Intelligent voice interaction and document automation generation method and system and intelligent chip

Abstract

The application provides a kind of intelligent voice interaction and document automation generation method, system and intelligent chip, belong to speech recognition technical field, specifically includes: template generation trigger and intelligent assembly: from template library automatically call corresponding template and intelligently assemble, generate personalized initial ward round record;Multi-source data integration and intelligent filling: combined with built-in professional medical knowledge base, the patient's physical data is logically analyzed and semantically aligned, to generate a preliminary, structured text description and targeted medical assistance prompts in the corresponding position of the personalized initial ward round record template, editing, auditing and archiving: provide a rich text editing interface, support nurses to modify text, voice modification, project check, common symptoms / signs of standardized terminology point selection operation, support standardized terminology selection and common symptom matching operation based on medical knowledge base, record locking and synchronously archived to electronic medical record system, improve the convenience of ward round record generation processing.

Description

Technical Field

[0001] This invention belongs to the field of speech recognition technology, and in particular relates to an intelligent speech interaction and document automated generation method, system and intelligent chip. Background Technology

[0002] In the current healthcare industry, medical documentation is a crucial component of medical and nursing work, especially nursing rounds records. Their accuracy, completeness, and timeliness directly impact the quality of patient care and medical safety. Particularly for brain diseases, traditional medical documentation templates are often generic, lacking personalized adaptation to the patient's specific nursing level, department, and disease. Medical staff must manually add numerous specialist observation items and disease-related assessments to these generic templates, a cumbersome process prone to overlooking crucial information.

[0003] To address the aforementioned technical problems, existing solutions combine ward round systems with hospital information systems that have databases, thereby resolving the issue of low efficiency in ward rounds and queries. However, these solutions have the following drawbacks: In the process of generating ward round records using speech recognition, a certain degree of recognition bias is inevitable. Therefore, it is urgent to solve the technical problem of determining the recognition strategy for modifying the data after speech recognition processing based on the risk of recognition bias in different disease types, and determining the management strategy for ward round records based on speech recognition in different disease types based on the recognition results, so as to reduce the risk of information errors caused by recognition bias.

[0004] Therefore, there is an urgent need for an intelligent voice interaction and document automation generation method, system, and intelligent chip. Summary of the Invention

[0005] To achieve the objectives of this invention, the following technical solution is adopted: Firstly, this application provides a method for intelligent voice interaction and automated document generation, specifically including: S1 uses the generated text of ward round records based on speech recognition results to determine the modified data in ward round records for different disease types. It uses the modified data of each nursing staff member to determine the identification deviation information item of the nursing staff member. It uses the similarity of the identification deviation information item of the nursing staff member in the disease type with different nursing staff members, and combines the identification deviation information item data of the nursing staff member to determine the identification risk disease type in the disease type. It uses the identification risk disease type in the disease type and the modified data of the nursing staff member to determine the identification processing scheme of the modified data. S2 uses the aforementioned identification processing scheme to determine semantic recognition deviation data under different information items, and combines the association between the information items and different identification risk disease types to determine the text generation control strategy for the identification risk disease types; S3 determines the text generation method for the ward round records of the disease type based on different text generation control strategies for identifying risky disease types and the degree of association between the ward round records of the disease type and the identified risky disease type in different information items.

[0006] The beneficial effects of this invention are as follows: By utilizing the risk identification of disease types and the modification data from nursing staff, a modification data identification and processing plan is determined. Based on the number of diseases with a high risk of speech recognition errors and the individual bias characteristics of nursing staff, the overall identification bias risk is determined, and the identification and processing targets that need to be identified during the modification process are dynamically adjusted according to the overall identification bias risk. This clarifies the reasons for the modifications, optimizes resource allocation, improves the overall medical record quality, and reduces the difficulty of identifying and processing modified data.

[0007] By utilizing semantic recognition bias data under different information items and the correlation between information items and different risk disease types, the text generation control strategy for risk disease types is determined. Based on semantic recognition bias data, the overall recognition bias risk is determined, and the recognition bias risk is used to identify risk disease types that cannot be generated using speech recognition. On the one hand, this reduces the risk of bias in ward round records due to the use of speech recognition for the aforementioned disease types. On the other hand, given the relatively high overall recognition bias risk, the control scope is dynamically adjusted to avoid the technical problem of low iteration rate of speech models due to the inability to effectively obtain negative samples because the control scope is too large.

[0008] Furthermore, the generated text of the ward round record is the result of filling in the information items of the ward round record based on the voice recognition results of the nursing staff.

[0009] Furthermore, the modified data in the ward round record is determined based on the modifications made by the nursing staff to the information items in the ward round record after they have been filled out.

[0010] Furthermore, the method for determining the risk disease type among the disease types is as follows: Using modification data from various nursing staff, determine the number of times each nursing staff member made modifications to different information items, and determine the identification deviation information items for each nursing staff member based on the number of modifications. By using the similarity between the identification deviation information item of the nursing staff in the disease type and different nursing staff, the nursing staff to which the information item belongs to the identification deviation information item are determined; by using the nursing staff to which the information item belongs to the identification deviation information item, the deviation risk information item in the information item is determined. Based on the modification data of each nurse in the deviation risk information item and the identification deviation information item, determine whether the disease type belongs to the identified risk disease type.

[0011] Furthermore, the method for determining the text generation method of the ward round record for the aforementioned disease type is as follows: Using different text generation control strategies for identifying risky disease types, the generation control disease type among the identified risky disease types is determined; Based on the semantic recognition deviation data of the ward round records for the disease type in different information items, determine a type of deviation risk information item in the information items of the disease type, and use it as a risk information item; Based on the identified risk disease types, the risk information items in the disease types, and the degree of correlation with the information items of the production management disease types, a method for generating the text of the ward round records for the disease types is determined.

[0012] Secondly, this application provides an intelligent voice interaction and automated document generation system, employing the aforementioned intelligent voice interaction and automated document generation method, specifically including: Template generation trigger and intelligent assembly: The system receives nursing level addition or change information pushed by the HIS system, or responds to daily scheduled tasks, and automatically triggers the ward round template generation process; the system automatically retrieves the corresponding template from the template library for intelligent assembly based on the patient's nursing level, department and diagnosis results, and generates a personalized initial version of the ward round record. Multi-source data integration and intelligent filling: Integrates multiple data sources, including automatically extracting current valid medical orders from the medical order system, automatically importing positive signs and nursing problems that need continuous monitoring, and also importing basic information such as patient name, bed number, hospital number, and allergy history. Combined with the built-in professional medical knowledge base, semantic mapping is performed on the extracted sign data, and structured descriptions and communication records between nurses and patients are generated in the corresponding positions of the personalized initial ward round record. Editing, reviewing, and archiving: The system provides a rich text editing interface, supporting nurses to modify text and voice, select items, and select standardized terms for common symptoms / signs. The system saves the automatically generated initial version and the final version modified by the nurse. After review and confirmation, the record is submitted, locked, and synchronously archived to the electronic medical record system.

[0013] Thirdly, this application provides a smart chip for the aforementioned intelligent voice interaction and document automation generation system, specifically comprising: The data acquisition interface is configured to acquire multimodal heterogeneous data from the user during the training process, and the storage unit is configured to store preset threshold parameters and program instructions. A processing unit, coupled to the data acquisition interface and the storage unit, is configured to execute the program instructions to achieve this.

[0014] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0016] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of a method for intelligent voice interaction and automated document generation; Figure 2 This is a flowchart illustrating the method for determining risk disease types within the disease categories; Figure 3 It is a flowchart of the method for determining the identification and processing scheme of modified data; Figure 4 This is a framework diagram of an intelligent voice interaction and automated document generation system. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0019] Example 1 like Figure 1 As shown, this application provides a method for intelligent voice interaction and automated document generation, specifically including: S1 uses the generated text of ward round records based on speech recognition results to determine the modified data in ward round records for different disease types. It uses the modified data of each nursing staff member to determine the identification deviation information item of the nursing staff member. It uses the similarity of the identification deviation information item of the nursing staff member in the disease type with different nursing staff members, and combines the identification deviation information item data of the nursing staff member to determine the identification risk disease type in the disease type. It uses the identification risk disease type in the disease type and the modified data of the nursing staff member to determine the identification processing scheme of the modified data. S2 uses the aforementioned identification processing scheme to determine semantic recognition deviation data under different information items, and combines the association between the information items and different identification risk disease types to determine the text generation control strategy for the identification risk disease types; S3 determines the text generation method for the ward round records of the disease type based on different text generation control strategies for identifying risky disease types and the degree of association between the ward round records of the disease type and the identified risky disease type in different information items.

