A tube feeding care complication search recommendation processing method and system

Abstract

The application provides a tube feeding care complication retrieval recommendation processing method and system, including the following steps: collecting multi-modal nursing data including structured medical record data and unstructured text data, and storing the data in a data buffer pool; forming a medical feature set; taking each medical feature point in the medical feature set as a node to form a knowledge graph containing community semantic labels; expanding a refined retrieval area based on the matching degree of the user query intention and the community theme; sorting the nodes in the refined retrieval area, and outputting a retrieval recommendation result. By proposing a local refined retrieval strategy, the system can perform deep retrieval and correlation analysis on only the key complication information according to the query intention and care needs of the care personnel, and perform simplified processing on other information, thereby significantly reducing the system calculation amount, improving the retrieval efficiency, and ensuring that the key complication information can be quickly presented and updated in real time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intensive care management, in particular to a tube feeding care complication retrieval recommendation processing method and system. BACKGROUND

[0002] At present, nursing information systems have been widely used in the management of tube feeding care complications, especially in the daily health monitoring of elderly patients, critically ill patients and long-term bedridden populations. However, most of the existing systems still rely on highly structured medical record data such as electronic health records (EHR), such as vital sign values, medication records and laboratory indicators, and ignore a large amount of important non-text and non-structured medical data such as medical images, real-time sensor monitoring data, nursing staff's handwritten or spoken care notes, patient complaints and behavior records. The incompleteness of such data integration makes it difficult for the system to fully capture the multidimensional and multi-modal characteristics of complications, thereby limiting its application effect in early warning and accurate decision support.

[0003] Specifically, although structured data is easy to standardize, it often cannot reflect the real-time state changes and subtle manifestations of patients. For example, a tube feeding patient may not have abnormal body temperature or blood pressure at a specific time point, but his or her feeding tolerance has decreased, the abdomen is slightly distended, the mood is agitated, or discomfort is conveyed through informal expression (such as moaning, gestures or facial expressions). These information is usually recorded in nursing shift reports, voice recordings or temporary observation notes, which are typical non-structured data. If the system only relies on structured data for analysis, it is easy to miss such early signals, and thus delay the identification and intervention of complications such as aspiration pneumonia, intestinal obstruction or gastric retention.

[0004] In addition, the existing information systems mostly use keyword-based retrieval mechanisms, such as searching for terms such as "vomiting" and "abdominal distension" to find relevant records. This method lacks natural language processing and semantic understanding ability, and cannot identify diverse expressions such as "feeling nauseous" and "abdominal distension is uncomfortable", resulting in incomplete information retrieval or low accuracy.

[0005] In terms of knowledge graph construction, traditional methods mostly rely on rule engines or simple graph structures, which are difficult to effectively express complex multi-hop relationships between medical concepts (such as "too fast tube feeding speed" may lead to "delayed gastric emptying", which in turn increases the "risk of aspiration"), and also have problems such as low computational efficiency, slow response, etc., which cannot support real-time decision-making needs. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application aims to provide a tube feeding care complication retrieval recommendation processing method and system to solve the problems raised in the background art.

[0007] To achieve the above object, the present application is realized by the following technical scheme: a tube feeding care complication retrieval recommendation processing method, comprising the steps of:

[0008] By collecting nursing data representing patient tube feeding complications, after natural language processing and semantic vectorization processing, a care behavior sequence is constructed with session ID as row and interaction event type as column to represent patient care needs;

[0009] Real-time monitoring of the care behavior sequence, when detecting user interaction behavior, extracting the original text data associated with the event as a dynamic risk first fragment, and matching the corresponding second fragment from the historical database accordingly; mapping the double-fragment text vocabulary with the medical terminology semantic library, assigning static and dynamic semantic vectors to the mapped concepts and unmapped vocabulary respectively, and constructing a weighted inter-fragment semantic similarity matrix; using dynamic programming algorithm to calculate the optimal semantic alignment path between the double fragments: initializing a semantic similarity matrix M sim with the same dimension as the score matrix M align and the path backtracking matrix M path , and setting the score matrix M align [0][0] to 0, and the first row and first column to the cumulative gap penalty value; iteratively calculating each cell (i,j) in the score matrix M align , where i and j start from 1, based on the maximum cumulative score of vocabulary matching and insertion gap, taking the maximum value as the value of M align [i][j], and recording the source direction of the score in the path backtracking matrix M path , completing the matrix filling; starting from the right bottom corner position [m][n], backtracking to the top left corner [0][0] according to the path backtracking matrix M path , finding the optimal semantic alignment path, at this time, determining the endpoint value M align [m][n] of the path, which is the overall semantic similarity score S segment of the two fragments; finally, comparing S segment with the preset fragment similarity threshold θ high , if S segment ≥ θ high , it is determined that the current first fragment and the second fragment constitute a high matching semantic pair, which is filtered out;

