present invention is to overcome the existing deficiencies and provide a method for structuring medical data, which comprises the following steps:
[0007] Acquire the voice recording data of the doctor entering the medical record; remove the interference sound of the voice recording data and save the medical record voice data;
[0008] Obtain the medical record voice data entered by the doctor, and transcribe the medical record voice data into corresponding medical record text information;
[0009] Construct a medical record speech transcription error checking model to perform medical record speech transcription error checking based on the transcribed medical record text information:
[0010] If the evaluation result based on the medical record speech transcription error checking model is normal, output the corresponding transcribed medical record text information;
[0011] If the evaluation result based on the medical record voice transcription error checking model is abnormal, re-transcribe the corresponding medical record voice data into medical record text information;
[0012] A medical record knowledge base is constructed, the medical record text information with a normal evaluation result is matched with the medical record knowledge base, and the matched medical terminology medical record entry information based on the medical record knowledge base is output.
[0013] As a further optimization of the above solution, the method for removing the interfering sound of the voice recording data includes the following:
[0014] Construct a voice feature model for obtaining medical record entry by doctors, wherein the influencing factor of the model is the single-segment impulse ratio corresponding to the medical record entry time and any unit time;
[0015] Mark any medical record recording voice acquisition start point to record the mark point, based on the mark point according to the voice acquisition frequency f, until any medical record recording voice ends to obtain the mark point number H, then the medical record recording time t is:
[0016]
[0017] The total impulse C of any medical record recording speech is marked, and is divided into x segments according to the speech acquisition frequency f based on the marked point. The impulse of any single segment of medical record recording speech divided into x segments is C i (i=1,2,…,x), the impulse ratio of any single segment of medical record recording speech divided into x segments is ε i (i=1,2,…,x), the single-stage impulse ratio calculation method is as follows:
[0018]
[0019] As a further optimization of the above solution, the method further includes:
[0020] Based on the constructed medical record entry time t and single-segment impulse ratio ε i , detect all collected recorded sounds:
[0021] If the detected single-segment impulse ratio first rises to the maximum value and then decreases, the corresponding recorded sound is retained; otherwise, all collected recorded sounds are deleted and the above steps are repeated;
[0022] Based on the retained recorded sound that first rises to a maximum value and then decreases, the sound corresponding to the maximum value of the medical record recording time t is retained as the target sound.
[0023] As a further optimization of the above scheme, the method for checking the medical record voice transcription error based on the transcribed medical record text information specifically includes the following:
[0024] Construct an arbitrary medical record text information distribution model L(b) as follows:
[0025]
[0026] Among them, L(b i |b i-p+1 ,…,b i-1 ) is the calculation object of the medical record text information distribution model L(b), represented by the i-th medical record entry word b i The distribution of p-1 is represented by the current medical record entry word b i The first p-1 medical record entry words;
[0027] Construct an arbitrary medical record text information fault-tolerant model T(b),
[0028]
[0029] Among them, K(b i-1 b i ) represents the number of times the corresponding medical record entry word appears in the medical record knowledge base;
[0030] Based on the constructed medical record text information distribution model L(b) and medical record text information fault tolerance model T(b), the next predicted word accuracy probability of any medical record input word in the medical record text information is calculated: where ω is the regularization constant, λ is the stability coefficient, and b i-1 b i Represents any combination of two medical record entry words;
[0031]
[0032] As a further optimization of the above scheme, the method for checking the medical record voice transcription error based on the transcribed medical record text information also includes the following:
[0033] Assuming that the evaluation index of the evaluation result based on the medical record speech transcription error checking model is Right, we have:
[0034]
[0035] in, It is represented as the current entry probability description of any medical record entry word, and Γ is the description weight.
[0036] As a further optimization of the above solution, the method for matching the medical record text information with a normal evaluation result with the medical record knowledge base specifically includes the following:
[0037] Generate any case text information that has been checked by medical record voice transcription, and mark all medical record input words that have been identified and detected as b i , then the set of case text information and medical record entry words is expressed as B = (b1, b2, b3…b m );
[0038] Based on the generated set B, a deep neural guided convolutional model Tend is constructed:
[0039] Tend=aggcn(B,B) (7)
[0040] Construct a mapping association Z between case text information and medical terms. The mapping association Z is constructed as follows:
[0041]
[0042]
[0043] Among them, the mapping association Z1 is the corresponding relationship between the jump degree Q1 as the first jump and the jump degree Q2 as the second jump, the mapping association Z2 is the corresponding relationship between the jump degree Q2 as the first jump and the jump degree Q1 as the second jump, and ∫(Q1,Q2) is expressed as the dynamic association of Q1 and Q2.