[0020] Furthermore, the generated text of the ward round record is the result of filling in the information items of the ward round record based on the voice recognition results of the nursing staff.

[0021] Furthermore, the modified data in the ward round record is determined based on the modifications made by the nursing staff to the information items in the ward round record after they have been filled out.

[0022] Specifically, such as Figure 2 As shown, the method for determining the risk disease type among the disease types is as follows: Using voice recognition technology to assist nursing staff in filling out ward round records has become an important means of improving work efficiency. However, due to factors such as accents, environmental noise, or unfamiliar medical terminology, voice recognition may have biases, leading to errors in the information fields of the ward round records. How to accurately identify which disease types have a higher risk of error in the process of generating ward round records through voice recognition, so as to conduct targeted review or optimization, is a problem that urgently needs to be solved.

[0023] S11 uses the modification data among various nursing staff to determine the number of modifications made by the nursing staff in different information items, and determines the identification deviation information item of the nursing staff based on the number of modifications; Nursing staff: Refers to medical personnel responsible for filling out ward round records, such as doctors and nurses. Information items: The basic data units that make up the ward round record, such as "changes in chief complaint," "vital signs," "physical examination," "auxiliary examination results," "medication adjustments," and "ward round comments." Number of modifications: Refers to the number of times a nursing staff member manually modifies the pre-filled results of the voice recognition for a particular information item. Recognition deviation information items: Refers to information items for which the number of modifications exceeds a preset modification threshold for a specific nursing staff member, indicating that this information item is prone to recognition errors for that nursing staff member.

[0024] The purpose of this step is to identify individual nurses' "prone errors" in speech recognition from their operational data. Because each nurse has a different accent, speaking speed, and language habits, the accuracy of the speech recognition system varies across different information items. By counting the number of corrections, this individual difference can be objectively quantified, providing foundational data for subsequent analysis to determine whether the issue stems from the "person" or the "disease." Its significance lies in achieving a precise profile of each nurse's individual recognition ability, avoiding the simplistic attribution of individual differences to disease type.

[0025] For example, during ward rounds in the cardiology department, when nurse A fills in the "vital signs" information field, the system frequently misidentifies "diastolic pressure" in the voice-recognized "blood pressure 120 / 80," requiring her to make corrections more than four times a week, exceeding the preset correction threshold (e.g., three times per week). Therefore, "vital signs" is identified as a recognition error field for nurse A. Similarly, when nurse B fills in "medication adjustment," the system frequently misidentifies "aspirin" as "atorvastatin," and the number of corrections also exceeds the threshold. Therefore, "medication adjustment" becomes a recognition error field for nurse B.

[0026] S12 uses the similarity between the identification deviation information item of the nurse in the disease type and different nurses to determine the nurses to whom the information item belongs to the identification deviation information item, and uses the nurses to whom the information item belongs to the identification deviation information item to determine the deviation risk information item in the information item; Disease Type: This refers to the disease classification corresponding to the ward round record, such as "coronary heart disease," "diabetes," or "pneumonia." Deviation Risk Information Item: This refers to an information item within a specific disease type that is marked as a personal identification deviation item by a majority of nurses (or more than a certain number of nurses), thus rising to the level of a common high-risk information item for that disease type. In short, if an information item is a "mistake-prone point" for different nurses handling the same type of disease, then that information item itself carries a high deviation risk.

[0027] The purpose of this step is to conduct a cross-sectional comparison of individual identification bias information items to identify which information items are generally difficult to recognize when dealing with specific diseases. If multiple unrelated nurses show a high modification rate for the same information item, the problem is likely not with a particular nurse's individual pronunciation, but rather with the medical terminology corresponding to that information item being easily confused in the context of a specific disease, or with the speech recognition model generally lacking the ability to recognize such terms. Its significance lies in achieving a shift from "individual bias" to "common risk," providing a basis for identifying high-risk information items.

[0028] Continuing the example above, under the disease type of "coronary artery disease," several nurses (such as A, C, and D) frequently modified the identification results regarding the description of "nitroglycerin" when filling out the "medication adjustment" form. In this case, "medication adjustment" is no longer merely an individual issue for nurse B, but has become a bias risk information item under the "coronary artery disease" disease type. Conversely, nurse A's issue with "vital signs" is not generally corroborated by other nurses under the "coronary artery disease" type; therefore, "vital signs" does not currently constitute a bias risk information item for "coronary artery disease."

[0029] S13 determines whether the disease type belongs to the identified risk disease type based on the modification data of each nurse in the deviation risk information item and the identification deviation information item.

[0030] It should be noted that the identification deviation information item is the information item whose modification count is greater than a preset modification count threshold.

[0031] Specifically, the deviation risk information item is an information item in which the number of nursing staff belonging to the deviation identification information item does not meet the requirements. For example, the information item in which the number of nursing staff belonging to the deviation identification information item is not less than 2 people.

[0032] Specifically, determining whether the disease type belongs to the risk-identifying disease types includes: Case 1: If the number of deviation risk information items is greater than the preset threshold for the number of risk information items, then the disease type is determined to belong to the risk identification disease type.

[0033] Number of deviation risk information items: refers to the total number of common high-risk information items identified in S12 for the current disease type.

[0034] When multiple information items for a given disease type are consistently reported to be difficult to identify, it indicates systemic risks in several key stages of the ward round record generation process for that disease type. This scenario is designed to quickly identify disease types with a wide range of problems. This is a "focus on the big picture" strategy, prioritizing the areas with the highest concentration of risk.

[0035] For example, for the disease type "pneumonia," analysis revealed that the three information items "vital signs," "physical examination," and "medication adjustment" were all identified as deviation risk information items. Assuming a preset threshold of 2 risk information items, the number of deviation risk information items for the "pneumonia" disease type (3) exceeds the threshold, therefore "pneumonia" is directly determined to be a disease type with identified risks.

[0036] Scenario 2: If the number of deviation risk information items is not greater than the preset threshold for the number of risk information items, the nurses who belong to the deviation information items in the deviation risk information items will be the deviation matching nurses. When the proportion of the deviation matching nurses in the nurses of the disease type is greater than the preset threshold for the proportion of nurses, the disease type is determined to belong to the identification risk disease type.

[0037] Bias-matched caregivers: These are caregivers identified as having personal identification bias in a particular bias risk information item in S12. Bias-matched caregiver percentage: This refers to the proportion of bias-matched caregivers out of all caregivers who have treated this type of disease.

[0038] The limited number of high-risk information items does not necessarily mean the disease type is risk-free. It's possible that while only one or two information items are common high-risk, almost all nursing staff treating the disease make mistakes on these items. This indicates a widespread problem in identifying these information items, severely impacting the quality of ward round records for this disease type. Therefore, it's necessary to examine the breadth of personnel involved; if most people make mistakes, even a single high-risk item is sufficient to classify the disease type as high-risk.

[0039] For example, for the disease type "diabetes," there is only one bias risk information item: "medication adjustment." However, there are 10 nurses who have handled this disease type, and 8 of them (such as A, B, C, D, etc.) are identified as biased in this information item, meaning there are 8 biased matching nurses. Assuming the preset threshold for the proportion of nurses is 50%, then the proportion of 8 / 10 = 80% is much greater than the threshold, therefore "diabetes" is determined to be a disease type with identified risk.

[0040] Scenario 3: If the proportion of the deviation-matched nurses among the nurses for the disease type is not greater than the preset nurse proportion threshold, the proportion of the deviation risk information item in the deviation information item is determined based on the number of identification deviation information items for different deviation-matched nurses, and the deviation risk weight value of the deviation-matched nurse is determined. If there are no deviation-matched nurses whose deviation risk weight value is greater than the preset weight threshold, it indicates that the nurses have identification deviations in many identification deviation information items, which may be due to their own pronunciation problems, resulting in a high probability of identification deviation. Therefore, it is determined that the disease type does not belong to the identification risk disease type.

[0041] Bias Risk Weight: This measures the extent to which a bias-matched caregiver's "problems" are caused by specific risk items (i.e., bias risk information items) for the current disease type, rather than their general speech recognition problems. It is calculated by dividing the number of bias information items marked as bias risk information items under the current disease type by the total number of their personal bias information items. A higher proportion indicates that the caregiver's problems are more focused on frequently encountered information items, and the higher the risk of bias in the information system's recognition.

