A management report text retelling method and deep learning system based on digital medical care

Abstract

The present invention discloses a management report text retelling method and deep learning system based on digital medicine, which belongs to the field of text processing technology. The present invention proposes an innovative solution for the interpretation of hospital management reports and the risk warning of adverse events. Through deep learning technology, it is possible to deeply explore the text semantics of hospital management reports, accurately capture core information, and provide a solid foundation for subsequent adverse event prediction and risk analysis. At the same time, the present invention can automatically identify and classify adverse events in reports, provide hospital managers with a clear and structured overview of adverse events, and significantly improve management efficiency and risk identification acuity. In addition, the present invention generates a local retelling annotation suggestion feature containing specific details of adverse events through a text retelling annotation algorithm, which is convenient for managers to quickly grasp the core information of the event. Finally, the present invention provides hospital managers with a comprehensive and objective risk assessment tool to help reduce the incidence of adverse events.

Description

Technical Field

[0001] The present invention belongs to the technical field of text processing, and in particular relates to a management report text retelling method and a deep learning system based on digital medicine. Background Art

[0002] In the context of the rapid development of digital medicine, hospital management reports are important documents that reflect hospital operations, patient service quality, and adverse events. Their interpretation and analysis are crucial for hospital managers. However, traditional hospital management report processing methods often rely on manual reading and analysis, which is not only inefficient, but also difficult to accurately mine potential adverse event risk points from massive texts.

[0003] Faced with this challenge, how to achieve intelligent processing of hospital management reports, automatically identify and classify adverse events, and provide hospital managers with a clear and structured risk overview has become a technical problem that needs to be solved urgently. Summary of the invention

[0004] The present invention provides a management report text repetition method and a deep learning system based on digital medicine, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0005] The present invention provides a management report text paraphrase method based on digital medical care, which is applied to a deep learning system. The method comprises: performing text semantic mining on a hospital management report to be processed to obtain the management report text semantics corresponding to the hospital management report to be processed under multiple target semantic attentions; determining the target adverse event label set corresponding to the management report text semantics of the hospital management report to be processed under each target semantic attention; encoding the key event state of the adverse event label set corresponding to the hospital management report to be processed in each target semantic attention, loading it into a text paraphrase annotation algorithm for paraphrase annotation analysis, and obtaining the local paraphrase annotation suggestion feature of the hospital management report to be processed under each target semantic attention; and determining the paraphrase annotation decision vector of the hospital management report to be processed by using the local paraphrase annotation suggestion feature under each target semantic attention.

[0006] The present invention provides a deep learning system, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above-mentioned method.

[0007] The present invention provides a readable storage medium, on which a program or instruction is stored, and the program or instruction implements the steps of the above method when executed by a processor.

[0008] The present invention proposes an innovative solution for the interpretation of hospital management reports and the early warning of adverse event risks. Through deep learning technology, the present invention can deeply explore the text semantics of hospital management reports, accurately capture core information, and provide a solid foundation for subsequent adverse event prediction and risk analysis. At the same time, the present invention can automatically identify and classify adverse events in reports, provide hospital managers with a clear and structured overview of adverse events, and significantly improve management efficiency and risk identification acuity. Furthermore, the present invention generates a local repetition annotation suggestion feature containing specific details of adverse events through a text repetition annotation algorithm, which facilitates managers to quickly grasp the core information of the event. Finally, the present invention provides hospital managers with a comprehensive and objective risk assessment tool to help reduce the incidence of adverse events. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 The present invention provides a flowchart of a method for repeating a management report text based on digital medicine.

[0010] Figure 2 A schematic diagram of the structure of a deep learning system provided by the present invention. DETAILED DESCRIPTION

[0011] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0012] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0013] Figure 1 A method for retelling a management report text based on digital medicine is shown, which is applied to a deep learning system. The method includes the following steps 110 to 140.

[0014] In the field of medical management, accurate understanding and analysis of hospital management reports are crucial to improving the quality of medical services and patient safety. Traditional manual review methods are not only time-consuming and laborious, but may also lead to information omissions or misunderstandings due to human factors. To this end, the present invention aims to automatically process hospital management reports, extract key information from them, and provide repetition annotation suggestions to help managers identify and respond to potential adverse events more efficiently.

[0015] Step 110: The deep learning system performs text semantic mining on the hospital management report to be processed to obtain the management report text semantics corresponding to the hospital management report to be processed under multiple target semantic attentions.

[0016] First, the deep learning system receives a pending hospital management report. The pending hospital management report contains various details of hospital operations, such as patient care, medical equipment maintenance, staff training, etc., as well as possible adverse event records. The deep learning system then uses deep learning models, especially models based on the Transformer architecture, to perform text semantic mining on the pending hospital management report.

[0017] The Transformer model, due to its powerful self-attention mechanism, can effectively capture long-distance dependencies in text and is suitable for processing complex hospital management reports. During the debugging phase, the Transformer model was debugged with a large number of annotated hospital management reports and learned how to map text to multiple preset target semantic attentions, such as "patient safety", "quality of medical services", "operational efficiency", etc. For the hospital management reports to be processed, the Transformer model can output the text semantics corresponding to each target semantic attention, that is, the semantic representation of the report in different management dimensions.

[0018] Step 120: The deep learning system determines the target adverse event label set corresponding to the management report text semantics of the hospital management report to be processed under each target semantic attention.

[0019] Next, the deep learning system further determines the target adverse event label set corresponding to each target semantic attention of the report based on the text semantics obtained in step 110. This step relies on a pre-built adverse event classification system (adverse event classification knowledge base), which covers various adverse events that may be encountered in hospital management, such as "medical error", "equipment failure", "infection control failure", etc.

[0020] The deep learning system identifies adverse events mentioned in the report by comparing text semantics with the adverse event classification system and classifies them into corresponding labels. For example, under the semantic attention of "patient safety", the deep learning system may identify adverse event labels such as "fall event" and "medication error".

[0021] Step 130: The deep learning system encodes the key event status of the hospital management report to be processed in the adverse event label set corresponding to each target semantic attention, loads it into the text paraphrase annotation algorithm for paraphrase annotation analysis, and obtains the local paraphrase annotation suggestion features of the hospital management report to be processed under each target semantic attention.

[0022] With the adverse event label set, the deep learning system enters the paraphrase annotation analysis stage. The purpose of this step is to generate specific paraphrase annotation suggestions for each adverse event in the report to help managers better understand the nature, cause and consequences of the event.

[0023] The deep learning system first determines the key event status codes for each adverse event label set, including the time, location, people involved, and execution measures of the event. These codes are then loaded into the text paraphrase annotation algorithm, which is also based on deep learning, especially the sequence-to-sequence (Seq2Seq) learning framework, and can generate natural and fluent paraphrase text.

[0024] Through the paraphrase annotation algorithm, the deep learning system can generate one or more paraphrase annotation suggestions for each adverse event. These suggestions are presented in a structured form, which makes it easier for managers to quickly grasp the core information of the event. For example, for the "medication error" event, the deep learning system may generate the following paraphrase annotation: "On March 15, 2023, a medication error occurred in the internal medicine ward. Patient Zhang took an excessive amount of antihypertensive drugs by mistake, resulting in a sharp drop in blood pressure. After timely treatment, it returned to normal."

[0025] Step 140: The deep learning system uses the local paraphrase annotation suggestion features under each of the target semantic attentions to determine the paraphrase annotation decision vector of the hospital management report to be processed.

[0026] Finally, the deep learning system uses the local paraphrase annotation suggestion features under each target semantic attention to determine the paraphrase annotation decision vector for the hospital management report to be processed. This step is an integration of all the previous analysis results, aiming to form a comprehensive paraphrase annotation report for managers to use in decision-making.

[0027] The paraphrase annotation decision vector is a high-dimensional vector, where each dimension corresponds to a specific paraphrase annotation suggestion or its attribute (such as importance, urgency). The deep learning system constructs this decision vector by comprehensively considering the paraphrase annotation suggestions under different semantic attentions, as well as the correlation and conflict between them. For example, if the same adverse event is mentioned under multiple semantic attentions and is labeled as high importance, then the corresponding dimension of this event in the decision vector will be given a higher weight.

[0028] Finally, the deep learning system generates a detailed paraphrase annotation report based on the paraphrase annotation decision vector. The paraphrase annotation report not only lists all the identified adverse events and their paraphrase annotation suggestions, but also sorts and classifies the events according to their importance and urgency. The paraphrase annotation report can greatly help hospital managers quickly understand the key information in the report, identify potential risk points, and formulate corresponding improvement measures, thereby improving the overall management level and service quality of the hospital.

[0029] Through the above steps, key adverse event information can be effectively mined from complex hospital management reports, and structured repetition annotation suggestions can be generated for these events. This not only improves the processing efficiency of hospital management reports, but also enhances managers' ability to identify and respond to potential risk points, which is of great significance for improving the quality of medical services and patient safety.

[0030] It can be understood that the above-mentioned management report text semantics, key event status coding, local paraphrase annotation suggestion features and paraphrase annotation decision vector are the core invention points of the present invention. Based on this, the following is an introduction and explanation through relevant examples.

[0031] First, the text semantics of the management report is obtained by mining the text semantics of the hospital management report through the above-mentioned deep learning system, especially the Transformer model. This step maps the report text to multiple preset target semantic attentions, such as "patient safety", "quality of medical services", etc. Numerically, this can be represented by a high-dimensional vector, where each dimension corresponds to a specific semantic attention, and the value of the vector reflects the semantic strength or relevance of the report under that semantics. For example, a three-dimensional text semantic vector [0.8, 0.4, 0.6] can represent that the report has a high semantic strength on "patient safety", a low semantic strength on "quality of medical services", and a medium level on another dimension (such as "operational efficiency").

[0032] Secondly, the critical event status code is a series of numerical codes generated for each adverse event mentioned in the report, which captures the key status information of the event, such as the time of occurrence, location, and personnel involved. Numerically, it can also be represented by a vector, where each dimension corresponds to a specific state attribute, and the value of the vector gives the specific value or code of the attribute. For example, a four-dimensional critical event state code vector [20230315, 2, 1001, 1] can represent that an adverse event occurred on March 15, 2023, the location code is 2 (such as the internal medicine ward), the personnel code is 1001 (such as Doctor Zhang), and the last dimension can represent the severity or urgency of the event.

[0033] Then, the local paraphrase annotation suggestion feature is generated by the text paraphrase annotation algorithm based on the key event state encoding, which reflects the core content of the paraphrase annotation suggestion. Numerically, this can be represented by a series of vectors, each of which corresponds to a different aspect of a paraphrase annotation suggestion, such as event description, cause analysis, consequence assessment, etc. For example, for the "medication error" event, two vectors may be generated: ["March 15, 2023, Internal Medicine Ward", "Patient Zhang accidentally took an overdose of antihypertensive drugs"] and ["Blood pressure dropped sharply", "Restored stability after timely treatment"], describing the basic information and consequences of the event respectively.

[0034] Finally, the paraphrase decision vector is a high-dimensional vector formed by integrating all local paraphrase suggestion features, which is used to represent the paraphrase decision of the entire hospital management report. Numerically, each dimension of this vector corresponds to a specific paraphrase suggestion or its attribute, and the value of the vector reflects the importance and urgency of the suggestion or attribute. For example, a five-dimensional paraphrase decision vector [1, 0.8, 0.6, 0.3, 0.9] can represent that five important adverse events were identified in the report, and they were assigned different weights to reflect their importance and urgency.