[0044] As a further optimization of the above solution, after completing the construction of the mapping association Z, the method further includes the following:
[0045] Obtain a mapping association Z, and input the mapping association Z into the medical term conversion set:
[0046]
[0047] in, is a sigmoid function, parZ is the input parameter matrix based on the medical record input words, Z 12 is the dot product of the mapping association Z1 and Z2, is the mapping associated offset vector.
[0048] As a further optimization of the above solution, the constructed model includes a starting point for entering medical record text information, relationship conversion, and an entry end point.
[0049] The present invention also discloses a medical data structured processing method system, the system comprising the following:
[0050] A voice filtering module is used to obtain the voice recording data of the doctor entering the medical record; remove the interference sound of the voice recording data and save the medical record voice data;
[0051] A voice transcription module is used to obtain the medical record voice data entered by the doctor and transcribe the medical record voice data into corresponding medical record text information;
[0052] The voice verification module is used to build a medical record voice transcription error checking model and perform medical record voice transcription error checking based on the transcribed medical record text information:
[0053] If the evaluation result based on the medical record speech transcription error checking model is normal, output the corresponding transcribed medical record text information;
[0054] If the evaluation result based on the medical record voice transcription error checking model is abnormal, re-transcribe the corresponding medical record voice data into medical record text information;
[0055] The transcription redirection module is used to build a medical record knowledge base, match the medical record text information with a normal evaluation result with the medical record knowledge base, and output the matched medical terminology medical record entry information based on the medical record knowledge base.
[0056] As a further optimization of the above solution, the voice filtering module includes the following:
[0057] A model building unit, used to build a voice feature model for obtaining a doctor's medical record entry, wherein the influencing factor of the model is a single-segment impulse ratio corresponding to the medical record entry time and any unit time;
[0058] The speech marking unit is used to mark the start point of any medical record recording speech acquisition and record the marking point. Based on the marking point, the speech acquisition frequency f is obtained until any medical record recording speech ends to obtain the marking point number H. The medical record recording time t is:
[0059]
[0060] The impulse calculation unit is used to mark the total impulse C of any medical record recording voice, and divide it into x segments according to the voice acquisition frequency f based on the marking point. The impulse of any single segment of medical record recording voice divided into x segments is Ci (i = 1, 2, ..., x), and the single segment impulse ratio of any single segment of medical record recording voice divided into x segments is ε i (i=1,2,…,x), the single-stage impulse ratio calculation method is as follows:
[0061]
[0062] The present invention adopts the above technical solution. Compared with the prior art, a medical data structured processing method and system have the following technical effects:
[0063] 1. The present invention designs a structured processing method for medical data, specifically by constructing a voice feature model for medical record entry, calculating the total impulse of any medical record entry time and medical record entry voice to calculate the single-segment impulse ratio, and based on the line graph of the single-segment impulse ratio, quickly and accurately screen out the required medical record entry voice data information.
[0064] 2. The present invention proposes a method for calculating the accuracy probability of the next predicted word of any medical record entry word in medical record text information. When the number of medical record entry words is 1, the accuracy probability of the next predicted word of any medical record entry word in the described medical record text information is the calculation result of any medical record text information fault tolerance model; when the number of medical record entry words is greater than 1, the calculation method designed based on the present invention can quickly obtain the corresponding occurrence probability. The present invention is designed to solve the problem that doctors record the patient's condition by voice transcription to form a medical record voice entry, and specifically designs a fault tolerance mechanism for medical record voice transcription, that is, by transcribing the medical record entry words from voice into text, and predicting the next medical record entry word for voice transcription based on any medical record entry words that have been transcribed, thereby forming medical record voice transcription data information with high recognition and more in line with the actual doctor's oral statement.