[0042] When the number of people involved is not broad enough, it is necessary to analyze the root cause of each person's problem in depth. If a nurse has identification biases on many information items (i.e., many information items with personal identification biases), and the risk items for the current disease type only account for a small portion of these, meaning the proportion of information items for which most people have problems is smaller, then their error in this area is more likely due to personal reasons (such as a strong accent) rather than the risk of the disease type itself. In this case, the disease type should not be classified as high-risk. This step aims to eliminate the interference of "problematic nurses" in the assessment of disease type risk, achieving refined screening.

[0043] For example, for the disease type "hypertension," there are two bias risk information items: "vital signs" and "ward round comments." Only two nurses, Zhang and Li, are matched for bias matching, and they represent a small percentage of all nurses handling "hypertension" cases (assumed to be 20%, less than the threshold of 50%). Zhang has 10 personal bias information items (such as changes in chief complaint, physical examination, medication adjustments, etc.), of which "vital signs" and "ward round comments" account for only 2 / 10 = 0.2 of his total bias items. Li has 5 personal bias information items, with the aforementioned two risk items accounting for 2 / 5 = 0.4. Assuming a preset weight threshold of 0.8, the calculated bias risk weight values ​​for both individuals do not exceed the threshold of 0.8. This means that Zhang and Li's problems are scattered, and their errors regarding "hypertension" may stem from their own pronunciation habits rather than the specific terminology recognition difficulties inherent in "hypertension." Therefore, "hypertension" is determined not to be a disease type requiring risk matching.

[0044] Scenario 4: If there are deviation care matching personnel whose deviation risk weight value is greater than the preset weight threshold, these personnel will be used as screening matching personnel. The risk value will be determined by the average of the proportion of deviation matching personnel among the nursing personnel of the disease type and the proportion of screening matching personnel among the deviation matching personnel. When the risk value of the disease type is greater than the preset risk threshold, the disease type will be determined to belong to the identified risk disease type.

[0045] Screened matching personnel: These are nurses whose personal questions are highly focused on the risk items of the current disease type (i.e., their bias risk weight value is high), and they are less likely to be misidentified due to pronunciation problems. Risk value: A comprehensive indicator used to ultimately determine the risk level of the disease type. It is obtained by calculating the arithmetic mean of two ratios: "the proportion of bias-matched nurses to all nurses" and "the proportion of screened matching personnel to bias-matched nurses."

[0046] When there are nursing staff highly focused on this type of disease, it indicates that there are indeed specific identification challenges for that disease type. At this point, two dimensions need to be considered comprehensively: the breadth of personnel involved (the first proportion) and the depth of problem focus (the second proportion). Averaging both allows for a more comprehensive risk assessment. If this overall risk value is sufficiently high, the disease type is ultimately determined to be high-risk. This step provides a more refined and balanced quantitative decision-making method, ensuring the scientific rigor and accuracy of the determination.

[0047] For example, regarding the disease type "heart failure," analysis revealed two bias risk information items: "vital signs" and "medication adjustment," with a quantity equal to the preset threshold of 2. Five nurses (N1 to N5) were matched for bias matching, representing 50% of all nurses handling this disease type (out of 10) (P1=0.5). The individual bias information items and weights for these five nurses are calculated as follows: N1: The only personal identification deviation information item is "vital signs" (1 item), which belongs to the deviation risk information item, with a weight of 1 / 1 = 1.0.

[0048] N2: Personal identification bias information items include "vital signs" and "medication adjustment" (2 items), both of which are risk items with a weight of 2 / 2 = 1.0.

[0049] N3: The only personal identification bias information item is "medication adjustment" (1 item), with a weight of 1.0.

[0050] N4: The personal identification deviation information items include "medication adjustment" and "change in chief complaint" (2 items), among which "medication adjustment" is a risk item with a weight of 1 / 2 = 0.5.

[0051] N5: Personal identification bias information items include "vital signs" and "physical examination" (2 items), among which "vital signs" is a risk item with a weight of 1 / 2 = 0.5.

[0052] Assuming a preset weight threshold of 0.8, the weights of N1, N2, and N3 are greater than 0.8, and they are identified as matched individuals. The proportion of matched individuals among the biased matched nursing staff is 3 / 5 = 0.6. The risk value R is calculated as (0.5 + 0.6) / 2 = 0.55. Since the preset risk threshold is 0.6, 0.55 < 0.6, therefore "heart failure" is not considered a risky disease type. If the weight of N4 is adjusted to 1.0 (i.e., its personal bias information item is only "medication adjustment"), the matched individuals become N1, N2, N3, and N4, a total of 4 people, with a proportion of 4 / 5 = 0.8. The risk value R = (0.5 + 0.8) / 2 = 0.65 > 0.6, and "heart failure" is considered a risky disease type.

[0053] This invention starts with modified data from individual caregivers, progressively identifying high-risk disease types, thus avoiding misjudgments caused by individual pronunciation differences. It can accurately identify systemic speech recognition risks caused by medical terminology, disease characteristics, etc.

[0054] This invention, by setting multiple judgment scenarios, comprehensively considers the number of high-risk information items, the breadth of personnel involved, and the depth of problem focus, forming a three-dimensional assessment system from macro to micro, ensuring the comprehensiveness and accuracy of risk assessment.

[0055] Specifically, such as Figure 3 As shown, the method for determining the identification and processing scheme of the modified data is as follows: Using speech recognition technology to assist nursing staff in completing ward round records has become an important means of improving work efficiency. However, due to factors such as accents, environmental noise, or unfamiliar medical terminology, speech recognition may be inaccurate, leading to errors in the information entered into the ward round records. In the previous steps, we were able to identify which disease types belong to the identification risk categories, that is, diseases with a high risk of speech recognition errors. Based on this, how to develop a reasonable and efficient data identification and processing plan based on the distribution of these risk disease types and the individual bias characteristics of nursing staff, that is, to clarify exactly which problems require modification, so as to optimize resource allocation and improve the overall quality of medical records, has become a problem that needs to be further solved in this field.

[0056] S21 uses the risk identification disease type data in the disease type to determine the proportion of risk identification disease types in the disease type, and uses the proportion of risk identification disease types in the disease type as the risk identification ratio. Identify high-risk disease types: These refer to disease types identified through step S13 that pose a high risk of error during voice recognition-based ward round record filling. Identify risk ratio: This refers to the proportion of all examined disease types identified as high-risk disease types.

[0057] This step aims to macroscopically assess the overall risk level of the speech recognition system in the current medical scenario. A high risk ratio indicates widespread recognition difficulties across most disease types, potentially suggesting systemic problems (such as poor overall adaptability of the speech model or prevalent environmental noise), requiring a comprehensive approach. A low risk ratio suggests the problem may be concentrated on a few disease types, allowing for more targeted, localized solutions. This ratio serves as the first watershed for subsequent decisions, determining the basic direction of the processing strategy.

[0058] A hospital has five common diseases: coronary heart disease, diabetes, pneumonia, hypertension, and heart failure. After analysis using step S13, coronary heart disease, pneumonia, and heart failure were identified as risk-prone disease types, while diabetes and hypertension were not. Therefore, the risk identification ratio is 3 / 5 = 0.6.

[0059] S22 determines the matched personnel for the disease type based on the modified data of the nursing staff; Screening and matching personnel: In step S13, case 4, this refers to the nursing staff whose deviation risk weight value is greater than the preset weight threshold. The identification deviation of these nursing staff is highly focused on specific risk information items of the current disease type, and their identification deviation on other information items is less. Therefore, it can be assumed that the probability of them having general pronunciation problems is low, and their errors are more likely to reflect the identification difficulties inherent in the disease type itself.

[0060] Screening and matching personnel is crucial for identifying the "critical minority" within the at-risk disease types. Their errors primarily stem from a higher probability of difficulty recognizing disease-specific medical terminology, rather than individual pronunciation issues. Identifying these individuals allows for the differentiation between "requiring comprehensive treatment of all individuals with deviations" and "only treating key individuals" when developing subsequent treatment plans, thereby ensuring precise resource allocation.

[0061] Based on the calculation results of S13, for the coronary heart disease (CHD) disease type, the bias-matched nursing staff are N1, N2, N3, N4, and N5, a total of 5 people. Among them, the bias risk weight values ​​of N1 and N2 are 1.0 and 0.9 respectively, both greater than the preset weight threshold of 0.8. Moreover, the personal identification bias information items of N1 and N2 are limited to the risk information items of CHD. Therefore, N1 and N2 are identified as the screening matching personnel for the CHD disease type. Conversely, N3 has more personal identification bias information items and lower weight values, and is considered to have a higher probability of having common pronunciation problems. Therefore, N3 is not included in the screening matching personnel.