[0010] Convert all high matching semantic pairs into sentence vectors, generate medical feature points representing different complication patterns through clustering algorithm, and assign main tags and auxiliary description tags to them by calculating their similarity with the medical terminology semantic library, to obtain a medical feature set;

[0011] A knowledge graph is constructed based on the medical feature set, and a community is divided and optimized; a top node is taken as a retrieval key to generate final recommended content according to the node weight order of the knowledge graph.

[0012] As a second aspect of the present application, a tube feeding care complication retrieval recommendation processing system comprises a memory and a processor, wherein the memory comprises a tube feeding care complication retrieval recommendation processing program, and the tube feeding care complication retrieval recommendation processing program is implemented when the processor is executed.

[0013] Compared with the prior art, the present application has the following beneficial effects:

[0014] The present application realizes deep fusion and accurate semantic alignment of multi-source heterogeneous nursing data by constructing a care behavior sequence and monitoring dynamic risks in real time; effectively solves the problem of semantic fragmentation of existing medical texts by using a dynamic programming algorithm to screen high-matching semantic pairs and combining clustering to generate medical feature points representing complication patterns; and further constructs a dynamically evolving knowledge graph through a graph attention network and community division, significantly improves the relevance of complication correlation, and greatly improves the accuracy, efficiency and practicability of tube feeding complication intelligent care recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0015] The disclosure of the present application will be described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes, and are not intended to limit the scope of protection of the present application. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0016] Figure 1 The overall processing flowchart of the tube feeding care complication retrieval recommendation processing method proposed in an embodiment of the present application is shown in the figure.

[0017] Figure 2 The structure diagram of the community C2 (cluster of complications of glucose metabolism disorder) obtained when the community division is performed in an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0018] It is easy to understand that, according to the technical solution of the present application, those skilled in the art can propose a plurality of structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the present application. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present application, and should not be regarded as the whole or as a limitation or restriction on the technical solution of the present application.

[0019] The present application will be further described in detail below with reference to the accompanying drawings, but not as a limitation of the present application.

[0020] As an embodiment of the present application, as Figure 1As shown, a tube feeding care complication retrieval recommendation processing method is proposed, which includes the following steps:

[0021] S1, multi-modal nursing data acquisition and semantic processing:

[0022] S1-1, collect multi-modal nursing data of tube feeding complications. First, deploy data collection interfaces, focus on collecting structured medical record data containing keywords such as "gastric retention", "diarrhea", "constipation", "aspiration", "reflux", "hyperglycemia", "refeeding syndrome", etc. and synchronously acquire unstructured text data, including at least: nursing notes describing complications, text descriptions of tube placement or lung infection in imaging reports, and related narrative data in consultation records. Subsequently, collect the interaction data of care personnel when querying tube feeding complications, including: query words input in the retrieval system (such as "What to do if abdominal distension after nasogastric feeding"), case IDs clicked to view, dwell time on specific case pages, and collection or labeling behaviors (such as labeling "Gastric retention treatment plan") and other behavior logs. Finally, send all collected multi-source heterogeneous data to a unified data buffer pool to complete the storage of original query data.

[0023] S1-2, vectorize the original query data. First, preprocess the original query data (such as "Severe diarrhea after nasogastric feeding") obtained from the data buffer pool. Use a Chinese word segmentation tool (such as Jieba word segmentation or a BERT-based word segmentation model) to perform word segmentation, obtaining the word sequence {"nasogastric feeding", "after", "diarrhea", "severe"}. Next, call a custom nursing / medical stopword list to remove stopwords without actual semantics (such as "after"), and perform stem extraction or morphological normalization on the remaining words (such as normalizing "severe" to "heavy"), finally generating the purified word sequence {"nasogastric feeding", "diarrhea", "heavy"}. At this point, this purified word sequence will serve as the basis for the "effective vocabulary set" in subsequent steps.

[0024] Second, use the pre-trained Clinical-BERT model to convert the processed word sequence into a fixed-dimensional query vector q, (such as 768 dimensions) to numerically represent the core semantics of this query.

