Neurosurgery nursing record key event extraction method and system based on natural language processing

Abstract

The invention provides a neurosurgery nursing record key event extraction method and system based on natural language processing, and relates to the technical field of artificial intelligence and natural language process.The method comprises the steps that candidate terminologies are recognized through a neurosurgery field terminology library based on a word vector sequence; based on the candidate terminologies, extracting core semantic vectors of the candidate terminologies, and associating corresponding entity type embedding vectors and event type embedding vectors; and performing multi-modal fusion and nonlinear transformation on the core semantic vector, the entity type embedding vector and the event type embedding vector to obtain a semantic enhanced term sequence. According to the method, the unstructured neurosurgical nursing records are automatically converted into the structured key event report capable of directly supporting clinical decisions, and the core problems of semantic fuzziness, information fragmentation and poor practicability in an original text are solved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, and in particular to a method and system for extracting key events from neurosurgical nursing records based on natural language processing. Background Technology

[0002] Neurosurgical intensive care units (ICUs) generate a large volume of unstructured nursing documents daily, containing crucial information for assessing patient conditions and guiding treatment. Currently, clinicians rely on manual sifting and summarizing key events from massive amounts of text, which is inefficient and may fail to meet the timeliness requirements of emergency care in the event of sudden changes in patient conditions. Although Natural Language Processing (NLP) technology offers the possibility of automated extraction, general-purpose models have the following limitations at the data processing level when dealing with neurosurgical specialized texts: First, at the data understanding level, the recognition rate for a large number of domain terms and abbreviations such as GCS scores and brain herniation is low, and it is unable to establish semantic connections within context. For example, associating a decrease in blood pressure with headache relief may lead to fragmented information. Second, at the data standardization level, it has poor adaptability to the diverse forms of expression in nursing records (such as ICP 25cmH2O, high intracranial pressure) and colloquial descriptions, which may lead to inconsistent extraction results. Finally, at the data output and application level, existing methods output event classifications with coarse granularity, are not deeply bound to specific indicators and timestamps, and are mostly offline batch processing, which cannot provide real-time, structured decision support. For example, in suspected cases of brain herniation, existing technology may have difficulty automatically associating the gradual change from slightly enlarged to dilated pupils with the fluctuation trend of blood pressure, and cannot quickly generate a coherent chain of evidence for the evolution of the disease. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and system for extracting key events from neurosurgical nursing records based on natural language processing, which automatically transforms unstructured neurosurgical nursing records into structured key event reports that can directly support clinical decision-making, thus solving the core problems of semantic ambiguity, information fragmentation and poor usability in the original text.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: In a first aspect, a method for extracting key events from neurosurgical nursing records based on natural language processing is provided. The method includes: identifying candidate professional terms based on word vector sequences using a neurosurgical terminology database; extracting the core semantic vector of the candidate professional terms and associating it with the corresponding entity type embedding vector and event type embedding vector; and performing multimodal fusion and nonlinear transformation on the core semantic vector, entity type embedding vector, and event type embedding vector to obtain a semantically enhanced terminology sequence. Based on semantically enhanced term sequences, a heterogeneous graph integrating syntactic and semantic information is constructed; the heterogeneous graph is evaluated for importance and its structure is divided to obtain topological features and metrics; attention weight coefficients are obtained based on topological features and metrics; structural features and key semantic features are extracted based on the heterogeneous graph and attention weight coefficients, respectively; the fusion weights of structural features and key semantic features are dynamically optimized and weighted to obtain optimized key event features. Based on the optimized key event features, normalization mapping is performed to obtain standardized events; core event sentences in the standardized events are identified, and missing arguments are filled in based on semantic similarity matching; the standardized events and the filled arguments are integrated to obtain a complete list of event arguments; Based on the complete list of event arguments, fine-grained event classification is performed to obtain event type labels; based on the event type labels, structured association information of each event is identified and extracted; the event type labels and structured association information are integrated to obtain structured key event data; Based on the structured key event data, a key event report is generated and output.

[0005] Secondly, a key event extraction system for neurosurgical nursing records based on natural language processing includes: The semantic enhancement module is used to identify candidate professional terms based on word vector sequences using a terminology database in the neurosurgical field; based on the candidate professional terms, its core semantic vector is extracted and associated with the corresponding entity type embedding vector and event type embedding vector; the core semantic vector, entity type embedding vector, and event type embedding vector are subjected to multimodal fusion and nonlinear transformation to obtain a semantically enhanced term sequence. The optimization module is used to construct a heterogeneous graph that integrates syntactic and semantic information based on semantically enhanced term sequences; to evaluate the importance and divide the structure of the heterogeneous graph to obtain topological features and metrics; to obtain attention weight coefficients based on topological features and metrics; to extract structural features and key semantic features based on the heterogeneous graph and attention weight coefficients; and to dynamically optimize the fusion weights of structural features and key semantic features and perform weighted fusion to obtain optimized key event features. The standardization and completion module is used to perform normalization mapping based on optimized key event features to obtain standardized events; identify core event sentences in the standardized events and complete missing arguments based on semantic similarity matching; integrate standardized events and completed arguments to obtain a complete list of event arguments; the classification and extraction module is used to perform fine-grained event classification based on the complete list of event arguments to obtain event type labels; identify and extract structured association information for each event based on the event type labels; and integrate the event type labels and structured association information to obtain structured key event data; The output module is used to generate and output a key event report based on the structured key event data.

[0006] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0007] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0008] The above-described solution of the present invention has at least the following beneficial effects: Candidate terms are precisely located based on a neurosurgical terminology database to ensure domain-specific adaptability of terminology recognition. By fusing core semantic vectors with entity and event type embedding vectors, terms not only carry their own semantics but also associate them with their category information. Multimodal fusion and nonlinear transformation of the semantically enhanced term sequence provide semantically clear and informationally complete foundational data for subsequent processing. Precise transformation from term sequence to optimized features is achieved, and the constructed heterogeneous graph integrates both syntactic and semantic information, making the relationships between terms more comprehensive. Importance assessment and structural partitioning focus on core data, and the generated attention weight coefficients guide feature extraction towards key regions. After extracting structural and key semantic features separately, the fusion weights are dynamically optimized to ensure the adaptability of the two types of features. The final optimized key event features retain topological relationships while highlighting core semantics, improving the accuracy of subsequent event processing. Standardization and completion operations enhance the standardization and completeness of event information, and normalization mapping transforms diverse expressions into standard forms. This system transforms event data, eliminating data confusion caused by differences in expression; core event sentence identification quickly locates high-value text, while argument completion based on semantic similarity aggregates scattered contextual information to the corresponding event; the integrated complete event argument list ensures that the core elements of each event are complete, providing comprehensive data support for subsequent classification processing; focusing on the structuring and refinement of event data, fine-grained event classification combines argument composition and semantic features, making event type division more aligned with actual clinical needs; targeted extraction of structured correlation information such as quantitative indicators and occurrence time clarifies the correspondence between events and core data; the integrated structured key event data deeply binds type labels with specific information, forming a clear and organized data format; it completes the transformation of structured data into practical results, with key event reports generated based on structured key event data serving as a centralized presentation of the results of previous data processing; the report format adapts to clinical review and decision-making needs, transforming scattered event data into a systematic and intuitive information carrier. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a method for extracting key events from neurosurgical nursing records based on natural language processing, as provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of a key event extraction system for neurosurgical nursing records based on natural language processing, provided by an embodiment of the present invention. Detailed Implementation

[0011] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0012] like Figure 1 As shown, embodiments of the present invention propose a method for extracting key events from neurosurgical nursing records based on natural language processing. The method includes the following steps: Step 100: Based on the word vector sequence, candidate professional terms are identified using a terminology database in the field of neurosurgery; based on the candidate professional terms, their core semantic vectors are extracted and associated with the corresponding entity type embedding vectors and event type embedding vectors; the core semantic vectors, entity type embedding vectors, and event type embedding vectors are subjected to multimodal fusion and nonlinear transformation to obtain a semantically enhanced term sequence. Step 200: Based on the semantically enhanced term sequence, construct a heterogeneous graph that integrates syntactic and semantic information; evaluate the importance and divide the structure of the heterogeneous graph to obtain topological features and metrics; obtain attention weight coefficients based on topological features and metrics; extract structural features and key semantic features based on the heterogeneous graph and attention weight coefficients respectively; dynamically optimize the fusion weights of structural features and key semantic features, and perform weighted fusion to obtain optimized key event features; Step 300: Based on the optimized key event features, perform normalization mapping to obtain standardized events; identify the core event sentences in the standardized events and complete the missing arguments based on semantic similarity matching; integrate the standardized events and the completed arguments to obtain a complete list of event arguments; Step 400: Based on the complete list of event arguments, perform fine-grained event classification to obtain event type labels; based on the event type labels, identify and extract structured association information for each event; integrate the event type labels and structured association information to obtain structured key event data; Step 500: Based on the structured key event data, generate and output a key event report.

[0013] In this embodiment of the invention, candidate professional terms are identified based on a neurosurgical terminology database, enhancing the domain adaptability of terminology identification and ensuring the targeted capture of professional expressions. By extracting core semantic vectors and associating them with entity and event type embedding vectors, the semantic representation dimensions of terms are enriched, making terminology information more complete. Multimodal fusion and nonlinear transformation processing methods achieve deep integration of vector information from different dimensions, improving the semantic expression accuracy of terminology sequences. A heterogeneous graph integrating syntactic and semantic information is constructed to achieve structured integration of the two key information types, making data association clearer. Topological features and metrics are obtained through heterogeneous graph importance assessment and structural partitioning, accurately locating core association information in the data. Structural features and key semantic features are extracted based on attention weight coefficients, highlighting the role of high-value information and reducing redundant data interference. Dynamic optimization of fusion weights and weighted fusion make feature combinations more closely match the essence of the data, improving the representativeness and effectiveness of key event features. The key event features are normalized and mapped to obtain... Standardized events unify data formats, improving data consistency and standardization; core event sentence identification quickly identifies the core expressive carriers of events, enhancing the focus of event extraction; semantic similarity matching completes missing arguments, supplementing event information and avoiding omissions of key information; standardized events and completed arguments form a complete event argument list, presenting event information in a structured manner; fine-grained event classification is performed based on the complete event argument list, enabling the division of event types and meeting the detailed clinical needs for event classification; structured correlation information is extracted based on event type tags, establishing a correspondence between event types and related data, making data logic clearer; event type tags and structured correlation information are integrated into structured key event data, achieving systematic organization of event information and improving data usability; reports are generated based on structured key event data, ensuring reliable source and clear organization of report content; the output key event reports directly present core information, improving information transmission efficiency.

[0014] In a preferred embodiment of the present invention, before step 100, the method further includes: Preprocessing of neurosurgical nursing record texts to generate standardized texts includes: First, performing the acquisition of nursing record texts by establishing a data interaction interface with the hospital's existing electronic medical record system to retrieve the original text data of neurosurgical nursing records in batches or in real time. The retrieved text data covers patient basic information, vital sign monitoring data, symptom descriptions, treatment operation records, and complication observation results, as well as other content related to the assessment of the patient's condition.

[0015] Secondly, noise reduction processing is performed on the acquired raw text. This step mainly removes redundant content in the text that is not related to medical information. Specifically, this includes deleting special symbols in the text, such as meaningless separators and misused punctuation marks, filtering out non-core information such as nursing staff signatures and notes, and removing blank characters and duplicate record fragments caused by input errors, so as to ensure that the text content focuses on the core information related to the patient's condition.