[0062] S23 determines the identification and processing scheme for the modified data based on the identified risk ratio and the distribution data of the screened and matched personnel in the disease type among the biased matched nursing personnel.

[0063] Modified data identification and processing plan: This refers to the specific processing strategy adopted for modified data in ward round records, determining which issues caused the modifications. For example, should a global review and model optimization be performed on bias-matched nurses across all disease types, or should targeted processing be applied only to nurses selected for matching in specific disease types? Bias-matched nurses: These are nurses who exhibit personal identification bias in the bias risk information item for a particular disease type.

[0064] This step integrates macro-level risk levels (identifying the proportion of risks) and micro-level individual characteristics (screening the distribution of matching personnel), automatically generating the most suitable treatment plan through multi-level condition judgment. This hierarchical decision-making mechanism can avoid excessive consumption of resources in low-risk situations while ensuring sufficient coverage of high-risk situations, reflecting the concept of refined management.

[0065] Specifically, if the identified risk ratio is greater than the preset identified risk ratio, then the probability of group identification bias is relatively high. Therefore, the identification and processing scheme for the modified data is determined to be to identify and process the modified data of the biased matching nurses in all disease types.

[0066] Group recognition bias: This refers to recognition errors that are widespread across multiple disease types. These errors may be caused by the speech recognition system itself, environmental factors, or common accent problems, rather than being specific to any particular disease type. Recognition processing: This involves specialized analysis and review of the modified data, which can be used to optimize speech recognition models or enhance personnel training.

[0067] When the proportion of high-risk disease types is excessively high, it indicates a widespread problem. In this case, localized treatment targeting individual disease types may not resolve the root cause. A global approach, uniformly processing all deviation-matched nursing staff modification data across all disease types, can maximize problem coverage, improve overall identification accuracy, and thus determine the true cause of the modification issues.

[0068] Assume the preset risk identification threshold is 0.7. If a hospital's risk identification ratio is calculated to be 0.8 (meaning 80% of disease types are high-risk), then a group identification bias is identified. The solution is to identify and process the modified data of each biased matching nurse across all disease types (including high-risk and non-high-risk).

[0069] Furthermore, if the identified risk ratio is not greater than a preset identified risk ratio, the following content is also included: Based on the screening and matching personnel data in different disease types, determine whether the proportion of screening and matching personnel in different disease types among the biased matching nurses is greater than the preset screening and matching proportion threshold. If so, determine that the modified data identification and processing scheme is to identify and process the modified data of biased matching nurses in all disease types. If not, proceed to the next step. The proportion of screened-match personnel among biased-match nurses: This refers to the percentage of screened-match personnel among all biased-match nurses for a given disease type. This proportion reflects the degree to which the problem is dominated by personnel with a lower probability of pronunciation issues (i.e., screened-match personnel) within that disease type. If this proportion is high across all disease types, it indicates that the identification bias for each disease type is primarily due to a higher probability of terminology issues related to the disease itself, rather than pronunciation problems of individual nurses, and that all disease types exhibit the same characteristics. In this case, a global approach should be taken.

[0070] This step determines whether the causes of recognition bias are consistent across different disease types and primarily stem from disease terminology. If the proportion of matched individuals is high across all disease types, it indicates that the recognition difficulties for each disease type can be focused on those with fewer pronunciation problems. This suggests a high probability of widespread deficiencies in the speech recognition model regarding these disease terms. At this point, global processing is performed, analyzing the modified data of all biased matched nurses. This allows for comprehensive model optimization and correction of errors caused by terminology issues, regardless of who caused these errors.

[0071] The risk identification ratio was set at 0.5, not exceeding a preset threshold of 0.7. The proportions of matched nurses for each disease type were as follows: coronary heart disease 0.8, diabetes 0.75, pneumonia 0.8, hypertension 0.85, and heart failure 0.9. The preset screening and matching ratio threshold was 0.7. The inspection revealed that the proportions for all disease types were greater than 0.7, indicating a high degree of consistency in the problems across all disease types, and a high probability that the issues were primarily caused by the disease terminology itself. Therefore, the determined treatment plan was to identify and process the modified data of all mismatched nurses across all disease types.

[0072] Based on the screening and matching personnel data in the disease type, determine the proportion of screening and matching personnel in the disease type in the deviation matching nursing personnel, and determine whether the proportion of screening and matching personnel in the disease type in the deviation matching nursing personnel is greater than the preset screening and matching proportion threshold. If so, identify and process the modified data of deviation matching nursing personnel in the disease type; otherwise, proceed to the next step. This step independently assesses each disease type, identifying those with a high percentage of matched personnel. For these disease types, the concentration of key personnel suggests the problem likely stems from disease terminology rather than individual pronunciation. By comprehensively addressing all mismatched nursing staff, all terminology-related errors are captured and optimized, while avoiding resource waste due to the limited number of these disease types. Unselected disease types proceed to the next step for more refined assessment.

[0073] Continuing with the previous example, the preset matching ratio threshold is 0.7. The ratios for each disease type are: coronary heart disease 0.4, diabetes 0.33, pneumonia 0.75, hypertension 0.33, and heart failure 0.5. Each type is judged individually: Coronary artery disease 0.4 < 0.7, no treatment, proceed to sub-step 2.3.

[0074] If the diabetes score is 0.33 < 0.7, no action is taken, proceed to sub-step 2.3.

[0075] Since the value of pneumonia is 0.75 > 0.7, all biased matching modifications to caregivers (hypothetically N1, N3, N5, N9) for the pneumonia disease type are identified and pneumonia is marked as a reliable disease type.

[0076] If the blood pressure is 0.33 < 0.7, no treatment is needed; proceed to sub-step 2.3.

[0077] Heart failure 0.5 < 0.7, no treatment, proceed to sub-step 2.3.

[0078] The disease types for which all deviation-matched nursing staff modification data are identified are considered reliable disease types. It is then determined whether the proportion of the reliable disease types in the risk disease types is greater than a preset risk disease type proportion threshold. If so, the modification data identification process is performed only on the remaining risk disease types for the selected matching personnel. If not, the modification data identification process is performed on the selected matching personnel in all remaining disease types.

[0079] Reliably identified disease types: These refer to the disease types that have been labeled in the sub-step. Remaining identified risk disease types: These are disease types that belong to the identified risk disease types but have not yet been labeled as reliably identified disease types. All remaining disease types: These are all disease types that have not yet been labeled as reliably identified disease types, including both identified and non-identified risk disease types.

[0080] This step involves final resource allocation for disease types that haven't been fully processed. The current processing coverage is determined by calculating the percentage of fully processed, reliably identified disease types among all high-risk disease types. If the coverage is already high (above a threshold), it indicates that the identification of disease types with a high probability of recognition errors due to speech is relatively comprehensive. The remaining high-risk disease types may primarily be due to pronunciation issues of individual caregivers. Therefore, processing should only target key personnel (matched personnel) within these disease types, avoiding over-intervention. If the coverage is insufficient, it indicates that many high-risk disease types have not been fully processed, potentially mixed with terminology and pronunciation problems. In this case, the scope should be expanded to identify and process key personnel across all remaining disease types (including non-high-risk ones) to comprehensively capture any potential terminology-related errors. Even if these errors occur in non-high-risk disease types, their probability of being reflected through matched personnel is high. This design maximizes resource utilization and optimizes risk control.

[0081] It is known that there are three risky disease types: coronary heart disease, pneumonia, and heart failure. In sub-step 2.2, only pneumonia is a reliable disease type. Therefore, the proportion of reliable disease types among the risky disease types is approximately 1 / 3 ≈ 0.33. A preset risky disease type proportion threshold is set to 0.5. Since 0.33 < 0.5, the process proceeds to the "No" branch, meaning that data modification processing is needed for the remaining disease types in the screening of matched individuals. The remaining disease types include: coronary heart disease, diabetes, hypertension, and heart failure (where diabetes and hypertension are not risky disease types). Therefore, the processing solution is as follows: For coronary heart disease: only the modified data of the matched individuals (N1, N2) are identified and processed.

[0082] For diabetes: Modifications to the screening and matching population (N6) are identified and processed only.

[0083] For hypertension: Modifications to the data of the selected matched individuals (N2) are identified and processed only.