[0065] 3. The evaluation index designed by the present invention based on the evaluation results of the medical record voice transcription error checking model can further provide feedback on the accuracy of the medical record voice transcription, intuitively express the medical record information that the doctor actually needs to transcribe, and improve work efficiency.
[0066] 4. The present invention creates a medical data structured processing method, which specifically radiates with the medical record status of actual symptoms as the core. Based on the established medical record knowledge base, it can quickly help doctors match and associate the attribute medical terms of actual symptoms, so that doctors can quickly and accurately complete the actual situation matching during the medical record status transcription process, with good results. BRIEF DESCRIPTION OF THE DRAWINGS
[0067] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:
[0068] Figure 1 It is a schematic diagram of the process of the present invention;
[0069] Figure 2 A schematic diagram of an embodiment of the present invention;
[0070] Figure 3 It is a schematic diagram of the transcription process of the present invention. DETAILED DESCRIPTION
[0071] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.
[0072] like Figure 1-2 As shown, an embodiment of the present invention discloses a method for structuring medical data, which includes the following steps:
[0073] Obtain the voice recording data of the doctor entering the medical record; remove the interference sound of the voice recording data and save the medical record voice data;
[0074] Obtain the medical record voice data entered by the doctor and transcribe the medical record voice data into corresponding medical record text information;
[0075] Construct a medical record speech transcription error checking model to perform medical record speech transcription error checking based on the transcribed medical record text information:
[0076] If the evaluation result of the medical record speech transcription error checking model is normal, the corresponding transcribed medical record text information is output;
[0077] If the evaluation result of the medical record voice transcription error checking model is abnormal, re-transcribe the corresponding medical record voice data into medical record text information;
[0078] Build a medical record knowledge base, match the medical record text information with normal evaluation results with the medical record knowledge base, and output the matched medical terminology medical record entry information based on the medical record knowledge base.
[0079] More specifically, after the present invention completes the output of medical terminology medical record information that is transcribed from the medical record voice data entered by the doctor after matching based on the medical record knowledge base, it is transmitted to an electronic touch device that supports interactivity. The doctor can match the required medical terms through point-and-click operations or modify them manually. This operation is based on existing conventional technology and will not be described in detail here. Those skilled in the art can quickly understand and master this operation based on the present technical solution.
[0080] Specifically, the method for removing the interfering sound of the voice recording data includes the following:
[0081] Construct a voice feature model for doctors to enter medical records. The influencing factor of the model is the single-segment impulse ratio corresponding to the medical record entry time and any unit time.
[0082] Mark any medical record recording voice acquisition start point and record the mark point. Based on the mark point, obtain the voice acquisition frequency f until any medical record recording voice ends and obtain the mark point number H. Then the medical record recording time t is:
[0083]
[0084] The total impulse C of any medical record recording speech is marked, and it is divided into x segments according to the speech acquisition frequency f based on the marked point. The impulse of any single medical record recording speech divided into x segments is C i (i=1,2,…,x), the impulse ratio of any single segment of medical record recording speech divided into x segments is ε i (i=1,2,…,x), the single-stage impulse ratio calculation method is as follows:
[0085]
[0086] Specifically, the above method also includes:
[0087] Based on the constructed medical record entry time t and single-segment impulse ratio ε i , detect all collected recorded sounds:
[0088] If the detected single-segment impulse ratio first rises to the maximum value and then decreases, the corresponding recorded sound is retained; otherwise, all collected recorded sounds are deleted and the above steps are repeated;
[0089] Based on the retained recorded sound that first rises to a maximum value and then decreases, the sound corresponding to the maximum value of the medical record recording time t is retained as the target sound.
[0090] More specifically, the present invention provides an embodiment in which doctors in hospitals summarize medical records based on the symptoms of patients and complete automatic input into the system. In this process, the voice input of the medical records summarized by doctors based on the patients is stable, that is, the single-segment impulse ratio has a concentrated time period, and its strength attenuation is a change that first rises and then falls. Based on the single-segment impulse ratio calculation method and the establishment of a line graph of time and single-segment impulse ratio, the sound that meets the concentrated time period and the strength attenuation that first rises and then falls can be quickly extracted from other interfering background sounds in hospitals;
[0091] It should be noted that when the patient is describing his or her symptoms, there are local sound characteristics that also satisfy the trend of having a concentrated time period and a strong and weak attenuation degree that first rises and then falls. Therefore, in order to more accurately isolate the patient's voice, it is necessary to further judge the medical record entry time t and obtain the maximum value of the medical record entry time t, that is, the corresponding sound is the doctor's continuous entry of the patient's medical record voice data information.