[0016] Furthermore, the standardization of textual descriptions was carried out, a process closely addressing the issue of diverse forms of expression in nursing records. On one hand, a table of terms and abbreviations in the neurosurgical field was established. Based on this table, abbreviations such as ICP were uniformly converted to the complete expression of intracranial pressure, and GCS was converted to the Glasgow Coma Scale. Simultaneously, colloquial descriptions such as "the patient is a little dizzy" or "the breathing is slightly rapid" were converted to standardized medical written descriptions such as "the patient complains of dizziness and a slightly rapid breathing rate." On the other hand, different expressions of the same indicator in the text were standardized. For example, expressions such as "high intracranial pressure," "ICP 25cmH2O," and "intracranial pressure exceeding the normal range" were all converted to a unified form including specific numerical values ​​and standard units, i.e., "intracranial pressure 25cmH2O," ensuring consistency in the expression of the same type of medical information.

[0017] Finally, the text undergoes grammatical and semantic correction. Grammatical errors caused by rapid data entry in the nursing records, such as reversed word order and incomplete components, are adjusted. For example, "patient given antihypertensive medication has high blood pressure" is corrected to "patient has high blood pressure and was given antihypertensive medication." At the same time, ambiguous expressions are clarified. Based on the expression habits of common neurosurgical diseases, expressions such as "pupils are slightly enlarged" are supplemented to "pupil diameter is slightly enlarged, about 3mm." Finally, a standardized text with uniform format, standard expression, and complete information is generated, providing high-quality input data for subsequent contextual semantic encoding.

[0018] Standardized text is semantically encoded to obtain word vector sequences containing dynamic semantics within the context. Specifically, this involves: First, the standardized text is segmented using a segmentation strategy based on a neurosurgical domain dictionary. This dictionary pre-includes common neurosurgical disease names such as brain herniation and cerebral hemorrhage, examination indicators such as GCS scores, and treatment procedures such as decompression craniectomy. This ensures that domain-specific vocabulary can be completely extracted during the segmentation process, avoiding situations where the prodromal symptoms of brain herniation are split into brain herniation, prodrome, and symptoms, which would affect semantic understanding.

[0019] Based on this, contextual semantic encoding is performed. The word sequence obtained after word segmentation is used as the processing object. The word sequence is truncated according to the natural paragraphs and sentence order of the text using a sliding window of preset length. The length of the sliding window can be set according to the characteristics of the sentences in the neurosurgical nursing record. It usually covers a single sentence and parts of the sentences before and after it, ensuring that the window contains complete contextual information of the current word. For example, when processing the word "blood pressure drop", the sliding window will simultaneously cover the patient's headache today is relieved compared to yesterday and the heart rate is stable, so that the encoding process can capture the semantic relationship between this contextual information and the blood pressure drop.

[0020] During the encoding process, we focus on the semantic changes of words in different contexts. Combining the medical logic of neurosurgery, we strengthen the semantic association weight between related medical information. For example, we increase the association weight between vital signs such as blood pressure and heart rate and symptoms such as headache and altered consciousness. This ensures that the word vectors generated after encoding not only contain the basic semantics of the words themselves, but also carry their dynamic semantics in the specific description of the condition and their association with surrounding information.

[0021] It should be noted that this encoding process, through comprehensive integration of contextual information, enables the generated word vector sequence to accurately reflect the logical relationships of the patient's condition description in the nursing record. For example, when processing text related to suspected cases of brain herniation, the semantics of words such as "slightly enlarged pupils," "fixed pupil dilation," and "fluctuating blood pressure" can be associated and mapped through encoding, laying the foundation for the association extraction of this information in subsequent steps, and finally outputting a word vector sequence containing dynamic semantics in the context.

[0022] In a preferred embodiment of the present invention, step 100 above involves identifying candidate professional terms based on a neurosurgical terminology database using a word vector sequence; extracting the core semantic vector of each candidate professional term and associating it with the corresponding entity type embedding vector and event type embedding vector; and performing multimodal fusion and nonlinear transformation on the core semantic vector, entity type embedding vector, and event type embedding vector to obtain a semantically enhanced terminology sequence, including: Step 101: Traverse the word vector sequence and extract the context word vector of each word as the semantics of each word. Specifically, this includes: using the word vector sequence as the calculation object, where each word vector in the sequence is presented as a fixed-dimensional numerical array; during calculation, a sequential traversal algorithm is first used to read the basic word vector corresponding to each word from the start position to the end position of the word vector sequence; to obtain the calculation result containing contextual semantics, a context window is set for each word. The window size is preset to 7 word vector units based on the average length of the neurosurgical nursing record sentences, that is, it contains the word vectors of the current word and the three adjacent words before and after it; then, through a vector extraction algorithm, the 7 word vectors in the window are extracted as a set of associated data. After extraction, the mean vector of the set of vectors is calculated by adding each element and taking the average. This mean vector is the calculation result of the current word carrying contextual semantics, which serves as the basis data for subsequent similarity comparison.

[0023] Step 102 involves comparing the semantics of each word with the semantic vectors of predefined standard terms, abbreviations, and synonyms in a neurosurgical terminology database. By calculating the semantic similarity between vectors, a set of similarity matching values ​​is generated for each word. Specifically, this includes: based on the semantics of each word, introducing a pre-constructed neurosurgical terminology database. The construction of this database and the pre-definition of semantic vectors are closely aligned with the clinical needs of neurosurgery, providing a reliable foundation for subsequent accurate comparisons. Specifically, the pre-definition process of the semantic vectors of predefined standard terms, abbreviations, and synonyms is as follows: First, by reviewing authoritative neurosurgical medical dictionaries, clinical guidelines, and real nursing records from the neurosurgical ICUs of tertiary hospitals within the past five years, core standard terms such as GCS score, brain herniation, and ICP are systematically collected, while commonly used clinical abbreviations are simultaneously compiled. The terminology is structured in several ways. First, it includes terms like ICP (intracranial pressure) and EVD (external ventricular drainage), along with synonyms such as pupillary dilation and pupillary vasodilation, forming a complete terminology set. Second, using this terminology set as the core, a corpus of 100,000 annotated neurosurgical nursing records is constructed. All texts in the corpus have been reviewed by clinical medical staff to ensure the authenticity and accuracy of the terminology's usage scenarios. Third, the corpus is trained using the Word2Vec tool, with a vector dimension of 256. A sliding window is used to capture the semantic relationships of terms in different contexts, generating initial semantic vectors. Finally, the initial semantic vectors are normalized to ensure that the semantic vectors of all terms, abbreviations, and synonyms are within the same numerical range. These are then categorized and stored in a neurosurgical terminology database, completing the pre-definition of semantic vectors.

[0024] Based on the pre-calculated and stored semantic vectors, during the calculation, a loop comparison program is initiated for the semantic vector of each word, sequentially calling the semantic vector of each predefined term in the terminology database and performing pairing calculations with the current word vector. The similarity calculation adopts the core logic of cosine similarity, which is suitable for this scenario because it can effectively reflect the degree of semantic association in the vector space. In the specific operation, the sum of the products of the corresponding dimension values ​​of the two vectors is first calculated to obtain the vector dot product; then the sum of the squares of the values ​​of each dimension of each vector is calculated, and the square root is taken to obtain the vector magnitude; finally, the dot product result is divided by the product of the two magnitudes to obtain a similarity value between 0 and 1. The closer the value is to 1, the stronger the semantic association between the two. After each pairing calculation is completed, the similarity value is bound to the identification information of the corresponding predefined term and stored in the matching value set of the corresponding word until the word is compared with all terms in the terminology database, and finally a similarity matching value set containing all pairing results is formed.

[0025] Step 103: From the set of similarity matching values, select all words corresponding to matching values ​​exceeding a preset similarity threshold, and mark these words as candidate professional terms. Specifically, this includes: using the set of similarity matching values ​​as the basis for calculation, firstly, a similarity threshold needs to be preset according to the clinical accuracy requirements of neurosurgical terminology recognition. The preset process of this threshold is clinically practical as the core objective and is achieved through multiple rounds of sample debugging. Specifically, the preset work first selects 5,000 nursing record samples from the neurosurgical ICU of a tertiary hospital. These samples cover mainstream diseases such as traumatic brain injury and post-operative cerebral hemorrhage, and include writings from nursing staff at different levels to ensure that the samples have sufficient representativeness. Then, the samples are divided into training... The training set and test set were used, with 70% allocated to initial threshold screening and 30% to threshold verification. Then, based on the core requirements of clinicians for terminology recognition—avoiding missed detection of key terms such as prodromal symptoms of brain herniation while minimizing non-terminology misjudgments—an accuracy of ≥92% and a recall of ≥88% were set as acceptable threshold criteria. Based on this, different thresholds ranging from 0.7 to 0.9 were tested sequentially in the training set, and the recognition accuracy and recall corresponding to each threshold were recorded. It was ultimately found that when the threshold was 0.8, the recognition accuracy on the test set reached 93.5% and the recall reached 89.2%, fully meeting clinical accuracy requirements. Therefore, 0.8 was determined as the preset similarity threshold.

[0026] After determining the threshold, a numerical filtering program is initiated for each word's matching value set during calculation. This program iterates through each element, comparing every similarity value in the set with the preset threshold of 0.8, accurately filtering out all similarity results with values ​​greater than 0.8. Based on this, an index-based correlation calculation is used to connect the qualifying results with the original data. Specifically, the position index table of the word vector sequence is called, converting the index information corresponding to the qualifying value into the start and end positions of the word in the original word vector sequence. Simultaneously, the original text content of the word and the standard term name with the highest matching degree are marked, forming correlation data of position, text, and standard term. To further ensure the effectiveness and uniqueness of candidate terms, deduplication calculations are performed on the marked words. A temporary storage list is constructed, and the correlation data is stored sequentially in the list. If the same word is marked repeatedly, its corresponding similarity values ​​are compared, and only the set of correlation data with the highest value is retained, while the remaining duplicates are removed. After the above series of calculation operations, a set of candidate professional terms with no duplication and high correlation is finally formed.

[0027] Step 104: Extract the context word vectors corresponding to each candidate term from the word vector sequence to which the candidate term belongs, as the core semantic vector; simultaneously, according to the predefined mapping relationships in the neurosurgical terminology database, query and obtain the entity type embedding vectors and event type embedding vectors corresponding to the candidate term. Specifically, this includes: performing a dual calculation operation on the candidate term, wherein the association calculation between the term and entity type, and event type, relies on the pre-constructed mapping relationship table in the neurosurgical terminology database. The pre-definition process of this mapping relationship is closely integrated with neurosurgical clinical practice. The diagnostic and treatment logic and data processing requirements provide core support for the accurate acquisition of type vectors. Specifically, the pre-definition of mapping relationships is based on neurosurgical clinical diagnosis and treatment guidelines, authoritative medical textbooks, and standardized nursing specifications of neurosurgery departments in tertiary hospitals. The specific scope of two core types is systematically sorted out and clarified: entity types cover vital signs such as blood pressure and heart rate, disease names such as brain herniation and intracranial hemorrhage, treatment devices such as EVD drainage tubes, and drug names such as mannitol; event types include changes in condition such as worsening or relief of symptoms, treatment procedures such as suctioning and medication administration, and the occurrence of complications such as pulmonary infection and epileptic seizures.

[0028] Based on the established category classification, further association rules between terms and types are established: For standard terms, the corresponding type is determined directly based on their clinical attributes, such as GCS score corresponding to assessment indicators in entity type, and brain herniation corresponding to disease name in entity type and condition change in event type; For abbreviations and synonyms, they are first associated with their corresponding standard terms, and then the type mapping relationship of the standard terms is followed, such as ICP first matching to the standard term intracranial pressure, and then determining its corresponding entity type as vital sign indicators; Subsequently, a review group composed of 5 physicians with more than 10 years of clinical experience in neurosurgery and 3 NLP technology experts is invited to verify the initially established association relationships one by one. For easily confused terms such as routine postoperative care, the event type is clarified as nursing operation rather than treatment operation in combination with clinical scenarios to correct association biases; Finally, the verified standard terms, entity types, and event type correspondences are stored in the form of a structured data table, in which each standard term corresponds to a unique entity type code and event type code, completing the pre-definition of the mapping relationship.