[0084] For heart failure: only modified data of its selected matching personnel (N1, N6) are identified and processed.

[0085] For pneumonia (already processed): Identify and process the modified data for all biased matching nursing staff (N1, N3, N5, N9).

[0086] Furthermore, the semantic recognition deviation data under the information item is determined by including the number of times semantic recognition deviation exists under the information item.

[0087] Furthermore, the method for determining the text generation control strategy for identifying risk disease types is as follows: Using voice recognition technology to assist nursing staff in filling out ward round records has become an important means of improving work efficiency. In the preliminary steps, we have been able to identify which disease types belong to the risk disease categories and formulate targeted data identification and processing plans. However, how to further delve into the information item level, analyze the specific manifestations of semantic recognition bias, and formulate differentiated text generation control strategies based on different risk levels to determine under what circumstances the use of voice recognition function should be suspended has become a key link in ensuring the quality of medical documents.

[0088] S31 uses semantic recognition deviation data under different information items to determine the number of times semantic recognition deviation exists under the information item, and determines the recognition deviation risk type of the information item based on the number of times semantic recognition deviation exists among different nursing staff; Semantic recognition bias: This refers to the error where the speech recognition system incorrectly identifies medical terms spoken by nurses as other words, resulting in a semantic change. For example, recognizing "palpitation" as "heart machine." Information item: The basic data unit constituting the ward round record, such as "changes in chief complaint," "vital signs," and "medication adjustments." Personnel with recognition bias: Nursing staff whose semantic recognition bias on a particular information item exceeds a preset threshold. Risk type of recognition bias: Based on the different ranges in which the number of personnel with recognition bias falls, the risk level of information items is divided into three categories: Category I, Category II, and Category III, with Category I having the highest risk and Category III the lowest.

[0089] This step aims to quantify the semantic recognition risk of each information item at a micro level. By statistically analyzing the number of semantic recognition errors made by different nursing staff on each information item, we can identify which information items are commonly "mistake-prone." Classifying the risk level based on the number of staff exhibiting errors provides a clear picture of the risk level of each information item, offering a foundation for developing subsequent control strategies. Its significance lies in achieving a refined granularity of risk identification, moving from disease type down to specific information items.

[0090] Set a preset threshold of 300 instances of semantic recognition deviation per month. Count the number of semantic recognition deviations in the last month's data modifications for each information item, and also count the number of nursing staff whose deviations exceed the threshold. The "Medication Adjustment" information item shows that 10 nursing staff members exceeded the threshold, with deviations of 500 times for N1, 400 times for N2, 500 times for N3, 400 times for N4, and 0 times for N5.

[0091] The following ranges are defined: 10 or more individuals are classified as Category I deviation risk, 3 to 9 as Category II deviation risk, and the remaining 0 to 2 individuals as Category III deviation risk. Therefore, "medication adjustment" and "vital signs" are Category I deviation risk, "changes in chief complaint" and "physical examination" are Category II deviation risk, and "ward round opinions" are Category III deviation risk.

[0092] S32 determines the identification bias risk type of different information items in the identified risk disease type based on the association between the information items and different identified risk disease types; Identify high-risk disease types: These refer to disease types identified through step S13 that have a high risk of error during the voice recognition process for filling out ward round records.

[0093] The importance of the same information item may vary across different disease types. For example, "medication adjustment" is a core information item in coronary heart disease and diabetes, but may be less critical in the common cold. This step links the risk level of an information item to its corresponding disease type, enabling a more accurate assessment of the risk composition within each identified risk disease type and providing a basis for developing differentiated management strategies.

[0094] The known risk disease types are coronary artery disease, pneumonia, and heart failure. For coronary artery disease, the core information items include "medication adjustment" (Category I), "vital signs" (Category I), and "changes in chief complaint" (Category II); for pneumonia, the core information items include "vital signs" (Category I), "physical examination" (Category II), and "medication adjustment" (Category I); for heart failure, the core information items include "vital signs" (Category I), "medication adjustment" (Category I), and "ward round comments" (Category III).

[0095] S33 determines the text generation control strategy for the identified risk disease type based on the identification deviation risk types of different information items and the identification deviation risk types of different information items in the identified risk disease type.

[0096] Text generation control strategy: This refers to the decision-making scheme regarding whether to continue using voice recognition for ward round record filling for identified high-risk disease types. When the risk is too high, text generation control is implemented, meaning that voice recognition is no longer used for ward round record filling, and manual input is adopted instead. This avoids ward round record problems caused by recognition errors that go undetected.

[0097] This step integrates the overall risk level (the proportion of a single risk item among all information items) and local risk characteristics (the distribution of risk items within each disease type), and automatically generates the most suitable control strategy through multi-level condition judgments. This mechanism enables precise control of high-risk disease types when the overall risk is controllable, and avoids the inefficiency caused by over-control when the overall risk is high, achieving a balance between safety and efficiency.

[0098] Specifically, the method for determining the identification bias risk type of the information item is as follows: Based on the number of semantic recognition deviations among different nursing staff, nursing staff whose number of semantic recognition deviations in the information item exceeds a preset threshold for the number of recognition deviations are identified and are regarded as personnel with recognition deviations. The identification bias risk type of the information item is determined based on the range of the number of people with identification bias.

[0099] It should be noted that the identification deviation risk type of the information item is determined according to different quantity ranges and user-preset risk types, specifically including a first-class deviation risk type, a second-class deviation risk type, and a third-class deviation risk type, wherein the first-class deviation risk type is greater than the second-class deviation risk type, and the second-class deviation risk type is greater than the third-class deviation risk type.

[0100] Specifically, based on the identification bias risk types of different information items and the identification bias risk types of different information items within the identified risk disease type, a text generation control strategy for the identified risk disease type is determined, including: The overall risk ratio is determined by the proportion of information items of one type of deviation risk among all information items. It is then determined whether the overall risk ratio is less than a preset risk ratio threshold. If so, text generation control processing is performed as long as there is an information item of one type of deviation risk among the identified risk disease types. That is, voice recognition is no longer used for filling in ward round records, thereby avoiding the occurrence of ward round record problems caused by the existence of identification deviations that are not detected. If not, it means that the overall identification deviation risk is relatively large. Therefore, a more lenient approach is needed to determine the identified risk disease types for text generation control processing, thereby improving the identification and processing efficiency of voice recognition deviations. Therefore, proceed to the next step. Overall risk ratio: This refers to the proportion of all examined information items that are assessed as having a type of bias risk. This indicator reflects the overall risk level of the speech recognition system in the current medical scenario.

[0101] When the overall risk ratio is low, it indicates that the system is performing well overall, with problems concentrated on a few information items. In this case, strict control measures should be implemented, suspending the use of speech recognition for disease types containing high-risk information items. This ensures medical safety without hindering data collection due to overly broad control. Conversely, when the overall risk ratio is high, it indicates widespread recognition problems. In this situation, a large-scale suspension of speech recognition would drastically reduce the amount of erroneous data available for model training, creating a vicious cycle of "the more afraid of errors, the less it's used; the less it's used, the worse the model becomes." Therefore, a more lenient strategy is needed, prioritizing data collection.

[0102] There are 10 information items in total. Among them, two items fall under the category of deviation risk: "Medication Adjustment" and "Vital Signs." The overall risk ratio is 2 / 10 = 0.2. The preset risk ratio threshold is 0.3. 0.2 < 0.3, therefore, the "Yes" branch is entered: as long as an information item of the deviation risk type is identified in the risk disease type, text generation control processing is performed. Coronary heart disease, pneumonia, and heart failure all contain either "Medication Adjustment" or "Vital Signs" as category 1 risk items; therefore, all three disease types are determined to require text generation control processing.

[0103] The disease types other than the identified risk disease types are used as other disease types. It is determined whether the proportion of the other disease types in the disease types is greater than the preset proportion value. If so, only other disease types can be used. During the iteration process of the speech recognition model, more training data is provided for the speech recognition model. Therefore, as long as there is a type of information item of deviation risk in the identified risk disease types, text generation control processing is performed, that is, speech recognition is no longer used for filling in the ward round record, thereby avoiding the occurrence of ward round record problems caused by recognition deviation that is not detected. If not, proceed to the next step. Other disease types: These refer to disease types that were not identified as high-risk disease types, i.e., disease types with lower risk. Preset percentage: This is used to determine if there are enough low-risk disease types to provide sufficient data for the iteration of the speech recognition model.