[0035] It can be seen that the composition and application of comprehensive management report text semantics, key event status encoding, local paraphrase annotation suggestion features and paraphrase annotation decision vectors enable the deep learning system to effectively mine key adverse event information from complex hospital management reports and generate structured paraphrase annotation suggestions for these events, thereby greatly improving the processing efficiency of hospital management reports and managers' ability to identify and respond to potential risk points.

[0036] It is worth noting that steps 120 to 140 are intended to automatically process hospital management reports, extract key information therefrom, and provide structured repetition annotation suggestions for managers.

[0037] First, step 120 determines the target adverse event label set corresponding to the hospital management report to be processed under multiple target semantic attentions. This step is the key bridge connecting text semantic mining and subsequent repetition annotation analysis. After the deep learning system completes the text semantic mining of the hospital management report, the deep learning system obtains a report interpretation rich in text semantics under multiple target semantic attentions. However, this interpretation is still relatively abstract and broad. It covers the semantic representation of the report in multiple dimensions such as "patient safety", "quality of medical services", and "operational efficiency", but does not specifically point out the adverse events hidden under these semantics.

[0038] Therefore, the task of step 120 is to further refine this interpretation and convert it into a specific set of adverse event labels. This conversion process relies on a pre-built adverse event classification system, which is a comprehensive and detailed framework that covers various adverse events that may be encountered in hospital management, such as "medical error", "equipment failure", "infection control failure", etc. By comparing the text semantics with this classification system, the deep learning system can identify the adverse events mentioned in the report and accurately classify them under the corresponding labels.

[0039] For example, under the semantic attention of "patient safety", the deep learning system may identify adverse event labels such as "fall event" and "medication error" from the report. These labels not only reflect the specific content of the report in terms of patient safety, but also provide clear direction and focus for subsequent retelling annotation analysis.

[0040] Next, step 130 encodes the key event status of the adverse event label set corresponding to each target semantic attention in the hospital management report to be processed, and loads it into the text paraphrase annotation algorithm for paraphrase annotation analysis. This step is the core link in the technical solution, which is responsible for converting abstract adverse event labels into specific and structured paraphrase annotation suggestions.

[0041] To achieve this transformation, the deep learning system first encodes the key event states for each adverse event label set. These states include the time, location, people involved, and implementation measures of the event. They are the core elements of adverse events and are also the information that must be included in the paraphrase annotation suggestions. By encoding these states, the deep learning system provides rich and orderly input data for the paraphrase annotation algorithm.

[0042] These codes are then loaded into the text paraphrase annotation algorithm. This algorithm is also based on deep learning, specifically the sequence-to-sequence (Seq2Seq) learning framework. It can receive the encoded key event states as input and generate natural and fluent paraphrases as output. Through this algorithm, the deep learning system can generate one or more paraphrases for each adverse event. These suggestions are presented in a structured form, making it easier for managers to quickly grasp the core information of the event.

[0043] For example, for a "medication error" event, the deep learning system may generate the following paraphrase annotation: "On March 15, 2023, a medication error occurred in the internal medicine ward. The patient Zhang's blood pressure dropped sharply due to an overdose of antihypertensive drugs. After timely treatment, it returned to normal." This paraphrase annotation suggestion not only includes the basic information of the incident (time, place, and people involved), but also describes the consequences of the incident and the measures taken, providing managers with a comprehensive and specific overview of the incident.

[0044] Finally, step 140 uses the local paraphrase annotation suggestion features under each target semantic attention to determine the paraphrase annotation decision vector of the hospital management report to be processed. This step is an integration and sublimation of all the previous analysis results, and it aims to form a comprehensive paraphrase annotation report for managers to use in decision-making.

[0045] The paraphrase tagging decision vector is a high-dimensional vector, where each dimension corresponds to a specific paraphrase tagging suggestion or its attribute (such as importance, urgency). In order to construct this vector, the deep learning system needs to comprehensively consider the paraphrase tagging suggestions under different semantic attentions, as well as the correlation and conflict between them. This is a complex and delicate process, which requires the deep learning system to have not only powerful computing power, but also deep medical knowledge and logical reasoning ability.

[0046] In the process of constructing the paraphrase annotation decision vector, the deep learning system evaluates the importance and urgency of each paraphrase annotation suggestion. This is achieved by analyzing the text content in the report, the nature of the adverse event, and the actual needs of hospital management. For example, if the same adverse event is mentioned in multiple semantic attentions and is annotated as high importance, then the corresponding dimension of this event in the decision vector will be given a higher weight.

[0047] Finally, the deep learning system generates a detailed paraphrase annotation report based on the paraphrase annotation decision vector. This report not only lists all the identified adverse events and their paraphrase annotation suggestions, but also sorts and classifies the events according to their importance and urgency. It presents the key information in the hospital management report in a clear and structured way, allowing managers to quickly understand the main content of the report, identify potential risk points, and formulate corresponding improvement measures.

[0048] It can be seen that based on steps 120 to 140, the key information in the report is first extracted through text semantic mining, and then converted into a specific adverse event label set. Next, the text paraphrase annotation algorithm is used to generate structured paraphrase annotation suggestions to provide managers with a detailed event overview. Finally, a comprehensive paraphrase annotation report is generated by constructing a paraphrase annotation decision vector for managers to use in decision-making. In this way, not only the processing efficiency of hospital management reports is improved, but also the managers' ability to identify and respond to potential risk points is enhanced, which is of great significance to improving the quality of medical services and patient safety.

[0049] In practical applications, the above technical solution can be widely used in various types of hospitals and medical institutions. It can help managers process and analyze hospital management reports more efficiently and accurately, and timely discover and solve potential problems and risks. At the same time, it can also provide medical institutions with a comprehensive and objective evaluation tool to help them continuously improve and enhance the quality of medical services.

[0050] In addition, this technical solution is highly scalable and customizable. According to different hospital management needs and report formats, the deep learning system can adjust and optimize the deep learning model and text paraphrase annotation algorithm to achieve more accurate and efficient processing results. At the same time, the deep learning system can also combine this technical solution with other medical information systems to achieve more comprehensive and intelligent hospital management.

[0051] In summary, the invention has brought significant beneficial effects to the interpretation of hospital management reports and the early warning of adverse event risks. First, by conducting in-depth text semantic mining on the hospital management reports to be processed, the invention can accurately capture the core text semantics in the management reports under multiple target semantic attention. This process not only improves the accuracy of information extraction, but also lays a solid foundation for subsequent adverse event prediction and risk analysis.

[0052] Next, the present invention can determine the corresponding target adverse event label set under each target semantic attention based on the mined text semantics. The realization of this step indicates that the present invention can automatically identify and classify adverse events in reports, providing hospital managers with a clear and structured adverse event overview, greatly improving management efficiency and risk identification acuity.

[0053] Furthermore, the present invention loads the key event status codes of the adverse event tag set into the text paraphrase annotation algorithm to perform paraphrase annotation analysis. This innovative application enables the present invention to generate local paraphrase annotation suggestion features under each target semantic attention. These features not only contain the specific details of the adverse event, but are also presented in a structured form, which makes it easier for managers to quickly grasp the core information of the event and formulate effective response measures.

[0054] Finally, by integrating the local paraphrase annotation suggestion features under each target semantic attention, the present invention can determine the paraphrase annotation decision vector of the hospital management report to be processed. The generation of this vector provides a comprehensive and objective risk assessment tool for hospital managers, helping them to consider the risk factors of adverse events more scientifically and comprehensively, thereby effectively reducing the incidence of adverse events at the health system level.

[0055] This design not only realizes the intelligent processing of hospital management reports and the accurate prediction of adverse events, but also brings forward-looking changes to hospital management and policy design. Through the application of this invention, hospitals will be able to more efficiently and accurately identify and respond to potential adverse event risks, providing strong support for improving the quality of medical services and patient safety.

[0056] Furthermore, in some optional embodiments, the step of performing text semantic mining on the hospital management report to be processed is implemented through a Transformer model, and the Transformer model is configured with a self-attention mechanism to capture long-distance dependencies in the hospital management report to be processed.

[0057] The method of determining the target adverse event label set corresponding to the management report text semantics of the hospital management report to be processed under each target semantic attention includes: comparing and analyzing the management report text semantics with a preset adverse event classification knowledge base to obtain the target adverse event label set corresponding to the management report text semantics under each target semantic attention.

[0058] The key event status codes of the adverse event label sets corresponding to the target semantic attentions of the hospital management report to be processed are loaded into the text paraphrase annotation algorithm for paraphrase annotation analysis to obtain the local paraphrase annotation suggestion features of the hospital management report to be processed under the target semantic attentions, including: determining the key event status codes of each adverse event label set; wherein the key event status codes include the time, location, involved personnel and implementation measures of the event; and performing structured paraphrase processing on the key event status codes through a text paraphrase annotation algorithm based on a Seq2Seq learning framework to obtain the local paraphrase annotation suggestion features of the hospital management report to be processed under the target semantic attentions.

[0059] The method of using the local paraphrase annotation suggestion features under each of the target semantic attentions to determine the paraphrase annotation decision vector of the hospital management report to be processed includes: generating the paraphrase annotation decision vector of the hospital management report to be processed based on the correlation and conflict between the local paraphrase annotation suggestion features under each of the target semantic attentions.

[0060] In a specific example, the relevant technical solutions for hospital management and policy design based on the hospital management report processing based on the above neural network algorithm / model are introduced as follows.

[0061] 1. Intelligent processing and accurate prediction of adverse events

[0062] 1) Text Semantic Mining - Transformer Model Example

[0063] In the processing of hospital management reports, the Transformer model is first used for text semantic mining. The self-attention mechanism of the Transformer model can be expressed as:

[0064] (Attention(Q,K,V)=text{softmax}(frac{QK^T}{sqrt{d_k}})V)

[0065] Where (Q), (K), (V) are the query vector, key vector, and value vector respectively, and (d_k) is the dimension of the key vector. Through this mechanism, the model can capture long-distance dependencies in the text.

[0066] For example, for a hospital management report containing information such as patient care and medical equipment maintenance, the Transformer model maps it to target semantic attention such as "patient safety", "medical service quality", and "operational efficiency". Assuming that the report contains a description of the medication process in patient care, the model can accurately determine its relevance in various semantic dimensions based on semantic associations, such as the semantic intensity in the dimension of "patient safety" may be higher.

[0067] 2) Determine the adverse event label set - matching based on the classification system

[0068] The deep learning system relies on a pre-built adverse event classification system, which can be seen as a knowledge base that contains the definitions and characteristics of various adverse events.

[0069] For example, when processing the text semantics in a report, if the text mentions "the patient had an abnormal reaction after injecting the drug", the system accurately classifies it by matching it with the "adverse drug reaction" label in the adverse event classification system. This matching process can be regarded as a rule-based classification process. Although more complex neural network layers may be used in actual deep learning systems to achieve this matching, it is essentially to establish a connection between text semantic features and adverse event label features.

[0070] 3) Paraphrase Annotation Analysis - Seq2Seq Learning Framework Example

[0071] For the determined adverse event label set, the system uses a text paraphrase annotation algorithm based on a sequence-to-sequence (Seq2Seq) learning framework. The Seq2Seq model consists of an encoder and a decoder.

[0072] The encoder takes the key event state encoding (such as the encoding of the time, location, and personnel involved in the event) as input, sets the key event state encoding vector to (X=[x_1, x_2, cdots, x_n]), and after processing by a multi-layer neural network layer, obtains the hidden state (H=[h_1, h_2, cdots, h_m]). The decoder generates a retelling annotation suggestion based on the hidden state. For example, for the event "a patient had an allergic reaction to an injection of a drug in a certain ward", the encoder processes the relevant state encoding, and the decoder may generate a retelling annotation such as "On April 10, 2023, patient Li in Ward 3 had an allergic reaction after injecting a certain drug, and the medical staff immediately performed anti-allergic treatment."