[0029] Based on the predefined mapping relationship, the dual computation operation is carried out in an orderly manner. On the one hand, the calculation of the core semantic vector aims to ensure the consistency of semantic representation. Based on the position information of the candidate terms in the original word vector sequence, the context window setting that has been verified in step 101 is called, that is, a 7-word window covering 3 words before and after the term, and the same vector extraction algorithm, to re-extract the context word vector group corresponding to the term. Then, the mean calculation method that is completely consistent with step 101 is used to add the word vectors in the window dimension by dimension and take the average value to obtain the core semantic vector that can fully carry the semantic context of the term, ensuring the semantic consistency during subsequent multimodal fusion.

[0030] On the other hand, the acquisition of type embedding vectors relies on a mapping table. During calculation, the standard terms associated with candidate terms are first used as search keywords. A matching query program is then launched to traverse the mapping table and quickly locate the corresponding entity type code and event type code. Considering that the core semantic vector is preset to 256 dimensions, to ensure the feasibility of vector fusion, the encoding conversion program adopts a combination of one-hot encoding and dimension mapping: first, the type code is converted into a one-hot vector, and then, through a preset linear transformation matrix, the one-hot vector is mapped into a 256-dimensional embedding vector. Finally, entity type embedding vectors and event type embedding vectors with dimensions completely consistent with the core semantic vector are obtained. Through the above dual calculation, the contextual semantic features of candidate terms are preserved, and the precise binding of terms and clinical type attributes is achieved, completing the association calculation of terms and type vectors, laying a solid foundation for subsequent multimodal fusion.

[0031] Step 105: For each candidate term, the core semantic vector, entity type embedding vector, and event type embedding vector are concatenated to form a multimodal fusion vector. Specifically, this involves: first, verifying the consistency of the dimensions of the core semantic vector, entity type embedding vector, and event type embedding vector. These vectors are pre-defined as 256-dimensional numerical arrays to ensure the feasibility of the concatenation calculation; second, using a vector order concatenation algorithm, following a fixed order of core semantic vector first, then entity type embedding vector, and finally event type embedding vector, the numerical arrays of these vectors are sequentially concatenated end-to-end to form a new 768-dimensional numerical array; third, real-time verification using a dimension counting program during the concatenation process to ensure that the concatenated vector dimension is the sum of the dimensions of the three types of vectors, with no numerical loss or duplicate concatenation issues, ultimately resulting in a structurally complete multimodal fusion vector.

[0032] Step 106 involves inputting each multimodal fusion vector into a fully connected neural network layer to perform a linear transformation, obtaining the corresponding linear features. Specifically, this includes: inputting the multimodal fusion vector into a preset fully connected neural network layer for linear transformation calculation. This fully connected layer contains 256 neurons, each corresponding to a set of preset weight parameters and bias values. These parameters are determined and fixed through training with 10,000 labeled neurosurgical nursing record samples. During calculation, for each dimension value of the fusion vector, a multiplication operation is performed with the weight parameters of the corresponding neuron. All multiplication results are then accumulated, and finally, the bias value of the neuron is added to obtain the output value of a single neuron. The above operations for all 256 neurons are completed through parallel computation, ultimately outputting a 256-dimensional linear feature vector, achieving dimensional compression and feature regularization of the fusion vector.

[0033] Step 107 involves inputting the linear features into a nonlinear activation function for processing to obtain the corresponding activated features. Specifically, this includes inputting the linear features into a ReLU nonlinear activation function for calculation. The core purpose of this step is to enhance the nonlinear expressive power of the features, and the ReLU nonlinear activation function is the key tool for achieving this goal. The core function of the ReLU nonlinear activation function is reflected in three aspects: First, it can break the limitation that linear features can only reflect simple data associations and capture the complex nonlinear relationship between the semantics of terms and clinical attributes in neurosurgical nursing texts, such as the implicit association between blood pressure changes and intracranial pressure fluctuations. Second, it can effectively alleviate the gradient vanishing problem by processing specific values, ensuring the stability of information transmission in subsequent feature processing and avoiding the weakening of key clinical features during transmission. Third, it can quickly filter meaningless feature signals, focus on key information valuable for disease assessment, and improve the efficiency and targeting of feature processing.

[0034] Specifically, the calculation employs an element-wise traversal approach, evaluating each value in the linear feature vector. This operation directly echoes the logic of the ReLU function. If the value is greater than 0, it indicates that it carries clinically meaningful features relevant to neurosurgery, and the value is directly retained as the activation result to highlight these key features. If the value is less than or equal to 0, it indicates that it is an invalid or interfering feature, such as a feature transformed from redundant statements unrelated to the condition. In this case, it is replaced with 0 as the activation result, achieving rapid filtering of invalid information. Each activated value generated here is a selected and enhanced effective feature unit. Together, they form an activated feature vector that can carry the core semantics and clinical relevance of the terminology.

[0035] This calculation process requires no additional parameter adjustment and is completed through simple numerical judgment and replacement, which improves computational efficiency and ensures the stability of the results. It can efficiently focus on key features that are clinically meaningful for neurosurgery. The final output activation feature vector has the same dimension as the linear feature. This dimensionality consistency provides an adaptation basis for the subsequent conversion of multimodal fusion vectors into semantically enhanced vectors. It ensures that subsequent processing can generate semantic representations that are more in line with clinical needs based on these high-quality features, providing strong support for the extraction of key events.

[0036] Step 108: Based on the activation features, each multimodal fusion vector is converted into a corresponding semantic enhancement vector. Specifically, this includes: performing semantic enhancement calculations on the multimodal fusion vector based on the activation features. First, through a feature weight allocation procedure, the value of each dimension in the activation features is used as the weight coefficient for the corresponding position. The value range of the weight coefficient has been normalized and limited to between 0 and 1. During calculation, the value of each dimension in the multimodal fusion vector is multiplied by the weight coefficient of the corresponding dimension of the activation feature to obtain the enhanced dimension value. After all dimensions have undergone multiplication, a new 768-dimensional vector is formed. This vector retains the complete information of the fusion vector and, through weight adjustment, highlights the key semantics pointed to by the activation features; this is the semantic enhancement vector. After calculation, error verification is performed to ensure that the enhanced vector value is within a reasonable range, without any abnormal amplification or reduction.

[0037] Step 109 involves arranging the semantic enhancement vectors of all candidate technical terms according to their original order in the text, forming a semantically enhanced term sequence. Specifically, this includes: performing a sorting calculation based on the semantic enhancement vectors and the candidate technical term position information recorded in Step 103. First, a key-value pair data structure containing term positions and semantic enhancement vectors is constructed, where term positions are represented by character offsets in the original nursing record text. Then, a sorting algorithm is initiated, using the numerical value of the term positions as the sorting criterion, to sort all key-value pairs in ascending order, ensuring that the sorting result is completely consistent with the order of term appearance in the original text. After sorting, the semantic enhancement vectors from all key-value pairs are extracted and sequentially combined in sorting order to form a two-dimensional vector array, which is the semantically enhanced term sequence. Finally, the calculation is validated by randomly selecting the position information of three adjacent terms to confirm that their order in the sequence is consistent with the original text, ensuring the accuracy of the calculation results.

[0038] In this embodiment of the invention, the word vector sequence is traversed and the context word vector of each word is extracted as semantics. This abandons the processing method that relies solely on the static semantics of the words themselves, allowing the semantics of words to carry the associated information in the textual context, making the semantic interpretation of individual words more consistent with the specific expression scenario of nursing records. The semantics of words are compared one by one with the semantic vectors of standard terms, abbreviations, and synonyms in the neurosurgical terminology database, and the similarity is calculated. This achieves comprehensive coverage of diverse expressions in the field, breaking the limitation of matching only a single expression in general processing and improving the matching coverage of professional terms and related variants. Words that meet the matching value are selected by setting a preset similarity threshold and marked as candidate professional terms. This establishes clear terminology selection criteria, avoids low-similarity words from being mixed into the candidate set, and reduces the interference of invalid data on subsequent processing. The context word vectors of candidate professional terms are extracted as core semantic vectors to ensure the contextual relevance of semantic representation. At the same time, entity type and event type embedding vectors are obtained based on the predefined mapping relationship of the domain terminology database, so that terminology information is deeply bound to the corresponding type attributes, enriching the semantic dimension of terms and laying the foundation for multimodal fusion.

[0039] This method concatenates core semantic vectors with two types of embedded vectors to form a multimodal fusion vector, achieving the organic integration of semantic and attribute information. This breaks the information limitations of a single vector dimension, allowing the fusion vector to simultaneously carry the contextual semantics of terms, entity attributes, and event attributes, thus enhancing the vector's information carrying capacity. A fully connected neural network layer performs a linear transformation on the multimodal fusion vector, unifying and standardizing the feature dimensions of vectors from different sources, eliminating dimensional differences and representational biases between vectors, and ensuring that the transformed linear features have a consistent data format. Finally, processing the linear features using a nonlinear activation function captures complex nonlinear correlation information within the vector features, highlighting… By overcoming the limitation that linear transformations can only reflect simple relationships, activation features are made to better align with the intrinsic connections between term semantics and attributes, thus enhancing the expressive power of features. Based on activation features, multimodal fusion vectors are converted into semantically enhanced vectors, achieving feature optimization and semantic enhancement of the original fusion vectors. This ensures that the final vector representation retains core information while highlighting key features, improving the accuracy of the vector in expressing term semantics. The semantically enhanced vectors are arranged according to the order of candidate professional terms in the original text, fully preserving the inherent semantic logic and temporal order of the text, avoiding logical confusion in the term sequence, and ensuring that the generated semantically enhanced term sequence is consistent with the logical description of the patient's condition in the nursing record.

[0040] In a preferred embodiment of the present invention, step 200 involves constructing a heterogeneous graph that integrates syntactic and semantic information based on a semantically enhanced term sequence; evaluating the importance and structurally partitioning the heterogeneous graph to obtain topological features and metrics; obtaining attention weight coefficients based on the topological features and metrics; extracting structural features and key semantic features based on the heterogeneous graph and attention weight coefficients; dynamically optimizing the fusion weights of the structural features and key semantic features, and performing weighted fusion to obtain optimized key event features, including: Step 201: Based on the semantically enhanced term sequence, parse the syntactic structure between terms in the sequence to obtain syntactic relation edges between terms; simultaneously, calculate the semantic similarity between the term vectors in the sequence, and generate semantic relation edges between terms according to a preset threshold. Specifically, this includes: based on the semantically enhanced term sequence, the core task is to mine the dual association relationships between terms to provide a foundation for subsequent heterogeneous graph construction; firstly, syntactic structure parsing is carried out, using the StanfordParser tool pre-trained on neurosurgical nursing text to divide the term sequence into paragraphs according to the original text sentences, and parsing the syntactic relationships between terms such as subject-predicate, verb-object, and modification segment by segment. For example, identify the relationship of blood pressure drop as subject and relief as predicate, or brain herniation as an attributive modifying symptoms. Mark each pair of terms with direct syntactic association as an ordered pair, generate corresponding syntactic relation edges, and label the attributes of the edges with specific syntactic types.