[0104] When the overall risk is high, control strategies need to be more cautious. If low-risk disease types account for a large proportion, it indicates that there is still a large amount of reliable speech recognition data available for collection and subsequent model optimization. In this case, it is safe to implement strict control measures on high-risk disease types; even if speech recognition is suspended, it will not affect the data source for model training. If there are insufficient low-risk disease types, data on high-risk disease types should not be easily abandoned, and further refined judgment is required.

[0105] Assuming the overall risk ratio is 0.4, which is greater than the threshold of 0.3, proceed to step B. There are 5 disease types in total: coronary heart disease, pneumonia, and heart failure are identified as risk disease types (3 types), while diabetes and hypertension are identified as other disease types (2 types). The proportion of other disease types = 2 / 5 = 0.4. The preset proportion value is 0.5. 0.4 < 0.5, therefore proceed to the next step.

[0106] The risk ratio threshold for identifying the disease type is determined based on the overall risk ratio. Based on the identification deviation risk type of different information items in the identified disease type, the proportion of information items of one type of deviation risk in the identified disease type is determined and used as the risk ratio. Using the risk ratio and the risk ratio threshold, it is determined whether the identified disease type needs to undergo text generation control processing.

[0107] It should be noted that if the risk ratio is greater than the risk ratio threshold, then the identification of the risk disease type requires text generation control processing.

[0108] Risk ratio: This refers to the proportion of information items classified as a biased risk type out of all core information items in a specific disease type. Risk ratio threshold: A judgment standard dynamically adjusted based on the overall risk ratio; the higher the overall risk, the higher the threshold setting, and the more lenient the control.

[0109] This step enables fine-grained control of dynamic thresholds, with the core logic being "the greater the risk, the looser the control." When the overall risk is high, the system urgently needs a large amount of erroneous data to improve the model. Therefore, it is necessary to raise the control threshold, allowing more disease types to continue using speech recognition, even if they carry some risk. This seemingly counterintuitive design is actually aimed at solving the problem at its root—only by collecting enough negative samples can a better model be trained, ultimately reducing the overall risk. This is a "space-for-time" strategy, tolerating some risk in the short term in exchange for long-term improvements in system performance.

[0110] The rule for determining the risk ratio threshold is: Risk ratio threshold = Overall risk ratio × 1.1. If the overall risk ratio is 0.4, then the risk ratio threshold = 0.4 × 1.1 = 0.44.

[0111] The proportion of the information item for a single type of bias risk in each identified risk disease type is calculated as follows: Coronary artery disease: There are 5 core information items (medication adjustment, vital signs, changes in chief complaint, physical examination, and ward round comments). Among them, there are 2 Class I risk items, namely medication adjustment and vital signs, with a risk ratio of 2 / 5 = 0.4.

[0112] Pneumonia: There are 4 core information items (vital signs, physical examination, medication adjustment, and changes in chief complaint). Among them, there are 2 risk items, namely vital signs and medication adjustment, with a risk ratio of 2 / 4 = 0.5.

[0113] Heart failure: There are 4 core information items (vital signs, medication adjustment, ward round comments, and changes in chief complaint). Among them, there are 2 risk items, namely vital signs and medication adjustment, with a risk ratio of 2 / 4 = 0.5.

[0114] The risk ratio was compared with a threshold of 0.6, therefore both pneumonia and heart failure disease types were subject to text generation control processing.

[0115] It is understandable that the higher the overall risk ratio, the higher the risk ratio threshold.

[0116] In this embodiment, the present invention achieves an inverse correlation between the intensity of control and the overall risk level through multi-level judgment logic—strict control ensures safety when the risk is low, and relaxed control collects data when the risk is high, forming an adaptive closed-loop management mechanism.

[0117] When the overall risk is high, by raising the control threshold, the speech recognition function of more disease types can be retained, ensuring that the model can continuously obtain negative sample data including misidentifications, fundamentally supporting the continuous optimization of model performance. The introduction of a dynamic threshold calculation formula based on the overall risk ratio enables the control standards to be smoothly adjusted with changes in system status, ensuring the scientific nature of decision-making and giving the system the ability to self-evolve.

[0118] Furthermore, the method for determining the text generation method of the ward round record for the aforementioned disease type is as follows: It should be noted that the disease types mentioned include all disease types except for those generated and controlled.

[0119] In the process of healthcare informatization, using speech recognition technology to assist nursing staff in filling out ward round records has become an important means of improving work efficiency. In the preliminary steps, we have been able to identify which disease types fall under the risk category and have formulated differentiated text generation control strategies to determine when the use of speech recognition should be suspended. However, for disease types that have not been suspended, how to further determine the text generation method for their ward round records—that is, whether deep recognition processing of modified data is needed to support model iteration—becomes a key issue in balancing medical safety and system optimization. In particular, when a large number of disease types are controlled, limiting the usability of speech recognition, there is an urgent need to accelerate model iteration to lift the controls as soon as possible. In this case, a more proactive data collection strategy is required.

[0120] S41 uses different text generation control strategies to identify risky disease types to determine the generation control disease type among the identified risky disease types; Generate controlled disease types: This refers to the identified risk disease types that are determined to require text generation and control processing in step S33. For these types of diseases, voice recognition for ward round record filling is suspended, and manual data entry is used instead.

[0121] The purpose of this step is to filter out the disease types that have already been controlled from the identified risk disease types, serving as a reference benchmark for subsequent decision-making. These disease types have been suspended from using speech recognition due to their excessively high risk and are therefore no longer involved in the subsequent text generation method determination. However, their number and related information (such as the risk distribution of information items) can be used to assess the severity of control and their association with other disease types.

[0122] It should be noted that the generated and controlled disease types are risk disease types that require text generation and control processing.

[0123] S42 uses the semantic recognition deviation data of the ward round records of the disease type in different information items to determine a type of deviation risk information item in the information items of the disease type, and uses it as a risk information item. Risk information item: refers to an information item that is rated as a type of bias risk within a certain disease type. Type 1 bias risk is the highest risk level, indicating that the information item has a widespread semantic recognition bias problem.

[0124] This step aims to identify high-risk information items within each disease type, which are the primary sources of overall risk for that disease type. The presence of these risk information items is an important consideration when subsequently deciding whether to identify and process modified data for that disease type.

[0125] Based on the classification results of S31, the information items for a type I deviation risk include "medication adjustment" and "vital signs". For coronary heart disease, these two type I risk items are included in its core information items, therefore the risk information items for coronary heart disease are "medication adjustment" and "vital signs"; the same applies to diabetes and hypertension.

[0126] S43 determines the text generation method for the ward round record of the disease type based on the generated control disease type in the identified risk disease type, the risk information item data in the disease type, and the degree of correlation with the information items of the production management disease type.

[0127] Text generation method: This refers to the identification and processing of whether modified data is required for the generation of ward round records for a specific disease type. Modified data identification and processing involves specifically analyzing modified data generated in all ward round records for that disease type to identify patterns of system errors and provide training data for the iterative optimization of the speech recognition model.

[0128] The core idea of ​​this step is that when the level of control is severe (i.e., a large number of disease types have had their speech recognition suspended), the usability of the speech recognition system is significantly reduced, necessitating the collection of more error data to accelerate model iteration and lift the control measures as soon as possible. Therefore, a comprehensive data collection strategy should be adopted at this time to identify and process modified data for all disease types (including those not under control). Conversely, when the level of control is manageable, the targets requiring processing should be selectively determined based on the risk characteristics of each disease type and its similarity to controlled disease types, achieving precise resource allocation.

[0129] It should be noted that if the number of disease types generated and managed in the identified risk disease types does not meet the requirements, the limitation is quite severe. Therefore, for all disease types, the method for generating and managing ward round records involves modifying the data recognition process to improve the efficiency of the speech recognition model's update and iteration processing.

[0130] Requirements Not Met: This indicates that the number of disease types generated for control exceeds the preset control limit. This situation suggests that too many disease types have been suspended, severely limiting the application scope of the speech recognition system. There is an urgent need to lift the restrictions through model optimization, thus requiring the collection of error data to the maximum extent possible.

[0131] When the number of controlled diseases exceeds a certain threshold, it indicates that the current speech recognition system is performing poorly across multiple disease types, increasing the burden of manual data entry and reducing work efficiency. In this situation, it is essential to accelerate model iteration, which requires a large amount of negative sample data containing misidentifications. Adopting a comprehensive data collection strategy, identifying and processing modified data for all disease types (including low-risk and uncontrolled high-risk diseases), can maximize the acquisition of negative samples, providing sufficient material for rapid model iteration and thus expediting the lifting of controls. This is a response measure under "emergency" conditions.