[0092] The present invention designs a structured processing method for medical data, specifically by constructing a voice feature model for medical record entry, calculating the total impulse of any medical record entry time and the medical record entry voice to calculate the single-segment impulse ratio, and based on a line graph of the single-segment impulse ratio, quickly and accurately screening out the required medical record entry voice data information.
[0093] Specifically, the method for checking the medical record voice transcription error based on the transcribed medical record text information specifically includes the following:
[0094] Construct an arbitrary medical record text information distribution model L(b) as follows:
[0095]
[0096] Among them, L(b i |b i-p+1,…,b i-1 ) is the calculation object of the medical record text information distribution model L(b), which is represented by the i-th medical record entry word b i The distribution of p-1 is represented by the current medical record entry word b i The first p-1 medical record entry words;
[0097] More specifically, p of the present invention is represented by the number of medical record entry items. When the number of medical record entry items is large, it means that the possibility of the medical record text information distribution model L(b) being used to describe the case voice transcription output becomes lower. Furthermore, as the number of medical record entry items increases, the parameters it contains increase, and the computational complexity of the medical record text information distribution model L(b) increases.
[0098] Construct an arbitrary medical record text information fault-tolerant model T(b),
[0099]
[0100] Among them, K(b i-1 b i ) represents the number of times the corresponding medical record entry word appears in the medical record knowledge base;
[0101] More specifically, the design of the present invention also provides a fault-tolerant model T(b) for arbitrary medical record text information, that is, when the medical record entry information is entered into a larger scale process, the result calculated based on the arbitrary medical record text information distribution model L(b) approaches the actual situation; but due to the increased complexity, there are medical record entry words that have not been recorded, that is, the theoretical probability of occurrence of the medical record entry word is 0.
[0102] Based on the above-described generation of medical record entry words that have not been recorded, that is, the actual occurrence of any medical record entry word is not zero, the present invention specifically designs a fault-tolerant model T(b) for arbitrary medical record text information, and designs the occurrence of the phonetic transcription of any medical record entry word as a possible probability.
[0103] Specifically, based on the constructed medical record text information distribution model L(b) and medical record text information fault tolerance model T(b), the next predicted word accuracy probability of any medical record entry word of the medical record text information is calculated: where ω is a regularization constant, and 0<ω<1; λ is a stability coefficient, so that the total probability of transcription prediction of this case entry remains 1; b i-1 b i Represents any combination of two medical record entry words; it should be noted that It is aimed at is the number of times all possible combinations of prefix and b appear, It is aimed at Prefix, including b i The number of times the possible combination of words appears, It is aimed at The number of times the possible combination of the prefix appears; b: For All different combinations of b under the condition that b is the antecedent prefix, i : For As the antecedent condition contains b i Different combinations of
[0104]
[0105] More specifically, the present invention proposes a method for calculating the accuracy probability of the next predicted word of any medical record entry word in medical record text information. When the number of medical record entry words is 1, the accuracy probability of the next predicted word of any medical record entry word in the described medical record text information is the calculation result of any medical record text information fault tolerance model T(b); when the number of medical record entry words is greater than 1, the calculation method designed based on the present invention can quickly obtain the corresponding occurrence probability. The present invention is designed to solve the problem that doctors record the patient's condition by voice transcription to form a medical record voice entry, and specifically designs a fault tolerance mechanism for medical record voice transcription, that is, by transcribing the medical record entry words from voice into text, and predicting the next medical record entry word for voice transcription based on any medical record entry words that have been transcribed, thereby forming medical record voice transcription data information with high recognition and more in line with the actual doctor's oral statement.