[0041] Simultaneously, the semantic similarity between terms is calculated. Taking the semantic enhancement vector of each term as the calculation object, the mature cosine similarity calculation logic in step 102 is used to pair term vectors in the sequence one by one. Combining the clinical characteristics of neurosurgical term association, a preset similarity threshold of 0.7 is set. When the similarity value of paired terms exceeds this threshold, it is determined that there is a semantic association between the two. For example, if the vector similarity between pupil dilation and intracranial pressure increases reaches the standard, a semantic relationship edge is generated. The attribute of the edge records the specific similarity value. Finally, the syntactic relationship edge set and the semantic relationship edge set are output simultaneously to ensure that both explicit syntactic associations and implicit semantic associations between terms are fully captured.

[0042] Step 202 involves using each term in the semantically enhanced term sequence as a graph node, and using the syntactic and semantic relation edges as edges connecting these nodes to construct a heterogeneous graph that integrates syntactic and semantic information. Specifically, this includes: first, defining graph elements by treating each term in the semantically enhanced term sequence as an independent graph node, with each node containing the semantic enhancement vector and corresponding clinical attribute labels such as signs, indicators, and disease names; then, establishing the association between edges and nodes by mapping the syntactic and semantic relation edges to two types of edges in the heterogeneous graph, where syntactic relation edges connect nodes with direct syntactic dependencies, and semantic relation edges connect nodes with semantic associations that meet the criteria.

[0043] To distinguish the information value of the two types of edges, attributes are labeled on the edges: syntactic relation edges are labeled with specific relation types such as subject-verb-verb-object, while semantic relation edges are labeled with corresponding similarity values. During the construction process, a graph structure verification procedure is used to ensure that each node is associated with at least one edge, avoiding the occurrence of isolated nodes. If a term has no directly associated edge, a semantic relation edge with similar terms is added based on clinical common sense, such as establishing an association between a GCS score that appears alone and a term of consciousness. The resulting heterogeneous graph simultaneously carries syntactic logic and semantic association, effectively solving the problem of information fragmentation and providing a structured carrier for subsequent associated feature extraction.

[0044] Step 203: Evaluate the importance of all nodes in the heterogeneous graph, calculate the centrality and semantic contribution of each node, and generate metrics. Specifically, this includes: conducting a node importance assessment for the heterogeneous graph, generating metrics that can quantify the value of nodes, and providing a basis for subsequent subgraph partitioning; the assessment is carried out from two dimensions: structure and semantics. The structural dimension calculates the node centrality, using a combination of degree centrality and betweenness centrality. Degree centrality counts the number of edges associated with a node, while betweenness centrality counts the frequency of a node being located on the shortest path to other nodes. For example, core terms such as brain herniation often have high centrality because they are associated with many signs and symptoms, and high betweenness centrality because they are located on the core path of disease description.

[0045] The semantic dimension calculates the semantic contribution value of nodes. Combining the priority of neurosurgical clinical diagnosis and treatment, the basic contribution weight of terms such as disease names and treatment procedures is set higher than that of ordinary descriptive terms. Then, the basic weight is adjusted by combining the semantic enhancement vector of the node with the average similarity of other nodes. For example, if the average similarity between the pupil dilation vector and the intracranial pressure fluctuation vector is high, its semantic contribution is further improved. After the calculation is completed, the centrality value and semantic contribution value of each node are normalized and uniformly mapped to the interval of 0 to 1. The two together constitute the metric of the node. The higher the metric value, the greater the value of the node in the assessment of the disease.

[0046] Step 204: Based on the node importance distribution and the strength of inter-node dependencies reflected by the metrics, the heterogeneous graph is subgraphed to obtain multiple internally closely related sub-regions. Specifically, this includes: subgraphing the heterogeneous graph according to the metrics, with the goal of aggregating closely related terms into sub-regions to form information units corresponding to specific disease scenarios; the subgraphing uses the Louvain algorithm, which can automatically identify closely related node clusters based on the principle of maximizing modularity, adapting to the needs of this scenario; during algorithm execution, the node metrics are incorporated as weight parameters into the modularity calculation, and a threshold for inter-node dependency strength is introduced. This threshold is determined through clinical sample debugging and is set to 0.6. When the attribute values ​​of the edges between two nodes, such as semantic similarity and syntactic association strength, are lower than the threshold, they are considered weak dependencies and restricted from being classified into the same subgraph.

[0047] During the segmentation process, rules are imposed to address the unique clinical scenarios in neurosurgery. For example, terms related to brain herniation, such as pupillary changes, blood pressure fluctuations, and state of consciousness, are forcibly included in the same subgraph candidate range to ensure that terms related to disease progression are not fragmented. After segmentation, multiple sub-regions are output. The terms in each sub-region are syntactically related and semantically point to the same clinical event. For example, terms related to postoperative infection, such as fever, elevated white blood cell count, and antibiotic use, form independent sub-regions, effectively solving the problem of broken evidence chains in disease progression.

[0048] Step 205: For each sub-region, calculate the distribution density of its nodes and the connectivity of its edges, and extract them as topological features. Specifically, this includes: For each sub-region, extracting topological features that can reflect the structural characteristics of the subgraph to provide a structural basis for subsequent attention weight calculation; First, calculate the node distribution density of each sub-region. The ratio of the number of nodes in the sub-region to the theoretical maximum number of edges (i.e., the number of nodes minus one) is the distribution density. The higher the density, the denser the term association within the sub-region, which usually corresponds to core clinical events.

[0049] The connectivity of edges is then calculated and quantified using a connectivity index. The proportion of paths between any two nodes within a sub-region is statistically analyzed, and the number of connected components is recorded. If a sub-region has only one connected component and a connectivity of 0.9 or higher, it is considered to have high connectivity, indicating good information integrity. If two or more connected components exist, the shortest distance between each component is further calculated as an auxiliary topological feature. After extraction, the distribution density, connectivity, and number of connected components of each sub-region are integrated into a topological feature vector. This vector is positively correlated with the clinical importance of the sub-region, providing support for subsequent feature focusing.

[0050] Step 206: Based on the topological features and metrics, a trainable parameterized function is used for comprehensive calculation and mapping to obtain attention weight coefficients for dynamically adjusting the feature fusion ratio. Specifically, this includes: based on the metrics and topological features, the core function uses a trainable parameterized function to achieve a comprehensive mapping between the two, generating attention weight coefficients. The so-called trainable parameterized function refers to a computational model that can dynamically adjust its internal parameters based on clinical sample data to adapt to the needs of neurosurgical scenarios. Its core function is to break down the information barriers between topological features and metrics, transforming topological features reflecting the structural characteristics of sub-regions and metrics reflecting the clinical value of nodes into quantitative weights of a unified dimension, achieving the organic fusion and accurate mapping of the two types of information. At the same time, sample training ensures that the output results conform to the clinical judgment logic. This function is based on a linear weighted model constructed from 5000 labeled clinical samples. The weight parameters within the model are determined through sample iterative debugging to ensure that the computational logic is consistent with the neurosurgeon's judgment criteria for the importance of events.

[0051] During calculation, the topological feature vector of each sub-region is first concatenated with the average metric of all nodes within that region to form a comprehensive feature vector that simultaneously carries structural and value information. This comprehensive feature vector is then input into the aforementioned trainable parameterized function. The function performs a linear transformation on the comprehensive feature vector using a preset weight matrix, amplifying the feature dimensions strongly correlated with clinically critical events and suppressing interference from redundant information. After normalization, the result is mapped to a reasonable range of 0 to 1, ultimately yielding the attention weight coefficient for that sub-region. The generated attention weight coefficient is not merely a numerical identifier; its core purpose is to provide a dynamic adjustment basis for subsequent feature extraction and fusion processes. The magnitude of the value directly reflects the clinical importance of the sub-region. Sub-regions with superior topological features and higher metric indicators, such as brain herniation-related sub-regions, will receive higher weight coefficients. This means that the features of this region will be allocated more computational resources in subsequent steps, ensuring that core clinical information is not weakened and achieving targeted and efficient feature processing.

[0052] Step 207: Based on the heterogeneous graph and attention weight coefficients, input the heterogeneous graph into a graph neural network to extract structural features containing graph topological information. Specifically, this includes: extracting structural features containing topological information based on the heterogeneous graph and attention weight coefficients, using a graph convolutional neural network (GCN) as the feature extraction carrier, which has been pre-trained and optimized using neurosurgical graph structure data; before inputting the heterogeneous graph, assigning attention weight coefficients to all edges of the corresponding sub-regions, causing the network to tilt towards high-weight sub-regions during convolution calculation, i.e., using higher convolution kernel weights for edges in high-weight sub-regions, thereby strengthening the extraction of structural information from the core region.

[0053] During network computation, the features of each node are aggregated with the features of neighboring nodes by combining the adjacency matrix with the weight coefficients. After processing through two convolutional layers, a structural feature vector that reflects the overall topology and core sub-region characteristics of the heterogeneous graph is output. This vector effectively captures the association patterns between terms, such as the linkage between disease symptoms and physical signs, providing structural support for constructing a coherent chain of disease evidence.

[0054] Step 208: Using the semantic enhancement vectors corresponding to each node in the heterogeneous graph as the node feature set, the node feature set and attention weight coefficients are input into the attention network to extract key semantic features that focus on key semantics. Specifically, this includes: first, collecting the semantic enhancement vectors of all nodes in the heterogeneous graph, dividing them into node feature subsets according to sub-regions, with each subset corresponding to a clinical event unit; inputting these feature subsets and their corresponding attention weight coefficients into the attention network composed of a multilayer perceptron, and the network allocates higher computing resources to the node features of high-weight sub-regions through a weight allocation mechanism.

[0055] Specifically, term vectors in high-weight sub-regions, such as pupil dilation and increased intracranial pressure, are given greater weight coefficients in network calculations, and their semantic information is enhanced; while the semantic contribution of non-core term vectors in low-weight sub-regions is reasonably suppressed; after network processing, the output is a key semantic feature vector that focuses on key clinical semantics. This vector effectively solves the problem of semantic information fragmentation and realizes the aggregation of core semantics.

[0056] Step 209: Based on the attention weight coefficient, dynamically modulate the contribution weights of the structural features and key semantic features in the fusion process to determine the final fusion weight. Specifically, this includes: firstly, decomposing the attention weight coefficient into two dimensions: structural weight and semantic weight. The structural weight is determined by the proportion of topological features of the sub-region, and the semantic weight is determined by the proportion of semantic contribution of the sub-region. For example, sub-regions with high connectivity and concentrated core terms have higher structural and semantic weights.

[0057] Subsequently, a weight modulation rule was established: when the association pattern reflected by the structural features and the clinical event reflected by the semantic features are highly matched, such as structurally closely related terms all pointing to intracranial hemorrhage semantically, the fusion weight of the two types of features is increased to the same level; when there is a local deviation between the two, such as some syntactically related terms having weak semantic association, the features corresponding to the high-weight sub-region are tilted according to the attention coefficient; through this modulation rule, the final fusion weight of the structural features and key semantic features is finally determined, realizing the dynamic adaptation of the fusion strategy.

[0058] Step 210: Based on the final fusion weight, perform weighted fusion on the structural features and key semantic features to obtain optimized key event features. Specifically, this includes: based on the final fusion weight, performing weighted fusion on the structural features and key semantic features to generate optimized key event features. The fusion operation adopts a vector weighted summation method, multiplying the structural feature vector and the key semantic feature vector by their respective final fusion weights, and then adding the two weighted vectors dimension by dimension to obtain the fused feature vector.

[0059] After fusion, the clinical validity is verified by comparing the key events pointed to by the fused feature vector with the manually labeled results to ensure that the fused features can accurately cover core event information such as disease changes, treatment procedures, and complications. The final optimized key event features not only retain the structural relationship between terms but also aggregate core semantic information, which can directly support subsequent event classification and structured data generation. This effectively solves the shortcomings of the one-sided feature extraction of existing methods and provides high-quality data support for clinical decision-making.