[0025] Finally, to quantify user behavior, construct a nursing behavior sequence M int , set the row of the sequence to represent the session ID, and the column to represent different interaction event types (such as keyword clicks, case viewing, literature retrieval), and record the timestamp and frequency of event occurrence in the corresponding position in the sequence whenever a user performs a specific behavior in a retrieval session.

[0026] S1-3, generating medical feature points representing care needs through semantic alignment, it is understood that the purpose of this step is to realize semantic fusion and standardization processing of tube feeding complication text data (such as nursing notes and image reports) from different care needs and different expressions. By constructing a dynamic and iterative semantic mapping framework, the enhanced weighted semantic similarity between heterogeneous text segments is calculated, and medical feature points representing specific complication patterns are generated.

[0027] The specific implementation steps include:

[0028] S1-31, the system monitors the care behavior sequence M int , set the frequency change threshold or the recent trigger time window, when it is found that the frequency of "keyword click" event type in a certain session (such as SessionID: S001) suddenly increases beyond the threshold, or the latest timestamp is located within the time window, it is indicated that the event (for example, the user clicks the keywords "choking cough" and "SpO2 drop") is triggered, it is determined that the user interaction behavior is active, then the original text data description (for example: "after nasal feeding, the patient has severe choking cough, and SpO2 drops to 88%") associated with the event is extracted, and it is defined as the first segment (that is, the segment containing the dynamic to-be-solved care risk event).

[0029] Then, taking the core medical terms (such as "choking cough" and "SpO2 drop") in the first segment as the retrieval key, a matching query is performed in the historical database in parallel, so as to obtain the corresponding standardized information units, such as "responding to choking cough: immediately stop tube feeding, take right lateral position …" retrieved from the standard nursing plan, or "chest X-ray shows patchy blurred shadow in the right lower lung field" matched from the historical image report. The above static structured knowledge unit is defined as the second segment (that is, the segment containing the static structured knowledge), which provides a reference solution and care basis for real-time processing of the first segment. In this embodiment, the historical database is a knowledge base containing structured knowledge units related to patient tube feeding complications (such as standard nursing plans, historical case data key information).

[0030] Based on the above technical concept, it should be noted that different sources of text (such as real-time nursing records and standard guidelines) have great differences in expression, detail, and term usage, such as a medical concept can have multiple expressions ("stomach distension", "abdominal distension", "all refer to", "abdominal distension"), and the same expression refers to different concepts in different contexts. Based on this, step S1-32 needs to be performed to perform vocabulary analysis on the two segments, and the vocabulary in the two segments is "translated" or "aligned" to a unified concept space, thereby providing a data basis for subsequent accurate calculation of the overall semantic similarity between the two segments.

[0031] S1-32、First, receive the original text data of the first segment and the second segment output in step S1-31, and independently perform natural language processing on the text of the two segments, respectively, denoted as the first segment valid vocabulary set V first and the second segment valid vocabulary set V second , such as {"nasogastric feeding", "severe", "choking cough", "SpO2", "decrease"} and {"chest X-ray", "right lower lung field", "patchy", "blurred", "density shadow", "inhale"}), so as to convert the data describing the care behavior into a vocabulary set representing the core medical semantics thereof;

[0032] Secondly, traverse the first segment valid vocabulary set V first , first , fi , j , , , , , fi , , , , j ,

[0033] It can be understood that the medical term semantic library is a pre-generated readable knowledge base containing authoritative standardized care medical terms and semantic relationships thereof, and the vectors thereof are obtained by training a deep learning model (such as BioBERT, Clinical-BERT) on a large amount of medical texts (such as textbooks, care medical guidelines, and high-quality medical records), and generally include symptoms, signs, diseases, operations, drugs, and other medical concepts representing care behaviors.

[0034] At this time, for a word w fi , an approximate nearest neighbor search algorithm needs to be used to calculate the cosine similarity between the vector of the word and the vector of each standard concept c j in the medical term semantic library, and a total of N similarity values are obtained, the maximum value in all similarity results is found in the N similarity values, and if the maximum cosine similarity exceeds a preset vocabulary concept mapping threshold θ map(e.g. 0.75), a unique vocabulary w is established fi to the standard concept c j . fi = c jmax , indicating that the term represented by the vocabulary has been successfully anchored to a standard concept, c jmax is the standard concept that achieves the maximum similarity Sim max ; otherwise, it means that a sufficiently similar corresponding concept cannot be found in the concept library, and it is marked as an unmapped vocabulary w fiu .