[0073] 2. Proactive Changes in Hospital Management and Policy Design

[0074] 1) Improve management efficiency and identify risks

[0075] In terms of hospital management, through the above process, managers can quickly obtain structured adverse event information. For example, a large hospital generates a large number of management reports every week. In the traditional way, manual review of these reports is time-consuming and prone to missing information.

[0076] Using the technology of the present invention, managers can quickly locate high-risk adverse events, such as the frequently occurring "medical device failure" event. According to the repetition annotation decision vector, for example, vector (D = [d_1, d_2, cdots, d_k]), where (d_i) represents the importance weight of different adverse events, if the (d_i) value related to medical device failure is high, managers can arrange equipment maintenance and update plans in time to improve management efficiency and reduce risks.

[0077] 2) Policy design basis

[0078] At the policy design level, hospitals can formulate relevant policies based on the statistics and analysis results of adverse events. For example, if the incidence of "patient falls" is high over a period of time, through the analysis of management reports, it is found that the main reasons are unreasonable layout of ward facilities and untimely inspections by nursing staff.

[0079] Hospitals can formulate policies such as re-planning ward layouts, adding protective facilities, and adjusting nursing staff's shift system to increase the frequency of rounds. This data-driven policy design is proactive and can prevent potential adverse events from happening, rather than taking remedial measures after the event occurs.

[0080] It can be seen that through the application of specific neural network algorithms such as the Transformer model and the Seq2Seq framework, the intelligent processing of hospital management reports and the accurate prediction of adverse events have been achieved. At the same time, it has also brought about a forward-looking transformation from passive response to active prevention for hospital management and policy design.

[0081] In some preferred embodiments, the debugging step of the above-mentioned text retelling annotation algorithm includes: obtaining the management report text semantics of multiple hospital management report examples under the same original semantic attention, performing event feature clustering on each of the management report text semantics belonging to the same original semantic attention, and obtaining the adverse event label set corresponding to each of the management report text semantics, and the key event status code of each of the adverse event label sets; based on the adverse event label set contained in the same original semantic attention, and the number of positive debugging examples and the number of negative debugging examples of the hospital management report examples to which the management report text semantics contained in each of the adverse event label sets belong, determining the risk trend prediction of each of the original semantic attentions. weight measurement; from the original semantic attention, sample the target semantic attention whose risk trend prediction weight meets the trend prediction requirements; encode the key event status of each hospital management report example in the adverse event label set corresponding to each target semantic attention, load it into the original text retelling annotation algorithm, and obtain the possibility of distinguishing whether each hospital management report example belongs to a positive debugging example or a negative debugging example; use the debugging example of each hospital management report example to distinguish annotations and the discrimination possibility, optimize the algorithm weight of the original text retelling annotation algorithm, and obtain a text retelling annotation algorithm, which is used to process the local retelling annotation suggestion features of the hospital management report to be processed under each target semantic attention.

[0082] It is understandable that the debugging step of the text paraphrase annotation algorithm is used to ensure the effectiveness and accuracy of the algorithm. The process first involves obtaining the management report text semantics of multiple hospital management report examples under the same original semantic attention. These management report text semantics represent the core information about the description of adverse events in the hospital management report and are the basis for subsequent analysis.

[0083] After obtaining the semantics of these management report texts, the next step is to cluster the event features of each management report text semantic that belongs to the same original semantic attention. The purpose of this step is to classify text semantics with similar event features together to facilitate the subsequent generation of adverse event label sets. Through event feature clustering, the deep learning system can obtain the adverse event label set corresponding to each management report text semantic, as well as the key event status code of each adverse event label set. These adverse event label sets and key event status codes are important bases for subsequent risk trend prediction and text retelling annotation.

[0084] Next, based on the adverse event label sets contained in the same original semantic attention, and the number of positive debugging examples and negative debugging examples of the hospital management report examples to which the management report text semantics contained in each adverse event label set belongs, the deep learning system needs to determine the risk trend prediction weights of each original semantic attention. This step is a key link in the algorithm debugging process, which determines the importance of different original semantic attentions in risk trend prediction. The risk trend prediction weight is a relative concept, which reflects the possibility and impact of adverse events under different original semantic attentions. By calculating the ratio of the number of positive debugging examples to the number of negative debugging examples, the deep learning system can obtain the risk trend prediction weights of each original semantic attention, thereby providing a basis for the subsequent sampling of target semantic attention.

[0085] After determining the risk trend prediction weights of each original semantic attention, the deep learning system needs to sample the target semantic attention whose risk trend prediction weights meet the trend prediction requirements from the original semantic attention. The purpose of this step is to screen out those original semantic attentions that have a significant impact on risk trend prediction so that they can be given more attention in the subsequent text retelling and annotation process. By sampling the target semantic attention, the deep learning system can further narrow the scope of analysis and improve the processing efficiency of the algorithm.

[0086] Then, the key event status encoding of each hospital management report example in the adverse event label set corresponding to each target semantic attention is loaded into the original text retelling annotation algorithm. This step is the core link of the text retelling annotation algorithm. It combines the key event status encoding with the original text retelling annotation algorithm to generate the discrimination probability of each hospital management report example belonging to a positive debugging example or a negative debugging example. This discrimination probability reflects the possibility and impact of adverse events in the hospital management report example, and is an important basis for subsequent algorithm weight optimization.

[0087] Finally, the debugging examples of each hospital management report example are used to distinguish annotations and discrimination possibilities, and the algorithm weights of the original text paraphrase annotation algorithm are optimized to obtain the final text paraphrase annotation algorithm. This step is the final link of the algorithm debugging process. It optimizes the weights of the original text paraphrase annotation algorithm so that the algorithm can more accurately generate local paraphrase annotation suggestion features when processing the pending hospital management reports. These local paraphrase annotation suggestion features not only contain the specific details of the adverse event, but are also presented in a structured form, which makes it easier for hospital managers to quickly grasp the core information of the event and formulate effective response measures.

[0088] It can be seen that the debugging steps of the text retelling annotation algorithm have achieved intelligent retelling and risk prediction of adverse events through in-depth analysis and processing of hospital management reports. The application of this algorithm will provide hospital managers with a comprehensive and objective risk assessment tool to help them consider adverse event risk factors more scientifically and comprehensively, thereby effectively reducing the incidence of adverse events at the health system level. At the same time, the implementation of this algorithm will also provide new technical support for the development of digital medicine and promote forward-looking changes in hospital management and policy design.

[0089] In the specific implementation process, in order to ensure the accuracy and effectiveness of the text retelling annotation algorithm, the deep learning system also needs to pay attention to the following aspects.

[0090] (1) Acquisition of management report text semantics is the basis for algorithm debugging. The deep learning system needs to ensure that the acquired management report text semantics can accurately reflect the core information about adverse events described in the hospital management report. To this end, the deep learning system can use natural language processing technology to perform text analysis and semantic mining on the management report, extract keywords, phrases and sentences related to adverse events, and form the management report text semantics.

[0091] (2) The accuracy of event feature clustering has an important impact on the generation of subsequent adverse event label sets and the determination of key event status codes. The deep learning system needs to select appropriate clustering algorithms and parameters to ensure that text semantics with similar event features can be accurately classified together. At the same time, the deep learning system also needs to manually review and adjust the clustering results to ensure the accuracy and completeness of the adverse event label set.

[0092] (3) The calculation of risk trend prediction weights is a key step in the algorithm debugging process. The deep learning system needs to ensure that the calculated risk trend prediction weights can accurately reflect the possibility and impact of adverse events under different original semantic attention. To this end, the deep learning system can use statistical methods and machine learning algorithms to calculate and analyze the ratio of the number of positive debugging examples to the number of negative debugging examples to obtain the risk trend prediction weights for each original semantic attention.

[0093] (4) Algorithm weight optimization is the final step of the algorithm debugging process. The deep learning system needs to use the debugging examples of each hospital management report example to distinguish annotations and discriminate possibilities to optimize the weights of the original text paraphrase annotation algorithm to improve the accuracy and efficiency of the algorithm when processing the pending hospital management reports. To this end, the deep learning system can use the optimization algorithm in the machine learning algorithm to iteratively optimize the weights of the original text paraphrase annotation algorithm until the preset convergence conditions are met.

[0094] In other preferred embodiments, event feature clustering is performed on each of the management report text semantics belonging to the same original semantic attention to obtain an adverse event label set corresponding to each of the management report text semantics, and a key event status code of each of the adverse event label sets, including: for each of the original semantic attentions, a semantic feature relationship network is determined through each of the management report text semantics corresponding to the original semantic attention; event feature label induction is performed on each of the management report text semantics according to the relationship path in the semantic feature relationship network to obtain multiple adverse event label sets under the original semantic attention, and the adverse event label set corresponding to each of the management report text semantics is determined; the key event status code of each of the adverse event label sets is determined by using the number of positive debugging examples and the number of negative debugging examples of the hospital management report examples to which the management report text semantics contained in each of the adverse event label sets belong.

[0095] It can be understood that the present invention adopts a more sophisticated and efficient method for the processing and analysis of management report texts, especially the process of labeling and status coding of adverse events therein. The core of this method is to first cluster the event features of the semantics of each management report text belonging to the same original semantic attention, and then obtain the adverse event label set corresponding to each management report text semantic, and the key event status coding of each of these adverse event label sets.

[0096] First, in order to effectively cluster event features of management report text, it is necessary to construct a semantic feature relationship network. This relationship network is based on raw semantic attention, that is, it is constructed based on the content in the management report text that can attract specific semantic attention. In this relationship network, nodes represent semantic units in the management report text, and edges represent the relationship between these semantic units. This relationship can be based on semantic similarity or semantic relevance, depending on the text content being analyzed and the analysis objectives.

[0097] The process of building a semantic feature relationship network is a complex but orderly task. It requires in-depth semantic analysis of the management report text, extracting the key semantic units, and determining the relationship between these semantic units. The implementation of this step may rely on natural language processing technologies, such as word segmentation, part-of-speech tagging, syntactic analysis, etc., as well as semantic analysis technologies, such as semantic role labeling, semantic dependency analysis, etc. Through these technologies, the management report text can be converted into a network structure composed of semantic units and relationships, providing a basis for subsequent event feature clustering.

[0098] Secondly, after the semantic feature relationship network is constructed, the semantics of each management report text can be summarized into event feature labels based on the relationship paths in this relationship network. This process is actually a clustering process on the graph structure, which requires the semantics of management report texts that are semantically similar or related to be clustered together to form adverse event label sets. These label sets represent different types of adverse events under the original semantic attention, and they are an abstraction and generalization of adverse events in management report texts.

[0099] When summarizing event feature labels, it is necessary to consider the various relationship paths in the semantic feature relationship network. These paths may represent different semantic association patterns, such as causal relationships, parallel relationships, transitional relationships, etc. By analyzing these relationship paths, different characteristics and patterns of adverse events in management report texts can be discovered, and then summarized into corresponding adverse event labels. The implementation of this step may rely on graph clustering algorithms, community detection algorithms, etc., which can find sets of nodes with similar characteristics in graph structures.

[0100] Then, after obtaining multiple adverse event label sets under each original semantic attention, the adverse event label set corresponding to each management report text semantic can be determined. This process is a process of matching the management report text semantics with the adverse event label, which requires assigning each management report text semantic to the adverse event label set that best matches its characteristics. The implementation of this step may rely on text classification algorithms, semantic matching algorithms, etc., which can classify management report text semantics into corresponding adverse event label sets based on their characteristics.