[0060] In this embodiment of the invention, syntactic relationship edges are generated by parsing the syntactic structure between terms, capturing explicit grammatical associations between terms and clarifying the expression logic; simultaneously, semantic similarity is calculated to generate semantic relationship edges, mining implicit semantic associations between terms, such as linking pupil changes with intracranial pressure fluctuations; the acquisition of dual relationship edges fully covers the multi-dimensional associations between terms, avoiding information isolation; a heterogeneous graph is constructed with terms as nodes and dual relationship edges as connections to achieve the organic integration of syntactic and semantic information; the graph structure transforms linear term sequences into a structured association network, intuitively presenting the complex dependencies between terms, making the scattered term information form an organic whole; node centrality and semantic contribution values ​​are calculated from both structural and semantic dimensions to quantify the importance of nodes in the graph; the centrality value reflects the structural core of the node, and the semantic contribution value reflects the clinical value of the node. The algorithm divides the graph into subgraphs based on node importance and dependency strength, aggregating closely related terms into sub-regions. Each sub-region corresponds to a type of clinical event unit; for example, postoperative infection-related terms form an independent sub-region. The node distribution density and edge connectivity of the sub-regions are calculated as topological features to quantify the structural characteristics of the sub-regions. Distribution density reflects the density of term associations, and connectivity reflects the completeness of information in the sub-region. Topological features provide a structural basis for judging the event completeness and coreness of the sub-region, assisting in the subsequent identification of key event units. Attention weight coefficients are generated by combining topological features and metrics through parameterized functions, achieving deep fusion of the two types of information. The weight coefficients can dynamically reflect the importance of different sub-regions, enabling subsequent feature processing to focus on high-value areas and ensuring that core clinical information is not weakened in feature extraction.

[0061] By incorporating attention weight coefficients into a heterogeneous graph neural network, the network prioritizes high-weight regions when extracting structural features. The extracted structural features contain topological information and core association patterns of the graph, capturing the structural characteristics of key sub-regions and providing structural support for understanding the association logic of events. Based on semantic enhancement vectors, these are input into an attention network using attention weight coefficients, enabling the network to prioritize the semantic information of high-weight nodes. The extracted key semantic features strengthen the semantic expression of core terms, filter redundant semantic interference, and ensure that semantic features are highly relevant to clinical key events. The contribution weights of structural features and key semantic features are dynamically modulated based on attention weight coefficients, allowing the fusion strategy to adapt to the feature requirements of different events. Higher weights are assigned to features corresponding to core sub-regions, achieving differentiated allocation of feature contributions and avoiding the dilution of core information caused by simple superposition of the two types of features. Weighted fusion of the two types of features is performed based on the final fusion weights, generating optimized key event features that simultaneously carry structural associations and core semantic information. These features retain the association logic of events while highlighting core semantic content, forming a complete and focused event representation, providing high-quality feature support for the accurate identification and classification of subsequent key events.

[0062] In a preferred embodiment of the present invention, step 300 involves: based on the optimized key event features, performing a normalized mapping of event descriptions to obtain standardized events; based on the standardized events, identifying core event sentences and completing missing event arguments based on semantic similarity matching; and integrating the standardized events and the completed event arguments to generate a complete list of event arguments, including: Step 301: Decode the corresponding event text description from the optimized key event features; match the event text description with the neurosurgical standard medical terminology database to identify non-standard and colloquial expressions; convert the non-standard and colloquial expressions into corresponding standard medical terms according to the predefined mapping rules in the neurosurgical standard medical terminology database to obtain standardized events. Specifically, this includes: using key event features as input, completing event text decoding and standardization to resolve inconsistencies in extraction results; the entire process is divided into three progressively advancing stages, with previous results providing support for subsequent operations and ensuring process continuity. Effective; the feature vectors are restored to clinical statements. A Transformer decoder pre-trained with 10,000 neurosurgical nursing records is used to accurately capture the clinical semantics in the feature vectors and reverse map the 64-dimensional key event feature vectors into natural statements. The decoding introduces a bundle search strategy with a bundle width of 5 and generates 5 candidate statements. The semantic overlap is calculated by the core term matching quantity and syntactic similarity. The statement with the highest overlap is selected for output. For example, the decoded statement is: The patient's headache is relieved today compared to yesterday, and the blood pressure of 100 / 60 mmHg is 15 mmHg lower than this morning. This ensures that the text fits the clinical description and the information is complete.

[0063] Based on a standard medical terminology database for neurosurgery, non-standard expressions are identified, paving the way for standardized conversion. This database contains 3,200 standard terms and 580 common non-standard / colloquial expressions. The standard terms cover core categories such as symptoms, signs, and assessment indicators. The matching employs a dual strategy of precision and fuzzy matching: precise matching locates explicit abbreviations such as ICP and EVD; fuzzy matching uses a character similarity threshold of 0.8 to capture colloquial descriptions such as headache, dizziness, and low blood pressure. Finally, the system outputs the location information of the non-standard expressions, the text content, and the initially matched standard terms.

[0064] The standardized conversion is based on predefined mapping rules within the terminology database, with rule construction closely aligned with clinical norms: Based on authoritative sources such as the *Guidelines for Clinical Nursing in Neurosurgery*, two types of mappings are identified: one-to-one and one-to-many. One-to-one mappings are for semantically unique expressions, such as ICP corresponding to intracranial pressure and GCS corresponding to the Glasgow Coma Scale, and are fixed in dictionary form. One-to-many mappings are for expressions requiring contextual understanding, such as blood pressure elevation associated with numerical grading: systolic blood pressure of 120-139 mmHg or diastolic blood pressure of 80-89 mmHg corresponds to mild elevation, 140-159 mmHg or 90-99 mmHg corresponds to moderate elevation, and greater than or equal to 160 / 100 mmHg corresponds to severe elevation, forming conditional rules for translating expressions and numerical conditions into standard terminology. These rules are reviewed and verified by senior neurosurgical clinicians and terminologists to correct association biases in easily confused expressions such as postoperative reactions and to supplement rules for special scenarios such as blood pressure grading in elderly patients. Verified rules are stored in the terminology database, including core information such as non-standard expression types and matching conditions, ensuring traceability.

[0065] Based on the above rules, a standardized transformation is performed: the rules corresponding to non-standard expressions are retrieved, and one-to-one mapping is used for direct transformation, such as EVD to external ventricular drainage tube; one-to-many mapping extracts contextual numerical matching conditions, such as high blood pressure with 150 / 95 mmHg to moderate elevation, 130 / 85 mmHg to mild elevation; after transformation, the results are back-verified through a terminology database to ensure standardization; finally, standardized events are output to provide a unified and standardized data foundation for subsequent core event sentence recognition.

[0066] Step 302: For the standardized events, key sentence detection processing is performed. The processing specifically includes: counting the number of standardized events in each sentence as a first indicator, calculating the ratio of key arguments to the total number of words in each sentence as a second indicator, and evaluating the completeness of the semantic structure of each sentence as a third indicator; weighting and summing the first, second, and third indicators according to preset weights to obtain the decision value of each sentence; identifying and marking sentences with decision values ​​exceeding a preset threshold as core event sentences, specifically including: for standardized events, locating core event sentences, focusing on sentences carrying key medical information, and the specific operations revolve around indicator calculation, weighted summation, and threshold screening.

[0067] First, three indicators are calculated. The first indicator is the number of standardized events. After splitting the nursing record into sentences, the standardized events output in step 301 are matched sentence by sentence, and the number of standardized events contained in each sentence is counted. For example, if a sentence contains two standardized events, namely pupillary dilation and increased intracranial pressure, the first indicator is set to 2. The second indicator is the proportion of key arguments. Key arguments are predefined as core information such as timestamps (e.g., 14:30), specific values ​​(e.g., GCS score of 12), and degree descriptions (e.g., obvious). Key argument words in the sentence are extracted using part-of-speech tagging tools, and the ratio of their number to the total number of words in the sentence is calculated. For example, if a sentence contains 15 words and 4 key arguments, the second indicator is set to 4 / 15≈0.27.

[0068] The third indicator is semantic structural completeness. Syntactic analysis tools are used to analyze the subject-verb-object structure of sentences. If a sentence contains a clear subject such as "patient", a verb such as "appearance", an object such as "symptoms", and at least one key argument, it is judged to be structurally complete and is assigned a value of 1. If it lacks an object or key argument, it is assigned a value of 0.5. If it consists only of scattered words without a complete syntactic structure, it is assigned a value of 0. For example, if a patient develops left pupil dilation at 14:30 with a diameter of 3mm, the sentence is syntactically complete and contains a time argument, and the third indicator is assigned a value of 1.

[0069] Then, a weighted summation calculation is performed. The preset weights were determined through testing with 5,000 clinical samples: the first indicator has a weight of 0.4, the second indicator has a weight of 0.35, and the third indicator has a weight of 0.25. The decision value is obtained by multiplying the three indicators of each sentence by their corresponding weights and then summing them. For example, if the three indicators of a sentence are 20.271, the decision value is calculated as 2×0.4+0.27×0.35+1×0.25≈1.14. Based on clinical needs, a preset decision threshold of 0.75 is set. Sentences with decision values ​​exceeding this threshold are marked as core event sentences to ensure that the marking results focus on key information such as changes in the condition and treatment procedures.

[0070] Step 303: Calculate the semantic similarity between each core event sentence and adjacent sentences in the nursing record; extract the event arguments contained in adjacent sentences with semantic similarity higher than a preset threshold that are missing in the current core event sentence. Specifically, this includes: focusing on the complete completion of event arguments based on the core event sentence to solve the problem of information fragmentation; to improve the accuracy of semantic association judgment, a spatial vector projection algorithm is integrated into the original logic to enhance the matching degree of semantic dimensions related to the core event and achieve accurate screening of candidate supplementary sentences. Specifically, this is carried out in three steps: semantic vector optimization, accurate similarity calculation, and targeted argument extraction. Each step closely relies on the technical achievements of the preceding steps.

[0071] First, the basic vector construction is completed, and then the core semantic association is strengthened through the projection algorithm. Sentence vector generation is performed first, following the mature word vector processing logic in step 101: the core event sentence and the two adjacent sentences before and after it (forming a five-sentence related text range) are split into words one by one. For each word in each sentence, the context window setting in step 101 is called (covering 3 words before and after the word) to extract the context word vector group. The initial sentence vector corresponding to each sentence is obtained by adding them dimension by dimension and taking the average. All vectors are uniformly set to 256 dimensions to ensure dimensional consistency.

[0072] Furthermore, a spatial vector projection algorithm is introduced to optimize vector representation. The initial sentence vector of the core event sentence is used as the reference vector of the target projection space. First, the reference vector is normalized by dividing the value of each dimension of the vector by the magnitude of the vector to obtain a unit reference vector with a magnitude of 1. Then, the initial sentence vectors of each adjacent sentence are projected into the space constructed by the unit reference vector. The specific operation of the projection is as follows: calculate the dot product of the adjacent sentence vector and the unit reference vector to obtain the projection coefficient; multiply the coefficient by the unit reference vector dimension by dimension to generate the projection vector of the adjacent sentence. Through this operation, semantic dimensions that are not related to the core event in adjacent sentences can be filtered out, such as dietary descriptions interspersed in nursing records, while strengthening the feature expression of core semantic dimensions such as blood pressure changes and pupil status, so that the projection vector is more focused on information related to the core event.

[0073] Using the optimized projection vector as the calculation object, the similarity judgment is improved. The calculation follows the core logic of cosine similarity, but the calculation object is replaced from the initial sentence vector to the projection vector. The specific process is as follows: Taking the unit reference vector of the core event sentence as a reference, the projection vectors of each adjacent sentence are paired and calculated one by one; First, the dot product of the unit reference vector and the projection vector is calculated, and then the magnitude of the two vectors is calculated separately. That is, the magnitude of the unit reference vector is already 1, so the calculation can be simplified directly. The dot product result is divided by the product of the magnitudes of the two vectors to obtain the similarity value between 0 and 1.