[0132] Additionally, it is understood that if the number of generated controlled disease types among the identified risk disease types meets the requirements, the following situations also apply: Case 1: If there is a risk information item among the associated information items in the disease type, then the method for generating and managing the ward round records of the disease type is determined to be the identification and processing of modified data, thereby improving the efficiency of the update and iteration processing of the speech recognition model.

[0133] Related information items: These refer to the core information items contained in this disease type. Risk information items: These are information items representing a type of deviation risk.

[0134] When a disease type inherently contains high-risk information items, it indicates a widespread difficulty in identifying that disease type, and the resulting modified data has high analytical value. Prioritizing the identification and processing of these disease types can accurately capture model defects and is the most direct and effective optimization path.

[0135] Case 2: If there is no risk information item in the associated information item of the disease type, obtain the number of disease types that have undergone data modification identification processing. If the number of disease types that have undergone data modification identification processing is greater than the preset disease type number threshold, then determine that the generation and management method of the ward round records of the remaining disease types is to no longer undergo data modification identification processing. Number of disease types requiring data modification: This refers to the cumulative number of disease types that have been identified as needing processing during the current decision-making process based on condition 1 or other criteria. Preset disease type threshold: This controls the upper limit of the total number of processes, preventing excessive resource allocation.

[0136] If a sufficient number of disease types have been identified and the threshold has been reached, it indicates that there is enough data available for model optimization. At this point, the remaining disease types that do not contain risk information items have relatively low data value for modification and do not need to be processed. Resources can then be concentrated on the selected key disease types.

[0137] Case 3: If the number of disease types requiring data modification is not greater than the preset disease type number threshold, the identification deviation risk value is determined based on the proportion of the information items of the second type of deviation risk in the associated information items of the disease type. It is then determined whether the identification deviation risk value of the disease type is greater than the preset risk threshold. If yes, proceed to the next step; otherwise, determine that the method for generating and managing the ward round records of the disease type is to not perform data modification identification processing. The proportion of information items classified as Category II risk level: This refers to the percentage of core information items for a given disease type that are assessed as Category II risk level. Risk identification value: This proportion measures the concentration of risk for that disease type on medium-risk information items.

[0138] When the number of cases already treated is insufficient and the disease type does not have a Class I risk item, it is necessary to examine the proportion of Class II risk items. If the proportion of Class II risk items is high (e.g., greater than 0.5%), it indicates that the disease type still has a certain degree of prevalence and may be worth including in the treatment scope, but further judgment is needed based on its correlation with the controlled disease type. If the proportion is too low, the risk is small and it can be directly excluded.

[0139] Based on the correlation between the information items in the ward round records of the disease type and the information items of the generated and controlled disease type, the proportion of information items belonging to both the disease type and the generated and controlled disease type in the information items of the ward round records of the disease type is determined, and this proportion is used as the correlation coefficient of the disease type. It is then determined whether the correlation coefficient of the disease type is greater than a preset correlation coefficient threshold. If it is, the generation and management method of the ward round records of the disease type is determined to be to perform data modification recognition processing, thereby improving the efficiency of the update and iteration processing of the speech recognition model. If not, the generation and management method of the ward round records of the disease type is determined to be to not perform data modification recognition processing.

[0140] Correlation coefficient: This refers to the proportion of information items shared by this disease type with all other disease types under control, relative to the total number of core information items of this disease type itself. This coefficient reflects the degree of similarity in information item composition between this disease type and already controlled high-risk disease types.

[0141] If a disease type highly overlaps with a disease type already under control in core information items (e.g., a correlation coefficient greater than 0.3), it indicates that they face similar identification challenges. Even if the disease type itself currently does not have a risk category, it may still pose a potential risk, or its data can provide a comparative reference for optimizing the model for controlling the disease type. Analyzing its modified data can provide more multi-dimensional information for model improvement. Therefore, disease types with high correlation coefficients are also worth including in the processing scope.

[0142] Scenario 1: The number of controls is moderate (meets requirements): Assuming the S33 judgment result is: the generated controlled disease types are C (pneumonia) and E (heart failure), a total of 2, which is exactly equal to the upper limit (not exceeded), so the requirements are met, and we proceed to the case-by-case discussion.

[0143] Step 1 (S41): Determine the generated disease types as C and E.

[0144] Step 2 (S42): Determine the risk information items for each disease type (Class I deviation risk type information items): Step 3 (S43): Determine whether each uncontrolled disease type (A, B, D) requires data modification.

[0145] For disease type A (coronary heart disease): Scenario 1: Risk information items (medication adjustment, vital signs) exist in A, so it is directly determined that A needs to be identified and processed for data modification.

[0146] For type B of disease (diabetes): Scenario 1: Risk information items exist in B, therefore it is directly determined that B needs to undergo data modification identification processing.

[0147] At this point, the number of disease types that need to be processed has been determined to be 2 (A and B), reaching the preset disease type count threshold T_disease_count=2.

[0148] For disease type D (hypertension): Since risk information items exist in D, it should be processed directly. Therefore, D is also directly identified as the type of data that needs to be modified. Ultimately, A, B, and D are all processed.

[0149] Scenario 2: Too many controls (not meeting requirements): Assuming the S33 judgment result is: the generated controlled disease types are A, C, and E, a total of 3 (i.e., coronary heart disease is also controlled), which exceeds the upper limit T_control_max=2, so "the requirements are not met", and the restriction is relatively severe at this time.

[0150] Therefore, in Scenario 2, although three disease types (A, C, and E) are generated for control, B and D are not yet controlled. Based on the principle of "more severe control," data modification is required for both B and D, without needing the screening process described in scenarios 1-3.

[0151] Final result: Scenario 1 (Moderate Control): A, B, and D are all handled.

[0152] Scenario 2 (Too much control): Both B and D are handled.

[0153] If there is a disease type without risk information in Scenario 1 (such as the assumed F), then further judgment should be made according to Cases 2 and 3.

[0154] Through the above steps, hospitals can dynamically adjust their data collection strategies based on the severity of the control measures: when the control scope is too large, comprehensive data collection is used to accelerate model iteration; when the control scope is controllable, high-value data sources are precisely selected to optimize resource allocation.

[0155] This step uses the number of disease types under control as the first dividing line. When there are too many control measures, comprehensive data collection is initiated, which reflects a positive response to the core requirement of "lifting control measures as soon as possible". When control measures are under control, disease types containing high-risk information items are prioritized for data collection through multi-level condition judgment, ensuring that model optimization resources are focused on the most valuable erroneous data.

[0156] Example 2 Secondly, such as Figure 4 As shown, this application provides an intelligent voice interaction and automated document generation system, employing the aforementioned intelligent voice interaction and automated document generation method, specifically including: Template generation trigger and intelligent assembly: The system receives nursing level addition or change information pushed by the HIS system, or responds to daily scheduled tasks, and automatically triggers the ward round template generation process; the system automatically retrieves the corresponding template from the template library for intelligent assembly based on the patient's nursing level, department and diagnosis results, and generates a personalized initial version of the ward round record. Multi-source data integration and intelligent filling: Integrates multiple data sources, including automatically extracting current valid medical orders from the medical order system, automatically importing positive signs and nursing problems that need continuous monitoring, and also importing basic information such as patient name, bed number, hospital number, and allergy history. Combined with the built-in professional medical knowledge base, semantic mapping is performed on the extracted sign data, and structured descriptions and communication records between nurses and patients are generated in the corresponding positions of the personalized initial ward round record. Editing, reviewing, and archiving: The system provides a rich text editing interface, supporting nurses to modify text and voice, select items, and select standardized terms for common symptoms / signs. The system saves the automatically generated initial version and the final version modified by the nurse. After review and confirmation, the record is submitted, locked, and synchronously archived to the electronic medical record system.

[0157] System Deployment and Initialization: The system of this invention is deployed in the intranet environment of a medical institution to ensure that the mobile terminal can run normally in a pure intranet environment. Basic templates, specialty templates, and disease-specific templates are pre-set in the template library: Basic templates, according to the national / hospital "Graded Nursing Standards," include basic observation items such as vital signs and level of consciousness for special, first, second, and third-level nursing care. Specialty templates are designed for different departments such as cardiology, neurosurgery, and obstetrics, adding specialty observation items to the corresponding nursing level's basic template. For example, cardiology adds observation items such as heart rate, heart rhythm, and signs of heart failure, while obstetrics adds observation items such as fundal height, abdominal circumference, and fetal heart rate. Disease-specific templates are designed for different diseases such as acute myocardial infarction, diabetes, and postoperative patients, setting specific assessment focuses. For example, the acute myocardial infarction template includes assessment items such as myocardial enzymes, electrocardiogram, and medication response, while the diabetes patient template focuses on assessment items such as blood glucose and foot skin. Simultaneously, data interface integration with the HIS system, medical order system, and electronic medical record system is completed to ensure smooth data transmission.