[0101] Finally, in order to more comprehensively describe the characteristics of the adverse event label set, it is also necessary to use the number of positive debugging examples and negative debugging examples of the hospital management report examples to which the management report text semantics contained in each adverse event label set belongs to determine the key event status code of each adverse event label set. This process is actually a process of quantitative description of the adverse event label set, which can reflect the severity, frequency of occurrence and other characteristics of the adverse event label set by counting the number of different debugging examples. The implementation of this step may rely on data statistics and analysis techniques, such as descriptive statistical analysis, setting tests, etc.

[0102] It can be seen that this embodiment provides an effective method for clustering event features of management report text and obtaining adverse event label sets and key event status codes. This method realizes a comprehensive analysis and description of adverse events in management report texts by building a semantic feature relationship network, summarizing event feature labels, determining adverse event label sets, and calculating key event status codes. This method not only improves the accuracy and efficiency of text analysis, but also provides strong data support and decision-making basis for hospital management.

[0103] In the next step, the key event status codes of each adverse event label set are determined by using the number of positive debugging examples and the number of negative debugging examples of the hospital management report examples to which the management report text semantics contained in each adverse event label set belong, including: determining the number of positive debugging examples and the number of negative debugging examples in the hospital management report examples to which the management report text semantics contained in each adverse event label set belong; determining a first calculation result of the number of positive debugging examples in each adverse event label set and the cumulative value of the positive debugging examples in the hospital management report examples, and a second calculation result of the number of negative debugging examples in each adverse event label set and the cumulative value of the negative debugging examples in the hospital management report examples; for each adverse event label set, performing quantitative feature mapping on the first calculation result and the second calculation result to determine the key event status codes of each adverse event label set.

[0104] In the actual application process, in order to more deeply understand and analyze the characteristics of the adverse event label set, the deep learning system needs to use the number of positive debugging examples and negative debugging examples of the hospital management report examples to which the management report text semantics contained in each adverse event label set belong to determine the key event status code of each adverse event label set. This process involves counting, calculating, and quantitative feature mapping of debugging examples in the management report examples, aiming to reveal the internal state and characteristics of the adverse event label set in a quantitative way.

[0105] First, the deep learning system needs to determine the number of positive debugging examples and negative debugging examples in the hospital management report examples that belong to the management report text semantics contained in each adverse event label set. This step is the basis for subsequent analysis. It requires the deep learning system to carefully review each hospital management report example and count the debugging behaviors (including positive debugging and negative debugging) described therein. Positive debugging examples refer to those report examples in which effective measures were taken to respond to or resolve adverse events, while negative debugging examples refer to those report examples in which effective response measures were not taken or the response was ineffective. Through this step, the deep learning system can obtain the specific number of positive debugging examples and negative debugging examples under each adverse event label set.

[0106] Secondly, the deep learning system needs to calculate the first operation result of the number of positive debugging examples in each adverse event label set and the cumulative value of positive debugging examples in the hospital management report example, as well as the second operation result of the number of negative debugging examples in each adverse event label set and the cumulative value of negative debugging examples in the hospital management report example. The purpose of this step is to reveal the relative position and importance of the label set in the whole by comparing the number of debugging examples in the adverse event label set with the cumulative value of debugging examples in the entire hospital management report example. The first operation result and the second operation result can respectively reflect the relative performance of the adverse event label set in terms of positive debugging and negative debugging, and provide a basis for the deep learning system to subsequently determine the key event status coding.

[0107] When calculating the first operation result and the second operation result, the deep learning system can use a variety of mathematical methods, such as ratio calculation, percentage calculation, etc., to obtain more comparable and interpretable results. For example, the deep learning system can calculate the ratio of the number of positive debugging examples in the adverse event label set to the cumulative value of positive debugging examples in the entire hospital management report example, and the ratio of the number of negative debugging examples to the cumulative value of negative debugging examples. In this way, the deep learning system can more intuitively understand the relative position and importance of each adverse event label set in terms of positive debugging and negative debugging.

[0108] Finally, for each adverse event label set, the deep learning system needs to perform quantitative feature mapping between the first operation result and the second operation result to determine the key event status code of each adverse event label set. This step is the core of the entire analysis process, which requires the deep learning system to convert the previously obtained operation results into quantitative features with practical significance, and use these features to determine the key event status code of the adverse event label set.

[0109] Quantitative feature mapping is a complex but crucial process. It requires the deep learning system to select appropriate quantification methods and mapping rules according to the actual situation and analysis objectives. For example, the deep learning system can divide the adverse event label set into different state categories, such as "active debugging dominant", "passive debugging dominant", "balanced state", etc., based on the size relationship between the first operation result and the second operation result. The deep learning system can then assign different codes to these state categories to represent their key event states.

[0110] When determining the status codes of key events, the deep learning system also needs to consider the readability and ease of use of the codes. The codes should be able to intuitively reflect the status characteristics of the adverse event label set, facilitating subsequent data analysis and decision support. For example, the deep learning system can use a combination of letters or numbers to represent different status codes, and attach corresponding explanations and instructions to ensure the ease of use and understandability of the codes.

[0111] It can be seen that the process of determining the key event status coding of each adverse event label set by using the number of positive debugging examples and the number of negative debugging examples of the hospital management report examples to which the management report text semantics contained in each adverse event label set belongs is a complex and meticulous task. It involves multiple links such as counting, calculation, and quantitative feature mapping of debugging examples in the management report examples, requiring the deep learning system to comprehensively use mathematical methods, quantitative techniques, and analytical skills to reveal the intrinsic state and characteristics of the adverse event label set. Through this process, the deep learning system can obtain more explanatory and practical key event status coding, providing strong data support and decision-making basis for hospital management.

[0112] In a replaceable embodiment, the risk trend prediction weight of each original semantic attention is determined based on the adverse event label set contained in the same original semantic attention, and the number of positive debugging examples and the number of negative debugging examples of the hospital management report example to which the management report text semantics contained in each adverse event label set belongs, including: determining a first operation result of the number of positive debugging examples in each adverse event label set and the cumulative value of the positive debugging examples in the hospital management report example, and a second operation result of the number of negative debugging examples in each adverse event label set and the cumulative value of the negative debugging examples in the hospital management report example; for each adverse event label set, comparing and analyzing the first operation result with the second operation result to obtain the comparison feature coefficient of each adverse event label set; determining the local trend deduction coefficient of each adverse event label set based on the key event status code of each adverse event label set contained in the same original semantic attention and the comparison feature coefficient of each adverse event label set; for each original semantic attention, integrating the local trend deduction coefficient of each contained adverse event label set to obtain the risk trend prediction weight of each original semantic attention.

[0113] In detail, the deep learning system adopts a comprehensive analysis and calculation method to determine the risk trend prediction weight of each original semantic attention based on the adverse event label set contained in the same original semantic attention, and the number of positive debugging examples and negative debugging examples of the hospital management report examples to which the management report text semantics contained in these adverse event label sets belong. This method not only considers the calculation results of the number of debugging examples in the adverse event label set and the cumulative value of debugging examples in the hospital management report example, but also further compares and analyzes these calculation results to obtain the comparison feature coefficient of the adverse event label set. Finally, combined with the key event state encoding and comparison feature coefficient of the adverse event label set, the deep learning system determines the local trend deduction coefficient of each adverse event label set, and obtains the risk trend prediction weight of each original semantic attention by integrating these coefficients.

[0114] First, the deep learning system needs to determine the first operation result of the number of positive debugging examples in each adverse event label set and the cumulative value of positive debugging examples in the hospital management report example, and the second operation result of the number of negative debugging examples in each adverse event label set and the cumulative value of negative debugging examples in the hospital management report example. This step is the basis for subsequent analysis, which requires the deep learning system to carefully count and calculate each adverse event label set to obtain its specific performance in terms of positive debugging and negative debugging. Through this step, the deep learning system can obtain two important operation results, which will serve as the basis for subsequent comparative analysis and determination of comparison feature coefficients.

[0115] Secondly, for each adverse event label set, the deep learning system needs to compare and analyze the first operation result with the second operation result to obtain the comparison feature coefficient of each adverse event label set. The purpose of this step is to reveal its inherent characteristics and trends by comparing the performance of the adverse event label set in active debugging and passive debugging. The comparison feature coefficient is an important quantitative indicator that can reflect the relative advantages and disadvantages of the adverse event label set in debugging behavior, and provide a basis for the deep learning system to subsequently determine the local trend deduction coefficient.

[0116] After obtaining the comparison feature coefficient, the deep learning system needs to further determine the local trend deduction coefficient of each adverse event label set based on the key event status coding of each adverse event label set contained in the same original semantic attention and the comparison feature coefficient of each adverse event label set. This step is the core of the entire analysis process. It requires the deep learning system to combine the comparison feature coefficient with the key event status coding, and obtain the deduction coefficient that can reflect the local trend of the adverse event label set through comprehensive analysis and calculation. The local trend deduction coefficient is a predictive quantitative indicator that can reveal the possibility and risk of the future development trend of the adverse event label set.

[0117] Finally, for each original semantic attention, the deep learning system needs to integrate the local trend deduction coefficients of the included adverse event label sets to obtain the risk trend prediction weight of each original semantic attention. This step is the ultimate goal of the entire analysis process. It requires the deep learning system to summarize and integrate the local trend deduction coefficients of each adverse event label set to obtain a prediction weight that can reflect the overall risk trend of the original semantic attention. The risk trend prediction weight is an innovative quantitative indicator that can provide important decision support and risk warning for hospital management.

[0118] Based on this embodiment, the deep learning system needs to make full use of mathematical methods, quantitative techniques and analytical skills to ensure the accuracy and reliability of the analysis results. At the same time, the deep learning system also needs to pay attention to maintaining the coherence and consistency of the analysis process to ensure that the final risk trend prediction weight can truly reflect the inherent characteristics and trends of the original semantic attention. Through this technical solution, the deep learning system can provide more accurate and effective risk prediction and decision support for hospital management.

[0119] Under some optional design ideas, the debugging examples of each of the hospital management report examples are used to distinguish between annotations and the discrimination possibilities, and the algorithm weights of the original text paraphrase annotation algorithm are optimized to obtain a text paraphrase annotation algorithm, including: using the debugging examples of each of the hospital management report examples to distinguish between annotations and the discrimination possibilities, generating semantic mining error indicators, and determining the indicator feature distribution of the semantic mining error indicators in a preset feature space as semantic paraphrase error features; using the debugging examples of each of the hospital management report examples to distinguish between annotations and the discrimination possibilities, as well as the semantic paraphrase error features, and optimizing the algorithm weights of the original text paraphrase annotation algorithm to obtain a text paraphrase annotation algorithm.

[0120] The deep learning system can adopt an innovative method to distinguish annotations and discriminate possibilities by using debugging examples of various hospital management report examples, and optimize the algorithm weights of the original text paraphrase annotation algorithm to obtain a more accurate and efficient text paraphrase annotation algorithm. This method not only involves in-depth analysis of hospital management report examples, but also combines the generation and application of semantic mining error indicators and semantic paraphrase error features, aiming to improve the accuracy and effect of text paraphrase annotation by optimizing the algorithm weights.

[0121] First, the deep learning system needs to use the debugging examples of each hospital management report example to distinguish annotations and discriminant possibilities to generate semantic mining error indicators. This step is the basis of the entire optimization process. It requires the deep learning system to carefully parse the hospital management report examples and extract the distinguishing annotations and discriminant possibilities of the debugging examples. Distinguishing annotations refers to clearly marking the debugging examples in the report for subsequent analysis and processing; while discriminant possibilities refer to further analysis and judgment of these debugging examples to determine their semantic accuracy and reliability. Through this step, the deep learning system can obtain a set of indicators that reflect the semantic mining errors of hospital management report examples, which will serve as an important basis for the weights of the subsequent optimization algorithm.