[0074] Based on the clinical characteristics of neurosurgical event associations, a preset similarity threshold of 0.65 was set. This threshold was determined through pairing samples of 1000 sets of core neurosurgical events and adjacent statements to ensure that it covers both strongly related statements such as blood pressure changes and intracranial pressure fluctuations, and excludes weakly related statements that only have literal overlap. The calculated similarity values ​​are compared with the threshold, and adjacent statements with values ​​exceeding 0.65 are selected as candidate supplementary statements and sorted from high to low similarity values ​​to provide a priority basis for subsequent argument extraction.

[0075] By combining projection vector features and argument templates, the precise extraction and attribution binding of missing arguments can be achieved. First, the missing arguments of the core event sentence are identified, and a neurosurgical event argument template is used for comparison. This template is constructed according to the event type. For example, the essential arguments for events related to changes in condition include subject, time, indicator name, specific value, and trend of change, while the arguments for events related to treatment operations include operation name, implementation time, and operation result. Taking the core event sentence "significant decrease in blood pressure" as an example, the missing time, specific value, and magnitude of decrease can be directly located through template comparison.

[0076] Subsequently, based on the ranking of candidate supplementary sentences, missing arguments are extracted first from the sentences with the highest similarity. During extraction, the dimensional features of the projection vector are used to assist in the judgment. The core event sentence vector has a higher weight in the semantic dimension corresponding to the blood pressure index. The extracted arguments need to form a strong correlation with the projection vector in this dimension. For example, in the adjacent sentence "14:30 blood pressure measured 90 / 55 mmHg, a decrease of 20 mmHg from before", 14:30 corresponds to the time dimension, 90 / 55 mmHg corresponds to the specific numerical dimension, and 20 mmHg corresponds to the magnitude of the decrease dimension. Moreover, the weights of these dimensions in the projection vector are all higher than 0.8, confirming them as valid supplementary arguments.

[0077] To ensure clear argument attribution, a three-dimensional association identifier is established, consisting of the core event sentence ID, argument type, and source sentence ID. For example, if the core event sentence ID is E001, the corresponding identifiers for the time argument are E001, time, and S003, where S003 is the source sentence ID. After extraction, a semantic consistency check is performed: the similarity between the vector corresponding to the completed argument and the projection vector of the core event sentence is calculated to ensure that the value exceeds 0.7, thus avoiding the extraction of irrelevant arguments. Finally, a set of completed arguments with association identifiers is output, realizing the binding of the core event sentence with the missing arguments.

[0078] Step 304 involves integrating all standardized events, core event sentences, and missing event arguments to obtain a complete list of event arguments. This includes: firstly, performing data association processing, using the core event sentence as the aggregation center, associating and binding the standardized events corresponding to the core sentence in step 301, and simultaneously associating the arguments completed in step 303 with the corresponding event items according to argument type, such as time value. For example, associating 14:30 and 90 / 55 mmHg with the standardized event of blood pressure decrease. During the association process, timestamp calibration is used to ensure that events and arguments of the same time dimension are aggregated into a group, such as integrating pupil changes, blood pressure changes, and corresponding indicators at 14:30 into a set of events at the same time node.

[0079] Then, structured integration is performed, filling in the content according to a preset event argument list template. This template includes fields such as event ID, event type, core event sentence, standard terminology, argument item time, argument value, argument change trend, and source statement identifier. After filling, redundancy is removed. If the events and arguments of different core sentences are completely duplicated, only the earliest record is retained. If there are differences in arguments, they are merged and supplemented to form a complete record. The final output complete event argument list is presented in the form of a structured table. Each record corresponds to a core clinical event. The argument information is complete and deeply bound to the timestamp, which can directly support clinical decision-making.

[0080] In this embodiment of the invention, standardized processing is used to unify and standardize event descriptions. After decoding key event features to obtain textual descriptions, non-standard and colloquial expressions are efficiently identified through precise matching with a standard neurosurgical terminology database. Conversion is performed according to predefined mapping rules, unifying the personalized expressions of different nursing staff into standard medical terminology and eliminating data ambiguity caused by differences in expression. Core event sentences are located through multi-indicator weighted evaluation. An evaluation system is constructed based on the number of standardized events, the proportion of key arguments, and the completeness of semantic structure. Decision values ​​are calculated using preset weights to ensure that the identification logic aligns with clinical standards for judging key information. Sentences that meet the decision value criteria are marked as core event sentences, enabling rapid focus on statements carrying core information such as changes in patient condition and treatment procedures, facilitating subsequent data processing. Resources are allocated to high-value texts to reduce redundant information interference and improve overall processing efficiency. Semantic similarity is used as a link to achieve complete event arguments. By calculating the semantic relationship between the core event sentence and adjacent sentences, sentences with supplementary value are accurately selected. Missing arguments in the core event sentence are extracted from qualified sentences, and key information such as timestamps and indicator values ​​scattered in the context are aggregated to the corresponding event, avoiding fragmented event information caused by argument fragmentation. Clear and usable event data results are formed through structured integration. Standardized events, core event sentences, and supplementary arguments are integrated according to event association logic to generate a complete list of event arguments. The list presents the core information and related arguments of the event in a structured form, clarifying the correspondence between various elements and solving the problems of scattered event information and ambiguous relationships.

[0081] In a preferred embodiment of the present invention, step 400 involves performing fine-grained event classification based on the complete list of event arguments to obtain event type labels; identifying and extracting structured association information for each event based on the event type labels; and integrating the event type labels with the structured association information to obtain structured key event data, including: Step 401: Based on the complete list of event arguments, an event classifier is used to analyze and classify the argument composition and semantic features of each event, assigning a fine-grained event type label to each event. Specifically, this includes: implementing fine-grained event segmentation through a customized event classifier to provide clear guidance for subsequent information extraction. This is achieved logically through classifier construction, feature input, classification calculation, and label output. First, the event classifier is constructed and pre-trained. This classifier adopts a two-tier structure of a feature extraction layer and a classification decision layer. The feature extraction layer uses the word vectors fine-tuned from the neurosurgical text in step 101. The model's classification decision layer is constructed using a fully connected network. The training process is based on 8,000 labeled neurosurgical nursing events, covering four core scenarios: changes in condition, treatment procedures, complications, and vital sign monitoring. These events are further subdivided into 28 fine-grained event types, such as blood pressure fluctuations, pupillary changes, increased intracranial pressure, decompression craniectomy, and postoperative infection, forming paired training data with event argument sets and fine-grained type labels. During training, the classification accuracy is optimized, and a mini-batch gradient descent method is used for iterative training for 60 rounds. Training stops when the classification accuracy on the validation set stabilizes above 94%, and the classifier parameters are solidified.

[0082] During classification calculation, the core information of a single event is first extracted from the complete list of event arguments to form an input feature set of argument composition features and semantic features. The argument composition features focus on the type combination of arguments, such as combinations containing blood pressure values, change trends, and time arguments, which directly point to blood pressure fluctuation events. The semantic features are obtained by taking the mean value of the semantic enhancement vectors corresponding to the event arguments (using the results of step 109) to strengthen the semantic association of neurosurgical professional terms. The input feature set is input into a pre-trained classifier, the feature extraction layer completes the deep encoding of the features, and the classification decision layer calculates the probability value of the event belonging to various fine-grained labels through the Softmax function, and selects the label with the highest probability (and greater than 0.85) as the type label of the event. For example, an event containing arguments such as pupil diameter 3mm, dilation compared to before, and 14:30 is assigned the fine-grained labels of pupil change and dilation, ensuring that the classification results accurately correspond to specific clinical scenarios.

[0083] Before output, reliability is improved through double verification: First, the similarity of the argument composition of similar events is calculated to ensure that the core arguments of similar events are consistent; second, for high-risk events such as brain herniation and intracranial hemorrhage, clinical experts sample and review the classification results, and output the associated data of event ID and fine-grained type label after the verification is passed.

[0084] Step 402: For each event type label and its corresponding event arguments, the information extraction unit identifies and extracts quantitative indicators, occurrence time, and symptom description information associated with the event. Specifically, this includes: relying on fine-grained type labels, the information extraction unit captures the structured association information of the event in a targeted manner. The structured association information includes quantitative indicators, occurrence time, and symptom description information, addressing the problem of loose binding between information and events. To further improve the accuracy of information extraction and eliminate scattered interference data, the DBSCAN spatial clustering algorithm is integrated into the original logic to achieve aggregation and optimization of extracted information. The overall implementation process revolves around the extraction rules. The process of constructing, targeted identification, cluster optimization, and information verification is carried out systematically. Using fine-grained type labels as the core anchor, customized extraction logic is developed according to event type, forming a structured correspondence library of labels, extracted elements, and identification rules, providing a clear basis for targeted identification. Specifically, for blood pressure fluctuation labels, the extracted elements are clearly defined as systolic blood pressure value, diastolic blood pressure value, amplitude of change, and time of occurrence; for intracranial pressure elevation labels, the extracted elements are intracranial pressure value, measurement method such as ICP monitoring, time of occurrence, and accompanying symptoms; for treatment operation labels such as decompressive craniectomy, the extracted elements are operation name, implementation time, operation result, and postoperative monitoring indicators.

[0085] The recognition rules are designed to closely match the characteristics of neurosurgical text: time information is recognized by matching formats such as HH:MM, X hours post-surgery, and X hour today; quantitative indicators are recognized by matching combinations of numerical values ​​and units, such as 25cmH2O and 12-minute GCS; symptom description information is recognized by fuzzy matching with a neurosurgical symptom terminology database, which covers more than 500 standard symptoms, with a matching similarity threshold set at 0.75 to ensure the accuracy of symptom association.

[0086] Targeted identification, guided by extraction rules, first retrieves matching extraction elements and identification rules from the corresponding relational database based on the fine-grained labels of the event. Then, it focuses on the argument set of the event and the associated original nursing statements to capture information. For example, for events tagged with blood pressure fluctuation and decline, it matches numerical information such as 90 / 55 mmHg, information on the magnitude of change such as a 20 mmHg decrease from before, time information such as 15:20, and symptom information such as dizziness and fatigue from the arguments and original sentences according to the rules. To avoid the one-sidedness of a single extraction, three independent identifications are performed on the same extraction element: the first is based on the event argument set, the second is based on the core clauses of the original nursing statement, and the third is based on the contextual content of the statement. After deduplication of the three identification results, a preliminary extraction set containing multiple sets of candidate information is formed. For example, the blood pressure numerical candidate set may contain multiple similar data such as 90 / 55 mmHg, 89 / 54 mmHg, and 91 / 56 mmHg, providing a data foundation for subsequent clustering optimization.

[0087] The DBSCAN spatial clustering algorithm is used to aggregate the quantitative indicator information in the initially extracted set, remove isolated outliers, and improve the accuracy of the information. The specific calculation process is as follows: First, feature vectors are constructed, and the core parameters of the quantitative indicators are transformed into two-dimensional feature vectors: the indicator value is the first dimension, and the frequency of the value in the nursing records is the second dimension. For example, if the blood pressure value of 90 / 55 mmHg appears twice, the corresponding vector is a quantified value of 0.92 and a frequency of 2 for 90 / 55 mmHg. The algorithm parameters are set in combination with the clinical characteristics of the neurosurgical quantitative indicators: the neighborhood radius ε is set to 0.15. This value is determined based on the reasonable clinical fluctuation range of similar indicators. For example, the reasonable fluctuation of blood pressure values ​​usually does not exceed 5 mmHg, and the corresponding difference threshold after quantification is 0.15. The minimum number of samples (min samples) is set to 3, that is, to form an effective cluster, it must contain at least 3 feature vectors to ensure the stability of the clustering results.