[0035] Finally, the same operation is performed on the vocabularies in V second , which is exactly the same as above, and the second segment effective vocabulary set V second is obtained. The mapping Map(w sk ) of each vocabulary w j to the standard concept c sk and the unmapped vocabulary w sku in the second segment effective vocabulary set V first is obtained. Thus, it is ensured that the vocabulary of each segment is bound to at most one standard concept, and a one-to-one or zero mapping relationship is finally formed, ensuring the fairness and consistency of the alignment of the vocabularies of the two segments in the same standard concept space.

[0036] Based on the above technical concept, it can be understood that the purpose of obtaining the mapped concept is to convert the above vocabulary into a standard concept, ensuring the accuracy and consistency of the core semantics of the system. For example, it is ensured that "stomach swelling" and "abdominal distension" point to a unique "Abdominal Distention", which is the basis for building reliable medical knowledge. The purpose of obtaining the unmapped vocabulary is to identify "edge" vocabularies that cannot be simply standardized. It should be noted that these vocabularies often contain key information rich in care details. Separating them out is to mark objects that need special processing, so that the system has more intelligent manifestations than traditional dictionary-based systems, and can understand free and diverse care descriptions.

[0037] S1-33, by calculating the optimal semantic alignment path of all vocabulary pairs (mapped standard concepts and unmapped vocabularies) between the first segment and the second segment, quantifying the overall semantic relevance of the two segments, to determine whether the dynamic to-be-solved care risk event and the static structured knowledge unit describe the same care demand problem.

[0038] Specifically, first, based on the processing results of S1-32, the effective unit list L first of the first segment is formed (containing mapped standard concepts Map(w fi ) and unmapped vocabularies w fiu) and the effective unit list L of the second segment second (contains Map(w sk ) and un-mapped words w sku ), load the predefined part-of-speech weight strategy (noun weight ω noun =1.0, adjective ω adj =0.8, verb ω verb =0.6), assign different parts of speech to the above effective unit list to represent their different importance in medical description.

[0039] Secondly, generate a unified high-dimensional semantic vector for all effective unit lists as the basis for calculation. For mapped standard concepts, directly read their pre-generated static semantic vectors from the medical terminology semantic library where they are located ; for un-mapped words, put them back into the context of their original segment (e.g. "patient has severe cough after nasogastric feeding"), input them into the pre-trained context-aware model (such as Clinical-BERT) in step S1-2, and take the hidden state (HiddenState) in the output layer of the model corresponding to the position of the word as its dynamic semantic vector .

[0040] Thirdly, construct an mxn inter-segment semantic similarity matrix M sim , set its rows to the effective unit list L of the first segment first , and its columns to the effective unit list L of the second segment second . For each element (u i , v j ) in the matrix, it is calculated by the weighted cosine similarity between the first segment effective unit u i and the second segment effective unit v j , the formula is: (2), where, is used to calculate the standard cosine similarity between vectors and , the value range is [-1, 1], respectively represent the weight coefficient assigned based on the part of speech of unit u i and v j , to ensure that the matching of key medical entities (nouns) representing care needs contributes most to the overall similarity, while the matching of modifying words (adjectives, verbs) contributes relatively less. For example, the matching weight of two nouns is (1.0+1.0) / 2=1.0, while the matching weight of one noun and one adjective is (1.0+0.8) / 2=0.9.

[0041] Finally, use the dynamic programming algorithm to find the maximum weighted sum of the semantic similarity matrix M simThe optimal semantic alignment path with the highest overall alignment score is found.

[0042] It can be understood that the logic of the dynamic programming algorithm proposed in the embodiment is to decompose it into a series of interrelated sub-problems by finding the optimal alignment, that is, the optimal score of each position [i][j] existing in the semantic similarity matrix M sim , and solve it step by step by filling a table, that is, M align , to avoid repeated calculation, and the purpose is to calculate the semantic correlation score S segment of the two segments as a whole. The specific process is as follows:

[0043] First, initialize a score matrix M align and a path backtracking matrix M path , and at the same time, set the dimensions consistent with the semantic similarity matrix M sim , the number of rows is equal to the length m of the effective unit list L first of the first segment, and the number of columns is equal to the length n of the effective unit list L second of the second segment.