[0122] Secondly, the deep learning system needs to determine the indicator feature distribution of the generated semantic mining error indicator in the preset feature space as the semantic paraphrase error feature. This step is a bridge connecting the semantic mining error indicator and the algorithm weight optimization. It requires the deep learning system to convert the semantic mining error indicator into a more abstract and general feature representation, namely the semantic paraphrase error feature. This feature representation can better reflect the errors and uncertainties in the semantic paraphrase of the hospital management report example, and provides a strong basis for the deep learning system to optimize the algorithm weights in the future. Through this step, the deep learning system can convert the specific semantic mining error indicator into a more general and computable feature form, laying the foundation for the subsequent optimization process.

[0123] Then, the deep learning system needs to use the debugging examples of each hospital management report example to distinguish annotations and discrimination possibilities, as well as the generated semantic paraphrase error features, to optimize the algorithm weights of the original text paraphrase annotation algorithm. This step is the core of the entire optimization process. It requires the deep learning system to combine the semantic paraphrase error features obtained earlier with the original text paraphrase annotation algorithm, and to improve the accuracy and effect of the algorithm by adjusting and optimizing the algorithm weights. During the optimization process, the deep learning system can use a variety of machine learning and data mining techniques, such as regression analysis, classification algorithms, clustering analysis, etc., to fine-tune and optimize the algorithm weights. Through this step, the deep learning system can obtain a more accurate and effective set of algorithm weights, thereby improving the accuracy and performance of the text paraphrase annotation algorithm.

[0124] Finally, through the above optimization process, the deep learning system can obtain a more accurate and efficient text paraphrase annotation algorithm. This algorithm not only considers the debugging example of the hospital management report example to distinguish annotations and discrimination possibilities, but also makes full use of the information of semantic mining error indicators and semantic paraphrase error characteristics. By optimizing the algorithm weights, it realizes the precise control and effective improvement of text paraphrase annotation. This algorithm can be widely used in medical text analysis, natural language processing, information retrieval and other fields, providing more accurate and effective text paraphrase annotation support for related applications.

[0125] In this way, by using the debugging examples of each hospital management report example to distinguish annotations and discrimination possibilities, generating semantic mining error indicators, and converting them into semantic paraphrase error features, and then optimizing the algorithm weights of the original text paraphrase annotation algorithm, the deep learning system can obtain a more accurate and efficient text paraphrase annotation algorithm. This algorithm not only improves the accuracy and effect of text paraphrase annotation, but also provides more powerful and flexible text processing capabilities for related applications.

[0126] In the following design idea, the debugging examples of each of the hospital management report examples are used to distinguish the annotations and the discrimination possibility, as well as the semantic paraphrase error characteristics, to optimize the algorithm weights of the original text paraphrase annotation algorithm to obtain a text paraphrase annotation algorithm, including: using the debugging examples of each of the hospital management report examples to distinguish the annotations and the discrimination possibility to determine the discrimination debugging offset; determining the word-level semantic confidence factor through the discrimination debugging offset gradient and the semantic paraphrase error feature gradient; determining the word-level semantic debugging error based on the discrimination debugging offset and the word-level semantic confidence factor, and optimizing the algorithm weights of the original text paraphrase annotation algorithm through the word-level semantic debugging error to obtain a text paraphrase annotation algorithm.

[0127] Among them, the deep learning system can adopt an innovative method for how to use the debugging examples of each hospital management report example to distinguish between annotation and discrimination possibilities, as well as semantic paraphrase error characteristics, to optimize the algorithm weights of the original text paraphrase annotation algorithm to obtain a more accurate and efficient text paraphrase annotation algorithm. This method not only involves in-depth analysis of hospital management report examples, but also combines the generation and application of key concepts such as discriminative debugging offset and word-level semantic confidence factors, aiming to improve the accuracy and effect of text paraphrase annotation by optimizing the algorithm weights.

[0128] First, the deep learning system needs to use the debugging examples of each hospital management report example to distinguish the annotations and discriminant possibilities to determine the discriminant debugging offset. This step is the basis of the entire optimization process. It requires the deep learning system to carefully parse the hospital management report examples and extract the distinguishing annotations and discriminant possibilities of the debugging examples. Discriminating annotations refer to clearly marking the debugging examples in the report for subsequent analysis and processing; while discriminant possibilities refer to further analyzing and judging these debugging examples to determine their semantic accuracy and reliability. Through this step, the deep learning system can obtain a set of indicators that reflect the discriminant debugging offset of the hospital management report examples, which will serve as an important basis for the subsequent optimization algorithm weights. The discriminant debugging offset is an important quantitative indicator that reveals the degree of offset and uncertainty of the debugging examples in the discrimination process, and provides a strong basis for the deep learning system to subsequently determine the semantic confidence factor at the word level.

[0129] Secondly, the deep learning system needs to determine the word-level semantic confidence factor through the discriminative debugging offset gradient and the semantic paraphrase error feature gradient. This step is a bridge connecting the discriminative debugging offset and the algorithm weight optimization. It requires the deep learning system to convert the discriminative debugging offset into a more abstract and general feature representation, namely the word-level semantic confidence factor. This feature representation can better reflect the confidence and uncertainty of the hospital management report example in the word-level semantics, and provides a strong basis for the deep learning system to optimize the algorithm weights in the future. Through this step, the deep learning system can convert the specific discriminative debugging offset into a more general and computable feature form, laying the foundation for the subsequent optimization process. In the process of determining the word-level semantic confidence factor, the deep learning system needs to comprehensively consider the influence of the discriminative debugging offset gradient and the semantic paraphrase error feature gradient. The discriminative debugging offset gradient reflects the degree and trend of the discriminative debugging offset between different words and sentences, while the semantic paraphrase error feature gradient reflects the distribution and characteristics of the semantic paraphrase error between different words and sentences. By comprehensively considering the influence of these two gradients, the deep learning system can obtain a more accurate and reliable word-level semantic confidence factor.

[0130] Then, the deep learning system needs to determine the word-level semantic debugging error based on the discriminant debugging offset and the word-level semantic confidence factor, and optimize the algorithm weights of the original text paraphrase annotation algorithm through this error. This step is the core of the entire optimization process. It requires the deep learning system to combine the word-level semantic confidence factor obtained above with the original text paraphrase annotation algorithm, and improve the accuracy and effect of the algorithm by adjusting and optimizing the algorithm weights. During the optimization process, the deep learning system can use a variety of machine learning and data mining techniques, such as regression analysis, classification algorithms, clustering analysis, etc., to fine-tune and optimize the algorithm weights. Through this step, the deep learning system can obtain a set of more accurate and effective algorithm weights, thereby improving the accuracy and performance of the text paraphrase annotation algorithm. The word-level semantic debugging error is an important quantitative indicator, which reveals the debugging error and uncertainty of the text paraphrase annotation algorithm in the word-level semantics. By minimizing this error, the deep learning system can obtain more accurate and effective algorithm weights, thereby improving the accuracy and effect of the text paraphrase annotation algorithm.

[0131] Finally, through the above optimization process, the deep learning system can obtain a more accurate and efficient text paraphrase annotation algorithm. This algorithm not only considers the debugging example of the hospital management report example to distinguish annotations and discrimination possibilities, but also makes full use of the information of semantic paraphrase error characteristics, discrimination debugging offset and word-level semantic confidence factor. By optimizing the algorithm weights, it realizes precise control and effective improvement of text paraphrase annotation. This algorithm can be widely used in medical text analysis, natural language processing, information retrieval and other fields, providing more accurate and effective text paraphrase annotation support for related applications.

[0132] In this way, by using the debugging examples of each hospital management report example to distinguish annotations and discrimination possibilities, as well as semantic paraphrase error characteristics, the deep learning system can determine the discrimination debugging offset and word-level semantic confidence factor, and optimize the algorithm weights of the original text paraphrase annotation algorithm based on this information, resulting in a more accurate and efficient text paraphrase annotation algorithm. This algorithm not only improves the accuracy and effect of text paraphrase annotation, but also provides more powerful and flexible text processing capabilities for related applications.

[0133] In some examples, the multiple hospital management report examples belong to a target digital hospital system; the method also includes: determining the initial screening discrimination possibility that does not meet the management model of the target digital hospital system from the discrimination possibilities obtained through the key event status coding; determining the target semantic attention matched by the key event status coding of the initial screening discrimination possibility as the semantic attention to be corrected; based on the semantic keywords of the semantic attention to be corrected, screening the correction strategy matching the semantic keywords to perform correction operations on the management report text semantics belonging to the semantic attention to be corrected, and obtaining the management report text semantics after the correction operation; performing event feature clustering again through the management report text semantics after the correction operation, and determining the adverse event label set corresponding to the management report text semantics after the correction operation.

[0134] When faced with multiple examples of hospital management reports, especially when these examples all come from the same target digital hospital system, the challenge faced by the deep learning system is not only to understand and process the information in these reports, but also to ensure that this information matches the management model of the hospital system, and how to identify and correct information that may not conform to this model. To this end, the method also covers a series of additional steps, aiming to extract the discrimination possibilities from the key event status codes, identify the initial screening discrimination possibilities that do not meet the management model of the target digital hospital system, and then correct these possibilities, and finally determine the adverse event label set corresponding to the corrected management report text semantics through event feature clustering.

[0135] First, the method involves determining the initial screening discrimination possibilities that do not meet the management model of the target digital hospital system from the discrimination possibilities obtained through the key event status coding. This step is the basis of the entire correction process. It requires the deep learning system to deeply analyze the key event status coding and extract those discrimination possibilities that may not meet the hospital system management model. The key event status coding is an abstract and quantitative representation of the key events that occurred in the hospital management report. It reflects the state, attributes and association relationship of the event with other events. By analyzing these codes, the deep learning system can identify those discrimination possibilities that do not meet the expected management model and provide targets for subsequent correction operations.

[0136] Secondly, the target semantic attention matched with the key event state encoding of the initial screening possibility is determined as the semantic attention to be corrected. This step is a bridge between the possibility of discrimination and the specific text semantics. It requires the deep learning system to convert the abstract possibility of discrimination into a specific text semantic representation, namely the semantic attention to be corrected. The target semantic attention refers to those parts of the text semantics that are directly related to the possibility of initial screening discrimination. They are the key objects that need to be corrected. By determining the semantic attention to be corrected, the deep learning system can concretize the target of the correction operation and provide clear guidance for subsequent processing.

[0137] Then, according to the semantic keywords of the semantic attention to be corrected, the correction strategy matching the semantic keywords is determined to perform correction operations on the semantics of the management report text belonging to the semantic attention to be corrected, and the semantics of the management report text after the correction operation is obtained. This step is the core of the entire correction process, which requires the deep learning system to formulate and implement corresponding correction strategies according to the specific semantic content of the semantic attention to be corrected. The correction strategy refers to a series of specific text processing operations aimed at modifying or adjusting the text semantics to make it conform to the management model of the target digital hospital system. Through the correction operation, the deep learning system can eliminate inconsistencies or errors in the text semantics and improve the quality and accuracy of the text.