[0106] Specifically, the method for checking the medical record voice transcription error based on the transcribed medical record text information also includes the following:
[0107] Assuming that the evaluation index of the evaluation result based on the medical record speech transcription error checking model is Right, we have:
[0108]
[0109] in, It is represented as the current entry probability description of any medical record entry word, Γ is the description weight, when the probability description value is greater than 0.26, Γ takes the value of 1, otherwise Γ takes the value of 0.
[0110] More specifically, in order to more concretely express the error checking of medical record voice transcription, the present invention, based on the above, specifically designs an evaluation index based on the evaluation results of the medical record voice transcription error checking model, that is, when the evaluation index of any medical record voice transcription text calculated is greater than 0.65, the corresponding medical record voice transcription sentence is marked as a valid voice transcription, which conforms to the actual description; otherwise, the corresponding medical record voice transcription sentence is marked as an invalid voice transcription, there is a transcription recognition error, and the corresponding sentence is re-transcribed. The evaluation index based on the evaluation results of the medical record voice transcription error checking model designed by the present invention can further feedback the accuracy of the medical record voice transcription, intuitively express the medical record information that the doctor actually wants to transcribe, and improve work efficiency.
[0111] Specifically, the method for matching the medical record text information with a normal evaluation result with the medical record knowledge base includes the following:
[0112] Generate any case text information that has been checked by medical record voice transcription, and mark all medical record input words that have been identified and detected as b i , then the set of case text information and medical record entry words is expressed as B = (b1, b2, b3…b m );
[0113] Based on the generated set B, a deep neural guided convolutional model Tend is constructed:
[0114] Tend=aggcn(B,B) (7)
[0115] Construct a mapping association Z between case text information and medical terms. The mapping association Z is constructed as follows:
[0116]
[0117]
[0118] Among them, the mapping association Z1 is the corresponding relationship between the jump degree Q1 as the first jump and the jump degree Q2 as the second jump, the mapping association Z2 is the corresponding relationship between the jump degree Q2 as the first jump and the jump degree Q1 as the second jump, and ∫(Q1,Q2) is expressed as the dynamic association of Q1 and Q2;
[0119] Get the mapping association Z and input the mapping association Z into the medical term conversion set:
[0120]
[0121] in, is a sigmoid function, parZ is the input parameter matrix based on the medical record input words, Z 12 is the dot product of the mapping association Z1 and Z2, is the mapping associated offset vector.
[0122] Specifically, the constructed model includes the starting point for entering medical record text information, relationship conversion, and the end point for entering.
[0123] Specifically, the technical solution disclosed in the present invention supports converting the transcribed medical terminology medical record entry information into a structured medical record display method, and the specific method is as follows:
[0124] A medical record structure knowledge graph is established in the medical record knowledge base, and the nodes include medical record structure nodes (chief complaint, current medical history, past medical history, allergy history, medication history, physical examination, examination, testing, etc.) and medical terminology nodes; the connection between the medical record structure nodes and the medical terminology nodes is the probability that the medical terminology node belongs to the node (a high probability indicates a high probability of belonging to the node), and the connection between medical terminology nodes is the relationship between term ontologies, and the medical terminology nodes define common attributes.
[0125] The above conversion process is as follows:
[0126] 1. The extracted medical terms are matched with the term nodes of the medical record knowledge graph by name, and unmatched terms are ignored;
[0127] 2. The term node obtains all synonymous terms based on the node synonym relationship, and the medical term used in the structured medical record has the highest probability of association with the structure node;
[0128] 3. Obtain the structural nodes associated with medical terms and merge them to form a medical record structural node set; merge term nodes into structural nodes; and merge unmatched terms following the previous input term.
[0129] 4. Reorganize the medical terms according to the medical record structure node set for interface display. Each medical record structure node constitutes a sub-area of the interface, and each sub-area displays the medical term set under the structure, and the corresponding transcribed medical terms are selected. Other medical terms are arranged from large to small according to the probability of similarity with the selected term and the probability of association with the medical record structure node.
[0130] More specifically, the present invention also provides a specific embodiment, which creates a medical data structured processing method, specifically radiating around the medical record status of stroke, and based on the established stroke medical record knowledge base, can quickly help doctors quickly match and associate stroke-related symptoms, contraindications, physical examinations, inspections, tests and other medical terms, so that doctors can quickly and accurately complete the actual situation matching during the medical record transcription process, with good results.