[0088] During clustering calculation, all feature vectors are traversed first. Taking each vector as the core, the number of other vectors in its ε-neighborhood is retrieved. If the number reaches the minimum samples, it is marked as a core point, and then expanded to form a cluster. If the number of vectors in the neighborhood is insufficient and not covered by the neighborhood of other core points, it is marked as an isolated point. After clustering, the effective clusters containing core points are retained, and isolated point data are removed. For each effective cluster, the average value of the index within the cluster is calculated as the final extraction result, which serves as the core quantitative indicator for the event. Symptom descriptions and time information are filtered based on the correlation between the indicators after clustering, retaining the time and symptoms that occur synchronously with the indicators of the effective clusters to ensure the correlation between information.

[0089] Information verification is conducted based on clustering optimization, forming a dual verification logic: First, clinical rationality verification confirms that the clustered quantitative indicators are within the normal clinical range. For example, blood pressure values ​​should be between 40 / 20 mmHg and 250 / 150 mmHg, and intracranial pressure values ​​should be within the normal range of 5 to 15 cmH2O. Values ​​outside the range are marked as abnormal and associated with the original statements. Second, association consistency verification checks whether the time information matches the timeline of the nursing records, and whether the symptom descriptions and quantitative indicators conform to the clinical pathological association. For example, a drop in blood pressure is usually associated with dizziness and weakness rather than restlessness. If a contradiction occurs, a manual review prompt is triggered. After the verification is passed, a five-tuple association data of event ID, type label, quantitative indicator, occurrence time, and symptom description is output. The quantitative indicator is accompanied by a clustering optimization label, which clarifies that the data has been processed by the DBSCAN algorithm, ensuring that each extracted piece of information can be clearly assigned to the corresponding event, and that the data accuracy meets the requirements of clinical application.

[0090] Step 403: Align and combine the event type label assigned to each event with the corresponding quantitative indicators, occurrence time, and symptom description information to obtain structured key event data. Specifically, this includes: information alignment using the event ID as a unique association identifier, matching the fine-grained type labels output in step 401 with the five-tuple data extracted in step 402 to ensure that the labels and information under the same event ID correspond completely; during the alignment process, erroneous associations are eliminated through consistency checks: if the type label of an event ID is elevated intracranial pressure, while the extracted quantitative indicator is blood pressure value, a check alarm is triggered, and the previous steps are returned to re-extract; if there are multiple records of occurrence time, such as operation start time and end time, they are marked as occurrence time start and occurrence time end, respectively, to avoid information confusion.

[0091] The structured data is presented in a clinically adapted, field-based format. The pre-defined structured data template includes six core fields: event ID, fine-grained event type, core quantitative indicators (including numerical values ​​and units), precise occurrence time, symptom description, and associated original statement ID. The aligned information is then filled in field by field. For example, the data for a given event might be: Event ID - E401, Event Type: Pupil Change - Dilatation, Quantitative Indicator: Pupil Diameter 4mm (Left Side), Occurrence Time: 2025-12-02 14:30, Symptom Description: Sluggish Light Reflex, Original Statement ID - S108. For fields without corresponding information, such as those where the event is monitored but no obvious symptoms are observed, they are marked as "no clear record" and associated with the monitoring data to ensure field completeness.

[0092] The final structured critical event data is generated through two layers of verification: the first layer is format verification, ensuring that the time format is consistent with year, month, day, hour, and minute, and that quantitative indicators include units; the second layer is clinical rationality verification, which is reviewed by neurosurgeons on a sample basis (10%) to confirm that the association between event types and information conforms to clinical logic. For example, the brain herniation warning label needs to be associated with multiple information such as pupil changes, increased intracranial pressure, and blood pressure fluctuations. The structured critical event data that passes the verification can be directly imported into the clinical decision support system, supporting rapid query and statistics by event type, time range, and other dimensions, effectively meeting the needs of real-time structured decision support.

[0093] In this embodiment of the invention, fine-grained classification achieves precise definition of event types. The event classifier focuses on analyzing the dual dimensions of event argument composition and semantic features, avoiding classification bias caused by relying solely on a single feature. Each event is assigned a unique fine-grained type label; for example, events related to changes in condition are further subdivided into blood pressure fluctuations, pupillary changes, and alterations in consciousness, making the attribute boundaries of different events clearly distinguishable. Targeted extraction achieves precise capture of event-related information. Using event type labels as the core filtering criterion, the information extraction unit specifically identifies and extracts quantitative indicators such as blood pressure values, GCS scores, occurrence time (e.g., 14:30, 2 hours post-surgery), and symptom descriptions (e.g., severe headache, dilated pupils), among other related information. The extraction process closely adheres to the event type attributes; for example, for blood pressure fluctuation events, the focus is on extracting blood pressure values ​​and changes. The data analysis focuses on extracting key data such as body temperature and white blood cell count for infection events, effectively filtering out redundant information irrelevant to the event and ensuring a high degree of matching between the extracted information and the event, thus enhancing the core value of the data. By aligning and combining data to construct structured event data, event type labels and quantitative indicators, occurrence time, and symptom descriptions are precisely aligned according to event affiliation, ensuring that each information element is clearly associated with its corresponding event. The resulting structured key event data presents a clear picture of the event with well-defined element relationships. For example, a blood pressure fluctuation label is linked to information such as a 20 mmHg decrease in blood pressure at 14:30 (90 / 55 mmHg) compared to the previous day, and worsening headache, resolving the problem of ambiguous associations between events and information elements. This structured format allows the data to be directly adapted to clinical data query, statistics, and decision support systems, improving the data's practical value and application efficiency.

[0094] In a preferred embodiment of the present invention, step 500, which generates and outputs a key event report based on the structured key event data, includes: Step 501: Based on the structured critical event data, the data is categorized and sorted according to the preset reporting dimensions and format requirements, and organized into an ordered set of report data. Specifically, this includes: First, completing the construction of the preset reporting dimensions and format requirements. The setting of the reporting dimensions is closely related to the clinical work focus of the neurosurgical ICU, referring to the "Guidelines for Neurosurgical Critical Care Nursing Practice" and the clinical reporting standards of tertiary hospitals, and determining three core dimensions: First, the time dimension, with the time of event occurrence as the core axis, supporting the presentation of the disease's evolution; Second, the event attribute dimension, divided into four categories according to event type: vital sign fluctuations, symptom changes, treatment procedures, and complication warnings. Each category is further associated with fine-grained sub-types, such as vital sign fluctuations including blood pressure fluctuations, intracranial pressure changes, etc.; Third, the urgency dimension, based on the clinical risk level corresponding to the event, divided into three levels: high-risk warning (e.g., brain herniation-related signals), medium-risk attention (e.g., postoperative infection signs), and routine monitoring (e.g., basal body temperature changes). The format requirements clearly define the data expression standards, such as using the year, month, day, hour, and minute format for time, indicating the unit for quantitative indicators and retaining one decimal place, and using standard medical terminology for symptom descriptions.

[0095] Based on this, data classification and sorting operations are carried out. Classification is based on the fine-grained event type labels and the quantitative index range of occurrence time in the structured key event data. Data is matched one by one to the corresponding dimension. For example, data labeled "pupil change - dilation," occurrence time 2025-12-02 14:30, and pupil diameter 4mm are classified into the time dimension of 14:30, the event attribute dimension of symptom change, and the urgency dimension of high-risk warning level. Sorting is performed hierarchically according to primary and secondary sorting criteria. The primary criterion is urgency, arranged in the usual order of high-risk, medium-risk, and general. Under the same urgency level, occurrence time is used as the secondary criterion, arranged in ascending order of time, ensuring that high-risk events are presented first and the disease progression is clearly presented chronologically. The resulting ordered report data set has a hierarchical structure of urgency, occurrence time, and event type, with complete event data corresponding to each level, providing logically clear data support for subsequent report filling.

[0096] Step 502: Based on the aforementioned report data set, populate the predefined critical event report template to obtain tabular event data. Specifically, the predefined critical event report template is constructed with clinical practicality as its core objective. It adopts a tabular design, setting a structure of fixed column fields and dynamic row data. The fixed column fields cover the core information required for clinical decision-making, specifically including seven columns: event number, occurrence time, fine-grained event type, core quantitative indicators and units, description of accompanying symptoms, related treatment operations, and urgency level indicator. The urgency level indicator uses color coding to distinguish between high-risk warnings (red), medium-risk warnings (yellow), and routine monitoring (blue) to improve visual recognition. The template also reserves a remarks column for supplementing explanations of special clinical scenarios, such as blood pressure fluctuations caused by low baseline blood pressure in elderly patients requiring special annotation. This template has been reviewed and confirmed by three chief neurosurgeons and two nursing experts to ensure that the field settings are highly consistent with clinical review needs.

[0097] The data population operation takes an ordered report dataset as input and is performed in a field-matching, data extraction, and precise population manner. Specifically, it extracts the corresponding information for a single event from the ordered dataset, uses the event ID as the event number, and extracts information such as occurrence time and fine-grained event type to populate the corresponding columns. Core quantitative indicators must be filled in with complete values ​​and units, such as intracranial pressure 25cmH2O. Accompanying symptom descriptions must integrate relevant content from the structured data and be concise, such as dizziness with sluggish pupillary light reflex. Field integrity checks are introduced during the population process. If a core field of an event, such as the occurrence time, is missing, it is marked as to be supplemented in the corresponding position and associated with the original nursing record ID, which facilitates subsequent manual review and improvement. The generated tabular event data after population is completed corresponds to a single event by row and presents the core attributes of the event by column. The information is intuitive and the boundaries are clear, which fully conforms to the reading habits of clinical reports.

[0098] Step 503 involves converting the tabulated event data into a data format conforming to the clinical information system interface specification and outputting it as a critical event report. Specifically, this includes: firstly, analyzing the clinical information system interface specification, collecting interface documents from the hospital's existing clinical information systems, primarily the HIS system and ICU monitoring system, and determining the data format requirements: data transmission uses XML format, field encoding follows the GB2312 standard, time fields use the ISO8601 format, numerical fields use floating-point data types, and text field length is controlled within 200 characters; simultaneously, the interface communication protocol is determined to be HTTP, and data transmission must include authentication information to ensure security; for specific neurosurgical needs, additionally confirming the system's field mapping rules for professional indicators such as GCS scores and ICP monitoring, such as GCS scores needing to be split into three sub-fields for eye-opening response, verbal response, and motor response for separate transmission.

[0099] The format conversion operation is based on tabular event data and adapts it according to the parsed specifications. Specifically, the column fields of the table are mapped one-to-one with the system interface fields. For example, the event number is mapped to event id, the occurrence time is converted to ISO8601 format time stamp, and the core quantitative indicators are split into three subfields: indicator name, value, and unit. During the conversion process, the data types are uniformly adjusted, converting text-type values ​​to floating-point types to ensure compliance with system data type requirements. After the conversion is completed, XML format report data is generated, and identity authentication header information is added for system access.

[0100] The final output is validated and published. Validation is two-layered: first, format validation, which uses a system interface simulation tool to test whether the report data can be successfully integrated into the clinical information system, ensuring no format errors; second, content validation, which involves sampling and verifying the consistency of information between the XML report and the tabular data, ensuring no data loss or mismatch during the conversion process. After successful validation, dual output formats are provided: first, an XML format file that can be directly integrated into the system is automatically uploaded to the designated clinical information system via the interface; second, a PDF format report that can be printed and archived, retaining the tabular presentation and color coding. The output critical event report meets the real-time access requirements of the clinical information system while also conforming to traditional paper archiving standards, fully supporting clinical diagnosis and treatment and case management.