[0044] It can be understood that, in the prior art, the method based on keyword matching or local similarity calculation can only capture isolated, point-to-point lexical associations (for example, only find that "choking cough" and "inhale" are similar in themselves), and cannot determine from the whole whether the dynamically to be solved care risk event description "after nasal feeding of the patient, severe choking cough occurs, SpO2 decreases" and the static structured knowledge unit description "deal with choking cough: immediately stop tube feeding... be vigilant about aspiration pneumonia" express the same care risk (i.e. aspiration risk).

[0045] Based on this, the initialization score matrix M align proposed by the application is to fill and record the cumulative optimal score in all possible alignment schemes from the beginning of the segment to any position (i, j) by the dynamic programming algorithm, so as to quantify the overall semantic correlation from the starting point to the ending point.

[0046] Further, the proposed initialization path backtracking matrix M path is to accurately record the best source direction (such as diagonal matching or insertion of empty space) of the score of each cell in the score matrix M align , so as to backtrack to construct the specific lexical alignment path that makes the overall alignment score highest after the calculation is completed (such as the path contains the pair of high-score matching "choking cough-inhale" and the possible empty space adjustment). Finally, according to the total path score S segment , the judgment is made to realize accurate semantic alignment and care demand problem identification.

[0047] It is to be noted that in the dynamic programming process, any position (i, j) refers to the same position in the score matrix M align and semantic similarity matrix M sim . In the initialization phase, the score matrix M align [0][0] is set to 0, and the first row and the first column are respectively accumulated with a gap penalty value gap (usually set to a fixed negative value, typically gap = -0.5, to handle the case where the units cannot be aligned).

[0048] Secondly, the score matrix M align is filled row by row and column by column. For each position (i, j), the score of the three directions is calculated, and the iteration is started row by row (i from 1 to m) and column by column (j from 1 to n) to fill each cell M align [i][j] of the matrix M align :

[0049] The diagonal score from position (i-1, j-1) can be understood as the score of aligning the i-th unit of the first fragment with the j-th unit of the second fragment, and its value is the score of the last diagonal position M align [i-1][j-1] plus the weighted cosine similarity M sim (i, j);

[0050] The upward score from position (i-1, j) (the unit of the first fragment is aligned with a gap), which is derived from the i-th unit of the first fragment not being aligned with any unit of the second fragment but with a gap, and its value is the score of the position directly above M align [i-1][j] plus a gap penalty value gap;

[0051] The leftward score from position (i, j-1) (the unit of the second fragment is aligned with a gap), which is derived from the j-th unit of the second fragment not being aligned with any unit of the first fragment but with a gap, and its value is the score of the position to the left M align [i ][j-1] plus the same gap penalty value gap;

[0052] Then, the maximum of the three is taken as the value of M align [i][j], i.e., , and the source direction is recorded in the path backtracking matrix M path .

[0053] Thirdly, after the filling is completed, the score matrix M alignThe lower right corner [m][n] starts backtracking to the upper left corner [0][0], collecting all the unit pairs along the maximum score path to form the optimal semantic alignment path, and obtaining the final score of the path:

[0054] S segment =M align [m][n], which is the overall semantic similarity score of the two segments, can be understood that its value directly depends on the cumulative and combination of all weighted cosine similarities in the semantic similarity matrix M sim , thereby realizing the quantification of the overall semantic correlation of the two segments, to determine whether the dynamic to-be-solved care risk event and the static structured knowledge unit description describe the same care problem, for example, if the path contains high matching pairs such as ("choking cough", "inhalation") and the overall score is high, it indicates that they are highly related. At this time, the lexical pair recorded on the path (such as "choking cough" of the first segment and "inhalation" of the second segment) is the most contributing key semantic alignment pair.

[0055] Fourthly, a higher segment similarity threshold θ high (e.g. 0.85) is set, and all to-be-processed (first segment, second segment) pairs are traversed, and only when S segment ≥ θ high , the key semantic alignment pair is identified as a "high matching semantic pair" and is screened out. Generally speaking, if the core medical entities (nouns) of the two segments are successfully matched and the score is high (the weighted weight is 1.0), the segment similarity threshold θ high is easy to reach.

[0056] For text data pairs that do not reach this threshold, based on the constructed feedback strategy: analyze the key lexical pairs in the optimal semantic alignment path, if there are high-frequency co-occurrences (co-occurrence times ≥ 5 times) but not successfully mapped lexical pairs (such as "gastric residual volume" and "gastric emptying disorder" are not mapped, but they commonly appear in multiple text pairs and the alignment score is high (the average alignment score exceeds the set threshold (≥ 0.7))), they are used as new concept candidates, and these new concept candidates are returned to the construction and updating process of the medical terminology semantic library to realize the continuous iteration and optimization of the knowledge base.