[0131] More specifically, the present invention also provides the following embodiments. When a doctor is transcribing a medical record, he or she uses the core key words of stroke to eliminate interference items and only retains medical terms that match stroke. Based on the deep neural guided convolutional model Tend, the mapping association Z and the medical term conversion set, the non-standardized medical dialogue is transcribed into standardized medical terms and the medical term medical record information is formed, as shown below:
[0132] Input 1: Stroke patient has had left lower limb weakness for 3 days, had cerebral infarction, is conscious, heart rate is 71
[0133] Transcript 1: [(left lower limb, weakness, 3, days) (past illness, cerebral infarction), (heart rate, 71), (consciousness)];
[0134] Structured 1: [(Current medical history: left lower limb weakness for 3 days) (Past medical history: cerebral infarction) (Physical examination: heart rate 71 bmp, clear mind)]
[0135] It should be noted that in the above embodiment, the starting point for entering text information in the medical record is stroke, the relationship is converted into symptoms and physical examination results, the end point heart rate is entered, and the structured processing results are the chief complaint, medical history, and physical examination. During the structuring process, the default unit (heart rate unit bmp) and the commonly used synonym conversion in the hospital (conscious-clear) are added. The rest of the situations are not described in detail here.
[0136] It should be noted that the technical means for implementing a medical data structured processing method system in another embodiment of the present invention are the same as those of a medical data structured processing method, and thus will not be described in detail herein.
[0137] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other.
Claims
1. A method for structuring medical data, characterized in that: The method comprises the following: Acquire the voice recording data of the doctor entering the medical record; remove the interference sound of the voice recording data and save the medical record voice data; Obtain the medical record voice data entered by the doctor, and transcribe the medical record voice data into corresponding medical record text information; Construct a medical record speech transcription error checking model to perform medical record speech transcription error checking based on the transcribed medical record text information: If the evaluation result based on the medical record speech transcription error checking model is normal, output the corresponding transcribed medical record text information; If the evaluation result based on the medical record voice transcription error checking model is abnormal, re-transcribe the corresponding medical record voice data into medical record text information; A medical record knowledge base is constructed, the medical record text information with a normal evaluation result is matched with the medical record knowledge base, and the matched medical terminology medical record entry information based on the medical record knowledge base is output.
2. A method for structuring medical data according to claim 1, characterized in that: The method for removing the interfering sound of the voice recording data comprises the following steps: Construct a voice feature model for obtaining medical record entry by doctors, wherein the influencing factor of the model is the single-segment impulse ratio corresponding to the medical record entry time and any unit time; Mark any medical record recording voice acquisition start point to record the mark point, based on the mark point according to the voice acquisition frequency f, until any medical record recording voice ends to obtain the mark point number H, then the medical record recording time t is: The total impulse C of any medical record recording speech is marked, and is divided into x segments according to the speech acquisition frequency f based on the marked point. The impulse of any single segment of medical record recording speech divided into x segments is C i (i=1, 2, ..., x), the impulse ratio of any single segment of medical record recording speech divided into x segments is ε i (i=1,2,...,x), the single-stage impulse ratio is calculated as follows:
3. A method for structuring medical data according to claim 2, characterized in that: The method further comprises: Based on the constructed medical record entry time t and single-segment impulse ratio ε i , detect all collected recorded sounds: If the detected single-segment impulse ratio first rises to the maximum value and then decreases, the corresponding recorded sound is retained; otherwise, all collected recorded sounds are deleted and the above steps are repeated; Based on the retained recorded sound that first rises to a maximum value and then decreases, the sound corresponding to the maximum value of the medical record recording time t is retained as the target sound.