[0158] Template generation and intelligent assembly example: Taking patient XX as an example, after receiving the patient's Level 1 nursing care information pushed by the HIS system, the system automatically triggers the template generation process. Based on the patient's nursing care level (Level 1 nursing care), department (cardiology), and primary diagnosis (acute myocardial infarction), the system retrieves the Level 1 nursing care basic template, cardiology specialty template, and acute myocardial infarction disease template from the template library for intelligent assembly. The Level 1 nursing care basic template provides 12 basic observation items, such as vital signs and level of consciousness; the cardiology specialty template provides 8 specialty observation items, such as heart rate, heart rhythm, and signs of heart failure; and the acute myocardial infarction disease template provides 6 specific observation items, such as myocardial enzymes, electrocardiogram, and medication response. Combined with the communication records between the nursing staff and the patient, a personalized initial ward round record for the patient is generated after assembly.

[0159] Intelligent data filling example: The system extracts Zhang Ming's current valid medical orders. If the orders include "continuous ECG monitoring" and "blood pressure monitoring Q6H", the blood pressure record item is highlighted in the template and a blood pressure-specific record template is generated. The system retrieves the previous ward round record and automatically inserts positive signs and nursing issues that require continuous monitoring, such as "left lower extremity edema still needs observation". The system compares the current and previous vital sign data and generates a trend description such as "body temperature decreased by 1.2℃ compared to yesterday". At the same time, the system inserts the patient's name, bed number, hospital number, allergy history, and other basic information. Based on the above data, structured text descriptions are generated in the corresponding positions of the personalized initial ward round record, such as: "Patient's consciousness [auto-fill in the previous record 'awake'], chief complaint: wound pain lessened compared to yesterday. Physical examination: T: [auto-fill in the latest temperature 36.5℃]℃, P: [auto-fill in the latest pulse 78 bpm] bpm, R: [auto-fill in the latest respiration 18 breaths / min] bpm, BP: [auto-fill in the latest blood pressure 135 / 85 mmHg] mmHg. The dressing on the left lower extremity wound is dry, with no exudate. [Auto-fill in the relevant medical orders: aspirin, clopidogrel, atorvastatin orally] treatment is administered as prescribed." All automatically generated content is presented in gray font and marked "Pending confirmation".

[0160] Examples of editing, reviewing, and archiving: View Zhang Ming's personalized initial ward round record through the system's rich text editing interface and review the automatically generated content. If the "Cardiac Function Classification" is not automatically filled, supplement it using the standardized terminology selection function or by voice selection of "NYHA Class II". The system records all modifications in real time, including the modifier, modification time (2023-05-15 09:30), and modification content. After the nurse reviews and confirms that everything is correct, click the submit button. The record is locked and synchronously archived to the electronic medical record system. At the same time, the system saves both the automatically generated initial version and the revised final version for future reference.

[0161] Example 3 Thirdly, this application provides a smart chip for the aforementioned intelligent voice interaction and document automation generation system, specifically comprising: The data acquisition interface is configured to acquire multimodal heterogeneous data from the user during the training process, and the storage unit is configured to store preset threshold parameters and program instructions. A processing unit, coupled to the data acquisition interface and the storage unit, is configured to execute the program instructions to achieve this.

[0162] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0163] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0164] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for intelligent voice interaction and automated document generation, characterized in that, Specifically, it includes: Using the generated text of ward round records based on speech recognition results, the modified data in ward round records for different disease types is determined. The modified data of each nursing staff member is used to determine the identification deviation information items of the nursing staff member in the disease type. The similarity between the identification deviation information items of the nursing staff member in the disease type and different nursing staff members is used. Combined with the identification deviation information item data of the nursing staff member, the identification risk disease type in the disease type is determined. Using the identification risk disease type in the disease type and the modified data of the nursing staff member, the identification processing scheme of the modified data is determined. Using the aforementioned identification processing scheme, semantic recognition deviation data under different information items are determined, and combined with the association between the information items and different identification risk disease types, the text generation control strategy for the identification risk disease types is determined. Based on different text generation control strategies for identifying risky disease types, and combined with the degree of correlation between the ward round records of the disease type and the risky disease type in different information items, the text generation method for the ward round records of the disease type is determined.

2. The intelligent voice interaction and automated document generation method as described in claim 1, characterized in that, The generated text of the ward round record is the result of the nursing staff filling in the information items of the ward round record based on the voice recognition results.

3. The intelligent voice interaction and automated document generation method as described in claim 1, characterized in that, The modifications in the ward round record are determined based on the changes made by the nursing staff after the information items in the ward round record have been filled out.

4. The intelligent voice interaction and automated document generation method as described in claim 1, characterized in that, The method for determining the risk-prone disease type among the aforementioned disease types is as follows: Using modification data from various nursing staff, determine the number of times each nursing staff member made modifications to different information items, and determine the identification deviation information items for each nursing staff member based on the number of modifications. By using the similarity between the identification deviation information item of the nursing staff in the disease type and different nursing staff, the nursing staff to which the information item belongs to the identification deviation information item are determined; by using the nursing staff to which the information item belongs to the identification deviation information item, the deviation risk information item in the information item is determined. Based on the modification data of each nurse in the deviation risk information item and the identification deviation information item, determine whether the disease type belongs to the identified risk disease type.

5. The intelligent voice interaction and automated document generation method as described in claim 4, characterized in that, The identification deviation information item is the information item whose number of modifications is greater than a preset modification number threshold.

6. The intelligent voice interaction and automated document generation method as described in claim 4, characterized in that, The deviation risk information item refers to the information item where the number of nursing staff who fall under the category of deviation information items does not meet the requirements.

7. The intelligent voice interaction and automated document generation method as described in claim 1, characterized in that, The method for determining the text generation method of the ward round records for the aforementioned disease type is as follows: Using different text generation control strategies for identifying risky disease types, the generation control disease type among the identified risky disease types is determined; Based on the semantic recognition deviation data of the ward round records for the disease type in different information items, determine a type of deviation risk information item in the information items of the disease type, and use it as a risk information item; Based on the identified risk disease types, the risk information items in the disease types, and the degree of correlation with the information items of the production management disease types, a method for generating the text of the ward round records for the disease types is determined.

8. The intelligent voice interaction and automated document generation method as described in claim 7, characterized in that, The generated and controlled disease types are the risk disease types that require text generation and control processing.

9. An intelligent voice interaction and automated document generation system, employing the intelligent voice interaction and automated document generation method according to any one of claims 1-8, specifically comprising: Template generation trigger and intelligent assembly: The system receives nursing level addition or change information pushed by the HIS system, or responds to daily scheduled tasks, and automatically triggers the ward round template generation process; The system automatically retrieves the corresponding template from the template library based on the patient's nursing level, department, and diagnosis results, and intelligently assembles them to generate a personalized initial version of the ward round record; Multi-source data integration and intelligent filling: Integrates multiple data sources, including automatically extracting current valid medical orders from the medical order system, automatically importing positive signs and nursing problems that need continuous monitoring, and also importing basic information such as patient name, bed number, hospital number, and allergy history. Combined with the built-in professional medical knowledge base, semantic mapping is performed on the extracted sign data, and structured descriptions and communication records between nurses and patients are generated in the corresponding positions of the personalized initial ward round record. Editing, reviewing, and archiving: The system provides a rich text editing interface, supporting nurses to modify text and voice, select items, and select standardized terms for common symptoms / signs. The system saves the automatically generated initial version and the final version modified by the nurse. After review and confirmation, the record is submitted, locked, and synchronously archived to the electronic medical record system.

10. A smart chip, used in the intelligent voice interaction and document automation generation system of claim 9, characterized in that, Specifically, it includes: The data acquisition interface is configured to acquire multimodal heterogeneous data from the user during the training process, and the storage unit is configured to store preset threshold parameters and program instructions. A processing unit, coupled to the data acquisition interface and the storage unit, is configured to execute the program instructions to achieve this.

Images