[0138] Finally, event feature clustering is performed again through the semantics of the management report text after the correction operation to determine the adverse event label set corresponding to the semantics of the management report text after the correction operation. This step is the verification and application of the correction results. It requires the deep learning system to use event feature clustering technology to further analyze and process the corrected text semantics. Event feature clustering is a technology that classifies events with similar features into one category. It can help the deep learning system identify different types of adverse events described in the text. Through event feature clustering, the deep learning system can obtain the adverse event label set corresponding to the semantics of the management report text after the correction operation. These label sets reflect the types and characteristics of the adverse events described in the text, providing an important reference for subsequent hospital management and decision-making.

[0139] In this way, by extracting the discrimination possibilities from the key event status codes, we can identify the initial screening discrimination possibilities that do not meet the target digital hospital system management model, and then correct these possibilities, and finally determine the adverse event label set corresponding to the corrected management report text semantics through event feature clustering. This process not only improves the quality and accuracy of the text, but also provides strong support for hospital management and decision-making.

[0140] In some examples, the semantic keyword of the semantic attention to be corrected is cross-departmental conflict risk, and the correction strategy matching the cross-departmental conflict risk is a correction strategy based on spatiotemporal resource scheduling, wherein the semantics of the management report text after the correction operation includes the semantics of the management report text after the spatiotemporal feature correction operation.

[0141] Understandably, faced with complex information in hospital management reports, especially those involving cross-departmental conflict risks, how to ensure the accuracy and consistency of this information so that it conforms to the management model of the target digital hospital system has become an important challenge. To this end, the deep learning system adopts an innovative technical solution, which aims to extract discrimination possibilities from key event status codes, identify initial screening discrimination possibilities that do not meet the hospital system management model, and then correct these possibilities, and finally determine the set of adverse event labels corresponding to the semantics of the corrected management report text through event feature clustering. In this process, the deep learning system pays special attention to the semantic keyword of cross-departmental conflict risk, and adopts a correction strategy based on spatiotemporal resource scheduling to perform correction operations.

[0142] First, the deep learning system determines the initial screening discrimination possibilities that do not meet the management model of the target digital hospital system from the discrimination possibilities obtained through the key event status coding. This step is the basis of the entire correction process. It requires the deep learning system to deeply analyze the key event status coding and extract those discrimination possibilities that may not meet the hospital system management model. The key event status coding is an abstract and quantitative representation of the key events that occurred in the hospital management report. It reflects the state, attributes and relationship of the event with other events. By analyzing these codes, the deep learning system can identify those discrimination possibilities that do not meet the expected management model and provide targets for subsequent correction operations.

[0143] After determining the possibility of initial screening, the deep learning system further matches these possibilities with the key event status codes, thereby determining the semantic attention to be corrected, namely the risk of cross-departmental conflict. The risk of cross-departmental conflict refers to the conflicts in responsibilities, interests or resource allocation between different departments during hospital management, which lead to problems such as low management efficiency, poor information communication or obstructed decision-making execution. The determination of this semantic keyword provides a clear direction and goal for the subsequent correction operation of the deep learning system.

[0144] Next, the deep learning system determines the correction strategy that matches the semantic keyword of the semantic attention to be corrected, that is, the risk of cross-departmental conflict, which is the correction strategy based on spatiotemporal resource scheduling. The correction strategy based on spatiotemporal resource scheduling is an effective solution to the risk of cross-departmental conflict. It reduces or eliminates conflicts and contradictions between different departments by rationally scheduling and optimizing the spatiotemporal resources in the hospital system. The implementation of this strategy requires the deep learning system to conduct in-depth analysis and research on the spatiotemporal resources in the hospital system, including the allocation and use of time resources (such as working hours, meeting time, etc.) and space resources (such as wards, operating rooms, etc.). Through reasonable spatiotemporal resource scheduling, the deep learning system can effectively alleviate or resolve the risk of cross-departmental conflict and improve the overall efficiency and effectiveness of hospital management.

[0145] After determining the correction strategy, the deep learning system corrected the semantics of the management report text that belonged to the semantic attention to be corrected, and obtained the semantics of the management report text after the correction operation. This step is the core of the entire correction process. It requires the deep learning system to make specific modifications and adjustments to the management report text according to the specific semantic content of the semantic attention to be corrected and the selected correction strategy. During the correction process, the deep learning system paid special attention to the important text semantic element of spatiotemporal features and performed corresponding correction operations on it. Spatiotemporal features refer to the time- and space-related event or behavior characteristics described in the management report, which are of great significance for understanding and analyzing cross-departmental conflict risks. Through the correction operation of spatiotemporal features, the deep learning system can more accurately describe and reflect the cross-departmental conflict risk situation in hospital management.

[0146] Finally, the deep learning system clustered the event features again through the semantics of the management report text after the correction operation, and determined the adverse event label set corresponding to the semantics of the management report text after the correction operation. This step is the verification and application of the correction results. It requires the deep learning system to use event feature clustering technology to further analyze and process the corrected text semantics. Event feature clustering is a technology that classifies events with similar features into one category. It can help the deep learning system identify different types of adverse events described in the text. Through event feature clustering, the deep learning system can obtain the adverse event label set corresponding to the semantics of the management report text after the correction operation. These label sets reflect the types and characteristics of the adverse events described in the text. For the semantic keyword of cross-departmental conflict risk, the corresponding adverse event label set may include "poor communication between departments" and "unequal resource allocation". The determination of these label sets provides an important reference for the subsequent hospital management and decision-making of the deep learning system.

[0147] In this way, the deep learning system achieves the correction and processing of cross-departmental conflict risks in hospital management reports through a series of steps, including extracting discrimination possibilities from key event status codes, determining preliminary screening discrimination possibilities, matching semantic attention to be corrected, selecting correction strategies, performing correction operations, and event feature clustering. This process not only improves the quality and accuracy of the text, but also provides strong support for hospital management and decision-making.

[0148] In other examples, the semantic keyword of the semantic attention to be corrected is the risk of leakage of hazardous chemicals; the correction strategy selected to match the semantic keyword performs a correction operation on the management report text semantics belonging to the semantic attention to be corrected, and obtains the management report text semantics after the correction operation, including: determining a correction strategy based on spatiotemporal resource scheduling to perform a correction operation on the management report text semantics belonging to the semantic attention to be corrected, and obtains the management report text semantics after the spatiotemporal feature correction operation; selecting a security level correction strategy to perform a correction operation on the management report text semantics belonging to the semantic attention to be corrected, and obtains the management report text semantics after the security level correction operation; wherein, the management report text semantics after the correction operation include the management report text semantics after the spatiotemporal feature correction operation, and the management report text semantics after the security level correction operation.

[0149] In addition, facing the important issue of hazardous chemical leakage risk involved in hospital management reports, how to ensure the accuracy and consistency of the report so that it conforms to the management model of the target digital hospital system has become an urgent problem to be solved. To this end, the deep learning system adopts an innovative technical solution, which aims to screen the correction strategy that matches the semantic keyword of hazardous chemical leakage risk, and perform correction operations on the management report text semantics belonging to the semantic attention to be corrected, so as to obtain the management report text semantics after the correction operation. In this process, the deep learning system pays special attention to the application of two strategies: correction strategy based on spatiotemporal resource scheduling and safety level correction strategy.

[0150] First, the deep learning system determines the correction strategy based on spatiotemporal resource scheduling to perform correction operations on the semantics of the management report text that belongs to the semantic attention to be corrected. The application of this strategy stems from the deep learning system's in-depth analysis of hazardous chemical leakage risk events in hospital management reports. The deep learning system found that such events are often closely related to specific time and space conditions, such as the storage location, use time, and transportation routes of hazardous chemicals. Therefore, through the reasonable scheduling and optimal configuration of spatiotemporal resources, the deep learning system can effectively reduce the probability of hazardous chemical leakage risks. In the correction operation, the deep learning system carefully combed and adjusted the spatiotemporal features in the management report to ensure that key information such as time and location in the report is consistent with the actual situation, thereby improving the accuracy and credibility of the report.

[0151] Secondly, the deep learning system determined the safety level correction strategy to perform correction operations on the semantics of the management report text that belongs to the semantic attention to be corrected. The application of this strategy is based on the deep learning system's deep understanding of the hospital's safety level management system. The deep learning system knows that different hazardous chemicals have different degrees of danger and leakage risks. Therefore, hospitals need to adopt different management measures and reporting requirements for hazardous chemicals of different levels. During the correction operation, the deep learning system strictly verified and corrected the safety level information in the management report to ensure that the level of hazardous chemicals in the report is consistent with the hospital's safety level management system. The implementation of this step not only improves the standardization of management reports, but also provides strong support for the hospital's safety management.

[0152] After determining the above two correction strategies, the deep learning system performed a comprehensive correction operation on the semantics of the management report text that belongs to the semantic attention to be corrected. In this process, the deep learning system made full use of text processing technology and natural language processing technology to extract, analyze and correct the key information in the report. Through the organic combination of spatiotemporal feature correction operation and safety level correction operation, the deep learning system obtained the semantics of the management report text after the correction operation. This semantics not only includes the semantics of the management report text after the spatiotemporal feature correction operation, but also includes the semantics of the management report text after the safety level correction operation, thereby achieving a comprehensive, accurate and consistent description of the important issue of hazardous chemical leakage risk.

[0153] It is worth noting that in the text semantics of the management report after the correction operation, the text semantics of the management report after the spatiotemporal feature correction operation and the text semantics of the management report after the safety level correction operation are interrelated and complementary. The spatiotemporal feature correction operation ensures that the time and space information in the report is consistent with the actual situation, providing accurate spatial and temporal positioning for the analysis and response of hazardous chemical leakage risks; while the safety level correction operation ensures that the hazardous chemical level in the report is consistent with the hospital's safety level management system, providing a strong institutional guarantee for the management and control of hazardous chemical leakage risks. The organic combination of these two correction operations not only improves the accuracy and credibility of the management report, but also provides comprehensive and strong support for the hospital's safety management.

[0154] In addition, the deep learning system further analyzed and processed the semantics of the management report text after the correction operation. The deep learning system classified and summarized the hazardous chemical leakage risk events in the report through event feature clustering technology, and obtained the feature label sets corresponding to different types of hazardous chemical leakage risk events. The implementation of this step not only helps the deep learning system to more deeply understand and analyze the laws and characteristics of hazardous chemical leakage risk events, but also provides strong data support for the hospital to formulate targeted risk response strategies.

[0155] Therefore, by screening the correction strategy that matches the semantic keyword of hazardous chemical leakage risk, a comprehensive correction operation was performed on the semantics of the management report text that belongs to the semantic attention to be corrected, and the semantics of the management report text after the correction operation was obtained. In this process, the deep learning system made full use of text processing technology and natural language processing technology to extract, analyze and correct the key information in the report. Through the organic combination of spatiotemporal feature correction operation and safety level correction operation, the deep learning system achieves a comprehensive, accurate and consistent description of the important issue of hazardous chemical leakage risk. The implementation of this technical solution not only improves the accuracy and credibility of management reports, but also provides comprehensive and strong support for the safety management of hospitals.

[0156] In some other optional embodiments, the method also includes: determining the first risk trend prediction weight of the semantic attention to be corrected through the management report text semantics after the spatiotemporal feature correction operation, and determining the second risk trend prediction weight of the semantic attention to be corrected through the management report text semantics after the security level correction operation; determining the management report text semantics corresponding to the target prediction weight from the first risk trend prediction weight and the second risk trend prediction weight as the candidate management report text semantics under the semantic attention to be corrected; performing event feature clustering on each of the candidate management report text semantics of the semantic attention to be corrected, and obtaining the adverse event label set corresponding to each of the candidate management report text semantics, and the key event status code of each of the adverse event label sets.