4. A method for structuring medical data according to claim 3, characterized in that: The method for checking medical record voice transcription errors based on transcribed medical record text information specifically includes the following steps: Construct an arbitrary medical record text information distribution model L(b) as follows: Among them, L(b i |b i-p+1 , ..., b i-1 ) is the calculation object of the medical record text information distribution model L(b), represented by the i-th medical record entry word b i The distribution of p-1 is represented by the current medical record entry word b i The first p-1 medical record entry words; Construct an arbitrary medical record text information fault-tolerant model T(b), Among them, K(b i-1 b i ) represents the number of times the corresponding medical record entry word appears in the medical record knowledge base; Based on the constructed medical record text information distribution model L(b) and medical record text information fault tolerance model T(b), the next predicted word accuracy probability of any medical record input word in the medical record text information is calculated: where ω is the regularization constant, λ is the stability coefficient, and b i-1 b i Represents any combination of two medical record entry words; 5. A method for structuring medical data according to claim 4, characterized in that: The method for checking medical record voice transcription errors based on transcribed medical record text information also includes the following: Assuming that the evaluation index of the evaluation result based on the medical record speech transcription error checking model is Right, we have: in, It is represented as the current entry probability description of any medical record entry word, and Γ is the description weight.
6. A method for structuring medical data according to claim 5, characterized in that: The method for matching the medical record text information with a normal evaluation result with the medical record knowledge base specifically includes the following steps: Generate any case text information that has been checked by medical record voice transcription, and mark all medical record input words that have been identified and detected as b i , then the set of case text information and medical record entry words is expressed as B = (b1, b2, b3…b m ); Based on the generated set B, a deep neural guided convolutional model Tend is constructed: Tend = aggcn(B, B) (7) Construct a mapping association Z between case text information and medical terms. The mapping association Z is constructed as follows: Among them, the mapping association Z1 is the corresponding relationship between the jump degree Q1 as the first jump and the jump degree Q2 as the second jump, the mapping association Z2 is the corresponding relationship between the jump degree Q2 as the first jump and the jump degree Q1 as the second jump, and ∫(Q1, Q2) is expressed as the dynamic association of Q1 and Q2.
7. A method for structuring medical data according to claim 6, characterized in that: After completing the construction of the mapping association Z, the method further includes the following: Obtain a mapping association Z, and input the mapping association Z into the medical term conversion set: in, is a sigmoid function, parZ is the input parameter matrix based on the medical record input words, Z 12 is the dot product of the mapping association Z1 and Z2, is the mapping associated offset vector.
8. A method for structuring medical data according to claims 1-7, characterized in that: The constructed model includes a starting point for entering medical record text information, relationship conversion, and an entry end point.
9. A medical data structured processing method system, characterized in that: The system comprises the following: A voice filtering module is used to obtain the voice recording data of the doctor entering the medical record; remove the interference sound of the voice recording data and save the medical record voice data; A voice transcription module is used to obtain the medical record voice data entered by the doctor and transcribe the medical record voice data into corresponding medical record text information; The voice verification module is used to build a medical record voice transcription error checking model and perform medical record voice transcription error checking based on the transcribed medical record text information: If the evaluation result based on the medical record speech transcription error checking model is normal, output the corresponding transcribed medical record text information; If the evaluation result based on the medical record voice transcription error checking model is abnormal, re-transcribe the corresponding medical record voice data into medical record text information; The transcription redirection module is used to build a medical record knowledge base, match the medical record text information with a normal evaluation result with the medical record knowledge base, and output the matched medical terminology medical record entry information based on the medical record knowledge base.
10. A medical data structured processing method system according to claim 9, characterized in that: The voice filtering module includes the following: A model building unit, used to build a voice feature model for obtaining a doctor's medical record entry, wherein the influencing factor of the model is a single-segment impulse ratio corresponding to the medical record entry time and any unit time; The speech marking unit is used to mark the start point of any medical record recording speech acquisition and record the marking point. Based on the marking point, the speech acquisition frequency f is obtained until any medical record recording speech ends to obtain the marking point number H. The medical record recording time t is: The impulse calculation unit is used to mark the total impulse C of any medical record recording voice, and divide it into x segments according to the voice acquisition frequency f based on the marking point. The impulse of any single segment of medical record recording voice divided into x segments is C i (i=1, 2, ..., x), the impulse ratio of any single segment of medical record recording speech divided into x segments is ε i (i=1,2,...,x), the single-stage impulse ratio is calculated as follows:
Citation Information
Patent Citations
Method for detecting and correcting error on text after voice recognition
CN101655837A
Voice input electronic medical record information system and operation method
CN116959453A
Cited By
Medical assistance method and system based on voice recognition and medical information knowledge base
CN119742020A