[0101] In this embodiment of the invention, data is organized and focused through targeted organization. Pre-defined report dimensions closely align with clinical diagnostic and treatment needs, such as categorization by time progression, urgency of events, and affected systemic parts, ensuring the classification direction matches doctors' review habits. Data is sorted based on format requirements, such as ascending order by event occurrence time, making the disease progression clearly discernible. The ordered report data set formed through classification and sorting eliminates the scattered distribution of structured data, allowing different types of key event data to be categorized and prioritized, providing a logically clear and targeted data foundation for subsequent report generation, improving report construction efficiency. Standardization and intuitiveness of data presentation are achieved through template filling; predefined key event report templates include event types, Fixed fields such as occurrence time and quantitative indicators are highly consistent with the standardized expression of clinical reports; the tabular event data generated after filling the ordered data set into the template ensures that the core information elements of each event correspond one-to-one and the boundaries are clear, presenting the time, value and accompanying symptoms of pupil change events in a single table; the compatibility and immediacy of data application are achieved through format conversion, and the format is adapted according to the clinical information system interface specifications, so that critical event reports can be directly connected to the hospital's existing HIS, LIS and other systems without the need for additional secondary adjustments to the data format; the output critical event reports can be quickly integrated into the clinical information process, supporting doctors to directly retrieve and view them in the diagnosis and treatment system, and can also be directly used as the basic data for disease analysis and case summary.

[0102] like Figure 2 As shown, embodiments of the present invention also provide a key event extraction system for neurosurgical nursing records based on natural language processing, comprising: The semantic enhancement module is used to identify candidate professional terms based on word vector sequences using a terminology database in the neurosurgical field; based on the candidate professional terms, its core semantic vector is extracted and associated with the corresponding entity type embedding vector and event type embedding vector; the core semantic vector, entity type embedding vector, and event type embedding vector are subjected to multimodal fusion and nonlinear transformation to obtain a semantically enhanced term sequence. The optimization module is used to construct a heterogeneous graph that integrates syntactic and semantic information based on semantically enhanced term sequences; to evaluate the importance and divide the structure of the heterogeneous graph to obtain topological features and metrics; to obtain attention weight coefficients based on topological features and metrics; to extract structural features and key semantic features based on the heterogeneous graph and attention weight coefficients; and to dynamically optimize the fusion weights of structural features and key semantic features and perform weighted fusion to obtain optimized key event features. The normalization and completion module is used to perform normalization mapping based on optimized key event features to obtain normalized events; identify the core event sentences in the normalized events and complete missing arguments based on semantic similarity matching; and integrate the normalized events and the completed arguments to obtain a complete list of event arguments. The classification and extraction module is used to perform fine-grained event classification based on the complete list of event arguments to obtain event type labels; based on the event type labels, identify and extract the structured association information of each event; and integrate the event type labels and structured association information to obtain structured key event data. The output module is used to generate and output a key event report based on the structured key event data. The above description is a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for extracting key events from neurosurgical nursing records based on natural language processing, characterized in that, The method includes: Step 100: Based on the word vector sequence, candidate professional terms are identified using a terminology database in the field of neurosurgery; based on the candidate professional terms, their core semantic vectors are extracted and associated with the corresponding entity type embedding vectors and event type embedding vectors; the core semantic vectors, entity type embedding vectors, and event type embedding vectors are subjected to multimodal fusion and nonlinear transformation to obtain a semantically enhanced term sequence. Step 200: Based on the semantically enhanced term sequence, construct a heterogeneous graph that integrates syntactic and semantic information; evaluate the importance and divide the structure of the heterogeneous graph to obtain topological features and metrics; obtain attention weight coefficients based on topological features and metrics; extract structural features and key semantic features based on the heterogeneous graph and attention weight coefficients respectively; dynamically optimize the fusion weights of structural features and key semantic features, and perform weighted fusion to obtain optimized key event features; Step 300: Based on the optimized key event features, perform normalization mapping to obtain standardized events; identify the core event sentences in the standardized events and complete the missing arguments based on semantic similarity matching; integrate the standardized events and the completed arguments to obtain a complete list of event arguments; Step 400: Based on the complete list of event arguments, perform fine-grained event classification to obtain event type labels; based on the event type labels, identify and extract structured association information for each event; integrate the event type labels and structured association information to obtain structured key event data; Step 500: Based on the structured key event data, generate and output a key event report.

2. The method for extracting key events from neurosurgical nursing records based on natural language processing according to claim 1, characterized in that, Before step 100, the following are also included: Preprocess the neurosurgical nursing record text to generate standardized text; Contextual semantic encoding is performed on standardized text to obtain a sequence of word vectors containing dynamic semantics within the context.

3. The method for extracting key events from neurosurgical nursing records based on natural language processing according to claim 2, characterized in that, Step 100 includes: Traverse the word vector sequence and extract the context word vector of each word as the semantics of each word; The semantics of each word are compared one by one with the semantic vectors of predefined standard terms, abbreviations and synonyms in the neurosurgical terminology database. By calculating the semantic similarity between the vectors, a set of similarity matching values ​​is generated for each word. From the set of similarity matching values, filter out all words corresponding to matching values ​​that exceed a preset similarity threshold, and mark these words as candidate professional terms; From the word vector sequence to which each candidate term belongs, extract the context word vector corresponding to the candidate term as the core semantic vector; at the same time, according to the predefined mapping relationship in the neurosurgical terminology database, query and obtain the entity type embedding vector and event type embedding vector corresponding to the candidate term.

4. The method for extracting key events from neurosurgical nursing records based on natural language processing according to claim 3, characterized in that, Step 100 also includes: For each candidate term, the core semantic vector, entity type embedding vector, and event type embedding vector are concatenated to form a multimodal fusion vector; Each multimodal fusion vector is input into a fully connected neural network layer to perform a linear transformation, resulting in the corresponding linear features; The linear features are input into a nonlinear activation function for processing to obtain the corresponding activation features; Based on the activation features, each multimodal fusion vector is converted into a corresponding semantic enhancement vector; Arrange the semantic enhancement vectors of all candidate terms in the original order of their appearance in the text to form a semantically enhanced term sequence.

5. The method for extracting key events from neurosurgical nursing records based on natural language processing according to claim 4, characterized in that, Step 200 includes: Based on the semantically enhanced term sequence, the syntactic structure between terms in the sequence is parsed to obtain the syntactic relation edges between terms; at the same time, the semantic similarity between the term vectors in the sequence is calculated, and the semantic relation edges between terms are generated according to a preset threshold. Using each term in the semantically enhanced term sequence as a graph node, and using the syntactic relation edges and semantic relation edges as edges connecting these nodes, a heterogeneous graph that integrates syntactic and semantic information is constructed. The importance of all nodes in the heterogeneous graph is evaluated, the centrality value and semantic contribution value of each node are calculated, and a metric is generated. Based on the node importance distribution and the strength of inter-node dependencies reflected by the aforementioned metrics, the heterogeneous graph is divided into subgraphs to obtain multiple internally closely related sub-regions. For each sub-region, the distribution density of its nodes and the connectivity of its edges are calculated and extracted as topological features; Based on the aforementioned topological features and metrics, a trainable parameterized function is used for comprehensive calculation and mapping to obtain attention weight coefficients for dynamically adjusting the feature fusion ratio.

6. The method for extracting key events from neurosurgical nursing records based on natural language processing according to claim 5, characterized in that, Step 200 also includes: Based on the heterogeneous graph and attention weight coefficients, the heterogeneous graph is input into a graph neural network to extract structural features containing graph topological information; Using the semantic enhancement vectors corresponding to each node in the heterogeneous graph as the node feature set, the node feature set and the attention weight coefficient are input into the attention network to extract key semantic features that focus on key semantics; Based on the attention weight coefficient, the contribution weights of the structural features and key semantic features in the fusion process are dynamically modulated to determine the final fusion weight. Based on the final fusion weight, the structural features and key semantic features are weighted and fused to obtain optimized key event features.

7. The method for extracting key events from neurosurgical nursing records based on natural language processing according to claim 6, characterized in that, Step 300 includes: From the optimized key event features, the corresponding event text descriptions are decoded; the event text descriptions are matched with the neurosurgical standard medical terminology database to identify non-standard and colloquial expressions; the non-standard and colloquial expressions are converted into corresponding standard medical terms according to the predefined mapping rules in the neurosurgical standard medical terminology database to obtain standardized events. For the standardized events, key sentence detection processing is performed; the processing specifically involves: counting the number of standardized events in each sentence as a first indicator, calculating the ratio of key arguments to the total number of words in each sentence as a second indicator, and evaluating the completeness of the semantic structure of each sentence as a third indicator; weighting and summing the first, second, and third indicators according to preset weights to obtain the decision value of each sentence; identifying and marking sentences with decision values ​​exceeding a preset threshold as core event sentences; Calculate the semantic similarity between each core event sentence and adjacent sentences in the nursing record; extract the event arguments contained in adjacent sentences with semantic similarity higher than a preset threshold that are missing in the current core event sentence. By integrating all the standardized events, core event sentences, and missing event arguments, a complete list of event arguments is obtained.

8. The method for extracting key events from neurosurgical nursing records based on natural language processing according to claim 7, characterized in that, Step 400 includes: Based on the complete list of event arguments, an event classifier is used to analyze and classify the argument composition and semantic features of each event, and assign a fine-grained event type label to each event. For each event type label and its corresponding event arguments, the information extraction unit identifies and extracts the quantitative indicators, occurrence time and symptom description information associated with the event. The event type label assigned to each event is aligned and combined with the corresponding quantitative indicators, occurrence time, and symptom description information to obtain structured key event data.

9. The method for extracting key events from neurosurgical nursing records based on natural language processing according to claim 8, characterized in that, Step 500 includes: Based on the structured key event data, the data is classified and sorted according to the preset report dimensions and format requirements, and organized into an ordered report data set; Based on the aforementioned report data set, a predefined key event report template is populated to obtain tabular event data; The tabular event data is converted into a data format that conforms to the clinical information system interface specification and output as a critical event report.

10. A system for extracting key events from neurosurgical nursing records based on natural language processing, the system implementing the method as described in any one of claims 1 to 9, characterized in that, include: A semantic enhancement module is used to identify candidate technical terms based on word vector sequences and a terminology database in the field of neurosurgery. Based on candidate terms, their core semantic vectors are extracted and associated with their corresponding entity type embedding vectors and event type embedding vectors. The core semantic vectors, entity type embedding vectors, and event type embedding vectors are then subjected to multimodal fusion and nonlinear transformation to obtain a semantically enhanced term sequence. The optimization module is used to construct a heterogeneous graph that integrates syntactic and semantic information based on semantically enhanced term sequences; Importance assessment and structural partitioning of heterogeneous graphs are performed to obtain topological features and metrics; Based on topological features and metrics, attention weight coefficients are obtained; Based on heterogeneous graphs and attention weight coefficients, structural features and key semantic features are extracted respectively. The fusion weights of structural features and key semantic features are dynamically optimized and then weighted to obtain optimized key event features. The standardization and completion module is used to perform normalization mapping based on the optimized key event features to obtain standardized events; Identify the core event sentences in standardized events and complete the missing arguments based on semantic similarity matching; integrate the standardized events and the completed arguments to obtain a complete list of event arguments; The classification and extraction module is used to perform fine-grained event classification based on the complete list of event arguments to obtain event type labels; Based on event type labels, the structured correlation information of each event is identified and extracted; the event type labels and structured correlation information are integrated to obtain structured key event data. The output module is used to generate and output a key event report based on the structured key event data.

Images