[0157] Based on this embodiment, the present invention further expands its scope and depth of application. It is not only limited to the correction operation of the management report text semantics, but also covers the prediction of risk trends and cluster analysis of adverse events. Specifically, the method first determines the first risk trend prediction weight of the semantic attention to be corrected through the management report text semantics after the spatiotemporal feature correction operation. This step aims to use the text semantics after spatiotemporal feature correction to evaluate the potential trends related to the risk of leakage of hazardous chemicals. Spatiotemporal features, such as the storage location, transportation route, and use time of hazardous chemicals, are important factors affecting the risk of leakage of hazardous chemicals. By correcting these features, the changing trend of the risk can be grasped more accurately.

[0158] Next, the second risk trend prediction weight of the semantic attention to be corrected is determined by the text semantics of the management report after the safety level correction operation. This step considers the impact of the safety level of different hazardous chemicals on the leakage risk from the perspective of safety level. The text semantics after safety level correction reflects the hospital's management requirements and attention to different hazardous chemicals, which is also an indispensable part of predicting risk trends.

[0159] After obtaining the first risk trend prediction weight and the second risk trend prediction weight, the management report text semantics corresponding to the target prediction weight is further determined from the two weights as the candidate management report text semantics under the semantic attention to be corrected. This step is actually a process of weighing and selecting. By comparing the size and importance of the two weights, the management report text semantics with greater predictive value are selected as candidates for subsequent analysis and processing.

[0160] Subsequently, event feature clustering is performed on each candidate management report text semantics to be corrected. Event feature clustering is a technical method that groups events with similar features into one category, which can help the deep learning system better understand and analyze the various adverse events described in the text. By clustering the event features of the candidate management report text semantics, the adverse event label sets corresponding to each candidate management report text semantics can be obtained. These label sets are a summary and classification of the adverse events described in the text, and they reflect the various problems and risk points in hospital management.

[0161] Finally, the key event status codes of each adverse event tag set were obtained. The key event status code is an abstract and quantitative representation of adverse events, which contains information such as the status, attributes, and associations with other events of the event. By encoding the key event status of the adverse event tag set, we can further explore and analyze the internal connections and laws between events, providing more accurate and powerful support for hospital management.

[0162] In this way, not only the semantic correction operation of the management report text is realized, but also the value and practicality of the hospital management report is further improved through risk trend prediction and adverse event clustering analysis. Specifically, this method comprehensively analyzes and processes the risk of hazardous chemical leakage in the management report through two strategies: spatiotemporal feature correction and safety level correction; at the same time, through the determination of risk trend prediction weights and the selection of candidate management report text semantics, it realizes the accurate prediction and grasp of potential risks; finally, through event feature clustering and key event status coding acquisition, it provides hospital management with a more in-depth and comprehensive adverse event analysis and response strategy. The implementation of this series of steps not only improves the accuracy and credibility of management reports, but also provides powerful data support and analysis tools for hospital safety management and decision-making.

[0163] In practical applications, this method can be used in hospitals for daily monitoring and management of hazardous chemical leakage risks. By regularly correcting and clustering management reports, hospitals can promptly discover and identify potential hazardous chemical leakage risks and take appropriate countermeasures for prevention and control. At the same time, through in-depth analysis of adverse event label sets and key event status codes, hospitals can further explore and understand the laws and causes of adverse events, providing a basis and support for the formulation of more scientific and effective safety management strategies.

[0164] In addition, this method can also be integrated and connected with other hospital management information systems to achieve data sharing and interaction. Through integration with other systems, this method can more conveniently obtain and manage various data and information in the hospital, further improving the accuracy and efficiency of risk prediction and adverse event analysis. At the same time, this method can also provide hospital managers with intuitive and visual analysis reports and tools to help them more deeply understand and grasp the hospital's safety management status and risk trends.

[0165] It can be seen that this embodiment provides a new idea and method for the analysis and processing of hospital management reports by comprehensively using multiple technical means such as text correction, risk prediction, event clustering, etc. This method not only improves the accuracy and credibility of management reports, but also provides powerful data support and analysis tools for hospital safety management and decision-making.

[0166] In some independent embodiments, the method also includes: obtaining multiple past hospital management report examples, and sampling target hospital management report examples that meet the debugging requirements from the multiple past hospital management report examples through the past adverse event processing records of each of the past hospital management report examples; using the report text data of each of the target hospital management report examples, sampling positive debugging examples belonging to the hospital management report examples that meet the requirements of the retelling task from each of the target hospital management report examples.

[0167] It can be understood that based on step 110-step 140, the present invention further expands its application scope and depth, not only covering the aforementioned text correction, risk prediction and adverse event cluster analysis, but also adding deep learning and debugging strategies for past hospital management reports. Specifically, this embodiment aims to obtain and utilize multiple past hospital management report examples, sample target hospital management report examples that meet specific debugging requirements, and thus provide strong support for the improvement and optimization of hospital management reports.

[0168] First, the deep learning system needs to obtain multiple examples of past hospital management reports. These past reports are valuable data generated by the hospital in the historical management process. They record the hospital's management practices, lessons learned, and handling of adverse events in different periods and situations. Through in-depth analysis and learning of these past reports, the deep learning system can discover problems and deficiencies in hospital management, and thus provide useful reference and guidance for current hospital management.

[0169] After obtaining past hospital management report examples, the deep learning system needs to sample target hospital management report examples that meet the debugging requirements from the multiple past hospital management report examples through the past adverse event processing records of each of the past hospital management report examples. This step is key, as it determines the direction and focus of subsequent debugging work. Specifically, the deep learning system needs to screen out those representative reports that can reflect common or typical problems in hospital management from past reports as target examples. These target examples will serve as the focus of subsequent debugging work for the deep learning system. Through in-depth analysis and processing of them, the deep learning system can better grasp the core issues and key links of hospital management.

[0170] After sampling the target hospital management report examples, the deep learning system needs to use the report text data of each of the target hospital management report examples to sample active debugging examples that belong to the hospital management report examples and meet the requirements of the retelling task from each of the target hospital management report examples. This step is the specific implementation stage of the debugging work. The retelling task refers to a detailed retelling and description of the management practices, lessons learned, and adverse event handling process in the target example in order to better understand and grasp the key information and core points. Active debugging examples refer to those examples that have positive significance and can provide useful reference and guidance for current hospital management. Through in-depth analysis and learning of these active debugging examples, the deep learning system can discover excellent practices and successful experiences in hospital management, and thus provide useful reference and inspiration for current hospital management.

[0171] In practical applications, this method can play an important role. For example, when a hospital faces new management challenges or problems, it can discover successful experiences and coping strategies in similar situations through deep learning and debugging of past hospital management reports, thereby providing useful reference and guidance for current problems. At the same time, by learning and promoting positive debugging examples, it can also promote the sharing and exchange of hospital management experience and improve the overall level and quality of hospital management.

[0172] In addition, the method can also be integrated and connected with other hospital management information systems to achieve data sharing and interaction. Through integration with other systems, the method can more conveniently obtain and manage various data and information in the hospital, and further improve the accuracy and efficiency of the analysis and debugging of past hospital management reports. At the same time, the method can also provide hospital managers with intuitive and visual analysis reports and tools to help them more deeply understand and grasp the historical management status and lessons learned of the hospital.

[0173] In this way, by comprehensively using a variety of technical means such as obtaining, sampling, retelling and debugging past hospital management reports, a new idea and method is provided for the mining and learning of hospital management experience. This method can not only help hospitals better understand and grasp their own management practices and lessons learned, but also provide useful reference and guidance for current hospital management. In future research and applications, deep learning systems can further explore and optimize this method to improve its applicability and effectiveness in different scenarios and fields. For example, deep learning systems can study how to more effectively sample target examples that meet debugging requirements from past hospital management reports, how to more accurately repeat and describe the key information and core points in the target examples, and how to formulate and implement more targeted debugging strategies. These studies will help promote the further development and application of this method in the field of hospital management.

[0174] In other independent embodiments, the method also includes: performing feature analysis on the management report text semantics of the hospital management report example under the same initial semantic attention to obtain a management report text semantic heat map of each initial semantic attention; performing cross-analysis on the management report text semantics of the hospital management report example under different initial semantic attentions to obtain a semantic involvement vector between each initial semantic attention and an event impact vector under each initial semantic attention; determining the original semantic attention from each initial semantic attention through the management report text semantic heat map of each initial semantic attention, the semantic involvement vector between each initial semantic attention and the event impact vector under each initial semantic attention.

[0175] In actual application, the present invention is not limited to the acquisition and debugging of past hospital management reports, but also goes deep into the level of fine analysis of the semantics of the management report text. The core of this embodiment is to reveal the key information, potential connections and event impacts in the report through in-depth analysis of the semantics of the hospital management report text, thereby providing more accurate and powerful decision support for hospital management.

[0176] First, the method performs feature parsing on the semantics of hospital management report text under the same initial semantic attention. Initial semantic attention refers to the core or basic semantic areas that are first focused on when performing semantic analysis on management reports. These areas usually contain the main points, important events or key data of the report. By performing feature parsing on these semantic areas, the method can capture the key information in the report text and display it in an intuitive way. This is the management report text semantic heat map. The management report text semantic heat map is a visualization tool that uses different colors or intensities to represent the attention and importance of different semantic areas in the report text. Through the heat map, the deep learning system can clearly see which parts are the focus of the report and which parts contain key information, thereby helping the deep learning system to understand and grasp the core content of the report more quickly.

[0177] Secondly, the method further cross-analyzes the text semantics of the hospital management report example under different initial semantic attentions. Cross-analysis is a statistical method used to explore the relationship between different variables. Here, the deep learning system cross-analyzes the text semantics of the management report under different initial semantic attentions, aiming to reveal the intrinsic connections and laws between them. Through cross-analysis, the deep learning system can obtain two important results: one is the semantic involvement vector between each initial semantic attention, which reflects the degree of association and mutual influence between different semantic attentions; the other is the event influence vector under each initial semantic attention, which reveals the degree and trend of the influence of events on the report text under different semantic attentions. These two results provide strong support for the deep learning system to deeply understand the report text.

[0178] Finally, by integrating the semantic heat map of the management report text of each initial semantic attention, the semantic involvement vectors between each initial semantic attention, and the event impact vector under each initial semantic attention, the deep learning system can determine the original semantic attention from the numerous initial semantic attentions. The original semantic attention refers to the most core and important semantic area in the report text, which plays a vital role in understanding and grasping the overall meaning and value of the report. The process of determining the original semantic attention is a comprehensive consideration process. The deep learning system needs to comprehensively consider multiple factors such as the attention distribution in the heat map, the degree of association in the semantic involvement vector, and the influence trend in the event impact vector. Through this process, the deep learning system can more accurately grasp the key information and potential connections of the report, and provide a more accurate and powerful decision-making basis for hospital management.

[0179] It can be seen that by comprehensively using multiple technical means such as feature analysis, cross-analysis, and determination of original semantic attention of management report text semantics, a new idea and method is provided for mining and learning hospital management experience. This method can not only help hospitals better understand and grasp their own management practices and lessons learned, but also provide useful reference and guidance for current hospital management. At the same time, through in-depth analysis and mining of management report text semantics, this method can also reveal key information and potential connections in the report, providing more accurate and powerful decision-making basis for hospital management.

[0180] Further, Figure 2 Schematic diagram of the structure of a deep learning system 200 provided by the present invention. Figure 2 The deep learning system 200 shown includes a processor 210, which can call and run a computer program from a memory 230 to implement the method of the present invention.

[0181] Alternatively, if Figure 2 As shown, the deep learning system 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method of the present invention.

[0182] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0183] Alternatively, if Figure 2 As shown, the deep learning system 200 may further include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices. Specifically, it may send information or data to other devices, or receive information or data sent by other devices.

[0184] Optionally, the deep learning system 200 may implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the present invention, which will not be described in detail here for the sake of brevity.

[0185] It should be understood that the processor of the present invention may be an integrated circuit chip having signal processing capabilities.

[0186] It is understood that the memory in the present invention can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.

[0187] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and the program or instruction implements the steps of the above method when executed by a processor.

[0188] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0189] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0190] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the purpose of the present invention and the scope of protection of the present invention, all of which are within the protection of the present invention.

Claims

1. A method for retelling management report text based on digital medicine, characterized in that: The method is implemented by a deep learning system, and the method includes: Performing text semantic mining on the hospital management report to be processed to obtain the management report text semantics corresponding to the hospital management report to be processed under multiple target semantic attentions; Determine a target adverse event label set corresponding to the management report text semantics of the hospital management report to be processed under each target semantic attention; The key event status codes of the hospital management report to be processed in the adverse event label set corresponding to each target semantic attention are loaded into the text paraphrase annotation algorithm for paraphrase annotation analysis to obtain the local paraphrase annotation suggestion features of the hospital management report to be processed under each target semantic attention; wherein the key event status codes include the time, location, involved personnel and execution measures of the event, and the local paraphrase annotation suggestion features reflect the core content of the paraphrase annotation suggestion, which is represented by a series of vectors, each of which corresponds to different aspects of a paraphrase annotation suggestion, including event description, cause analysis and consequence evaluation; The paraphrase annotation decision vector of the hospital management report to be processed is determined by utilizing the local paraphrase annotation suggestion features under each of the target semantic attentions; wherein the paraphrase annotation decision vector is a high-dimensional vector formed by integrating all the local paraphrase annotation suggestion features, and the paraphrase annotation decision vector is used to represent the paraphrase annotation decision of the entire hospital management report; numerically, each dimension of the paraphrase annotation decision vector corresponds to a paraphrase annotation suggestion or its attribute, and the value of the paraphrase annotation decision vector reflects the importance and urgency of the paraphrase annotation suggestion or its attribute.

2. The method according to claim 1, characterized in that The step of performing text semantic mining on the hospital management report to be processed is implemented through a Transformer model, and the Transformer model is configured with a self-attention mechanism to capture long-distance dependencies in the hospital management report to be processed.

3. The method according to claim 2, characterized in that The step of determining the target adverse event label set corresponding to the management report text semantics of the hospital management report to be processed under each target semantic attention includes: The management report text semantics is compared and analyzed with a preset adverse event classification knowledge base to obtain a target adverse event label set corresponding to the management report text semantics under each target semantic attention.

4. The method according to claim 3, characterized in that The key event status encoding of the hospital management report to be processed in the adverse event label set corresponding to each target semantic attention is loaded into the text paraphrase annotation algorithm for paraphrase annotation analysis to obtain the local paraphrase annotation suggestion features of the hospital management report to be processed under each target semantic attention, including: Determine the key event status codes for each adverse event label set; The key event state encoding is structuredly paraphrased by a text paraphrase annotation algorithm based on a Seq2Seq learning framework to obtain local paraphrase annotation suggestion features of the hospital management report to be processed under each target semantic attention; The method of using the local paraphrase annotation suggestion features under each target semantic attention to determine the paraphrase annotation decision vector of the hospital management report to be processed includes: Based on the correlation and conflict between the local paraphrase annotation suggestion features under each of the target semantic attentions, a paraphrase annotation decision vector for the hospital management report to be processed is generated.

5. A method for retelling management report text based on digital medicine, characterized in that: The method is implemented by a deep learning system, and the method includes: Acquire the management report text semantics of multiple hospital management report examples under the same original semantic attention, perform event feature clustering on each of the management report text semantics belonging to the same original semantic attention, and obtain the adverse event label set corresponding to each of the management report text semantics, and the key event status code of each of the adverse event label sets; wherein the key event status code includes the time of occurrence, location, involved personnel and execution measures of the event; Based on the adverse event label set contained in the same original semantic attention, and the number of positive debugging examples and the number of negative debugging examples of the hospital management report examples to which the management report text semantics contained in each adverse event label set belongs, determine the risk trend prediction weight of each original semantic attention; From the original semantic attention, sampling the target semantic attention whose risk trend prediction weight meets the trend prediction requirement; The key event status encoding of each of the hospital management report examples in the adverse event label set corresponding to each of the target semantic attentions is loaded into the original text retelling annotation algorithm to obtain the possibility of each of the hospital management report examples belonging to a positive debugging example or a negative debugging example; wherein the positive debugging example refers to a report example in which effective measures are taken to deal with or resolve adverse events, and the negative debugging example refers to a report example in which effective response measures are not taken or the response is ineffective; The debugging examples of each of the hospital management report examples are used to distinguish annotations and the discrimination possibilities, and the algorithm weights of the original text paraphrase annotation algorithm are optimized to obtain a text paraphrase annotation algorithm, which is used to process the local paraphrase annotation suggestion features of the hospital management report to be processed under each of the target semantic attentions; wherein the local paraphrase annotation suggestion features reflect the core content of the paraphrase annotation suggestion, which is represented by a series of vectors, each vector corresponding to a different aspect of a paraphrase annotation suggestion, including event description, cause analysis and consequence evaluation.

6. The method according to claim 5, characterized in that The event feature clustering is performed on each of the management report text semantics belonging to the same original semantic attention to obtain the adverse event label set corresponding to each of the management report text semantics and the key event status code of each of the adverse event label sets, including: For each of the original semantic attentions, determining a semantic feature relationship network through the semantics of each of the management report texts corresponding to the original semantic attention; According to the relationship path in the semantic feature relationship network, the event feature label induction is performed on the semantics of each management report text to obtain multiple adverse event label sets under the original semantic attention, and the adverse event label sets corresponding to each of the semantics of the management report text are determined; Determine the key event status code of each of the adverse event label sets by using the number of positive debugging examples and the number of negative debugging examples of the hospital management report examples to which the management report text semantics contained in each of the adverse event label sets belong; The key event status codes of each of the adverse event label sets are determined by using the number of positive debugging examples and the number of negative debugging examples of the hospital management report examples to which the management report text semantics contained in each of the adverse event label sets belong, including: Determine the number of positive debugging examples and the number of negative debugging examples in the hospital management report examples to which the management report text semantics contained in each of the adverse event label sets belongs; Determine a first calculation result of the number of positive debugging examples in each of the adverse event label sets and the cumulative value of the positive debugging examples in the hospital management report examples, and a second calculation result of the number of negative debugging examples in each of the adverse event label sets and the cumulative value of the negative debugging examples in the hospital management report examples; For each of the adverse event label sets, the first operation result and the second operation result are subjected to quantitative feature mapping to determine the key event status code of each of the adverse event label sets.

7. The method according to claim 5, characterized in that The method of determining the risk trend prediction weight of each of the original semantic attentions based on the adverse event label set contained in the same original semantic attention and the number of positive debugging examples and the number of negative debugging examples of the hospital management report examples to which the management report text semantics contained in each of the adverse event label sets belongs, comprises: Determine a first calculation result of the number of positive debugging examples in each of the adverse event label sets and the cumulative value of the positive debugging examples in the hospital management report examples, and a second calculation result of the number of negative debugging examples in each of the adverse event label sets and the cumulative value of the negative debugging examples in the hospital management report examples; For each of the adverse event label sets, comparing and analyzing the first operation result and the second operation result to obtain a comparison characteristic coefficient of each of the adverse event label sets; Determine the local trend deduction coefficient of each adverse event label set based on the key event state code of each adverse event label set contained in the same original semantic attention and the comparison feature coefficient of each adverse event label set; For each of the original semantic attentions, the local trend deduction coefficients of the included adverse event label sets are integrated to obtain the risk trend prediction weights of each of the original semantic attentions.

8. The method according to claim 5, characterized in that The method of using the debugging examples of each of the hospital management report examples to distinguish annotations and the discrimination possibilities, optimizing the algorithm weights of the original text paraphrase annotation algorithm, and obtaining the text paraphrase annotation algorithm includes: Using the debugging examples of each of the hospital management report examples to distinguish annotations and the discrimination possibilities, a semantic mining error index is generated, and the indicator feature distribution of the semantic mining error index in a preset feature space is determined as a semantic paraphrase error feature; Using the debugging examples of each of the hospital management report examples to distinguish annotations and the discrimination possibilities, as well as the semantic paraphrase error characteristics, the algorithm weight optimization of the original text paraphrase annotation algorithm is performed to obtain a text paraphrase annotation algorithm; The method uses the debugging examples of each of the hospital management report examples to distinguish annotations and the discrimination possibilities, as well as the semantic paraphrase error features, to optimize the algorithm weights of the original text paraphrase annotation algorithm to obtain a text paraphrase annotation algorithm, including: Determine a discrimination debugging offset by using the debugging example discrimination annotations and the discrimination possibility of each of the hospital management report examples; Determining a word-level semantic confidence factor by discriminating a debugging offset gradient and the semantic repetition error feature gradient; The word-level semantic debugging error is determined according to the discrimination debugging offset and the word-level semantic confidence factor, and the algorithm weight of the original text paraphrase tagging algorithm is optimized by the word-level semantic debugging error to obtain a text paraphrase tagging algorithm.

9. The method according to claim 5, characterized in that The multiple hospital management report examples belong to the target digital hospital system; the method also includes: Determine the initial screening possibility of not meeting the management mode of the target digital hospital system from the discrimination possibilities obtained through the key event status coding; The target semantic attention matched with the key event state encoding of the possibility of the initial screening is determined as the semantic attention to be corrected; According to the semantic keywords of the semantic attention to be corrected, a correction strategy matching the semantic keywords is selected to perform a correction operation on the semantics of the management report text belonging to the semantic attention to be corrected, so as to obtain the semantics of the management report text after the correction operation; Performing event feature clustering again through the semantics of the management report text after the correction operation to determine the adverse event label set corresponding to the semantics of the management report text after the correction operation; Among them, the semantic keyword of the semantic attention to be corrected is the risk of leakage of hazardous chemicals; the correction strategy screened and matched with the semantic keyword performs a correction operation on the semantics of the management report text belonging to the semantic attention to be corrected, and obtains the semantics of the management report text after the correction operation, including: determining a correction strategy based on spatiotemporal resource scheduling to perform a correction operation on the semantics of the management report text belonging to the semantic attention to be corrected, and obtains the semantics of the management report text after the spatiotemporal feature correction operation; determining a security level correction strategy to perform a correction operation on the semantics of the management report text belonging to the semantic attention to be corrected, and obtains the semantics of the management report text after the security level correction operation; wherein the semantics of the management report text after the correction operation includes the semantics of the management report text after the spatiotemporal feature correction operation, and the security level correction operation. The management report text semantics after the operation; the method also includes: determining the first risk trend prediction weight of the semantic attention to be corrected through the management report text semantics after the spatiotemporal feature correction operation, and determining the second risk trend prediction weight of the semantic attention to be corrected through the management report text semantics after the security level correction operation; determining the management report text semantics corresponding to the target prediction weight from the first risk trend prediction weight and the second risk trend prediction weight as the candidate management report text semantics under the semantic attention to be corrected; performing event feature clustering on each of the candidate management report text semantics of the semantic attention to be corrected, and obtaining the adverse event label set corresponding to each of the candidate management report text semantics, and the key event status code of each of the adverse event label sets.

10. A deep learning system, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images