[0057] Fifthly, in order to aggregate the dispersed "high matching semantic pair" text segments (such as "patient choking cough with SpO2 drop" and "chest X-ray suggests aspiration pneumonia") into unified concepts with clear nursing significance, this step performs medical feature point generation and labeling:

[0058] Firstly, all the text segments (either the first segment or the second segment) in the "high matching semantic pairs" are converted into fixed-dimensional sentence-level semantic vectors by the sentence encoder (Sentence-BERT), forming a high-quality semantic vector pool.

[0059] Secondly, the DBSCAN density clustering algorithm (with the neighborhood radius parameter eps = 0.6 and the minimum sample number min samples = 5) is applied to the vector pool for clustering analysis to automatically discover semantic-dense clusters from the data distribution itself. Each identified dense cluster is defined as a medical feature point representing a specific complication pattern, and the core vector representation is determined by the centroid (mean vector) of all sentence vectors in the cluster. It should be noted that the centroid vector in this embodiment aims to provide an accurate and computable numerical definition for the abstract "medical concept". For example, when a patient queries "what to do if breathing is not good after nasogastric feeding", the system can also convert the query statement into a vector and quickly find the medical feature point (such as "aspiration pneumonia" and "atelectasis") that is "closest" to it (i.e., with the highest cosine similarity) in the semantic space, thereby achieving accurate recommendation.

[0060] Thirdly, a rich semantic description is generated for each generated medical feature point: first, calculate the cosine similarity between the centroid vector and all concept vectors in the medical term semantic library, and determine the standard concept (such as "aspiration pneumonia") with the highest similarity as the main label; second, perform word segmentation and part-of-speech filtering on all text in the cluster, extract nouns and gerunds, and count the word frequency, and the highest frequency term (such as "choking cough" and "aspiration") is used as an auxiliary description label.

[0061] Finally, all these medical feature points constitute a medical feature set FP i In this embodiment, the feature point data includes: {FP_001: {"main label": "aspiration pneumonia", "auxiliary description label": ["choking cough", "aspiration", "nasogastric feeding", "SpO2 drop"], "centroid vector": [0.12, -0.05,..., 0.78] / / 768-dimensional float array}; FP_002: {"main label": "gastric retention", "auxiliary description label": ["abdominal distension", "gastric residual volume", "nasogastric feeding", "stop feeding"], "centroid vector": [-0.23, 0.15,..., -0.42] / / 768-dimensional float array}}.

[0062] For ease of understanding, as shown in Table 1 below, it can be intuitively obtained that the medical feature set FP i Detailed description of each generated medical feature point.

[0063] Table 1:

[0064]

[0065] S2, construct a complication knowledge graph to map the medical feature set FP generated in S1 stage i Construct a multi-level and dynamically evolving knowledge graph to enable care reasoning.

[0066] The specific implementation steps include:

[0067] S2-1, based on semantic similarity matrix M sim Construct an initial graph to convert discrete medical features into a graph with a basic topology.

[0068] The specific process is as follows: with the generated medical feature set FP i As a graph node, set each node to represent a care complication feature point (such as "aspiration pneumonia" and "gastric retention"), set the node attributes to include its main label, auxiliary description label semantic information, calculate the cosine similarity score of each feature point in the semantic space, determine whether to establish a connection according to the preset initial edge construction threshold, establish a directed edge for the node pair that exceeds the threshold, and set the similarity value as the initial weight of the edge, thereby forming an initial, undirected and weighted initial graph G in , which preliminarily embodies the explicit association between complications.

[0069] S2-2, introduce a multi-head graph attention network (GAT) to optimize the structure of the initial graph G in to form an optimized graph Gop with clearer structure and richer semantics. The specific process uses the GATConv layer provided by the existing graph neural network library (such as PyTorchGeometric) to realize:

[0070] First, input the semantic features of each node of the initial graph G in into the GAT network (consisting of the centroid vectors of each feature point, which can be understood as each medical feature point (such as FP_001) becoming a node of the knowledge graph, and its centroid vector being the core attribute of this node) to perform linear transformation and feature mapping using the ready-made GATConv layer (with a number of heads set to 8) to project it to a low-dimensional semantic subspace (such as 128 dimensions) to obtain the transformed node feature representation, while using the LeakyReLU activation function (with a default negative slope of 0.2) inside the layer to calculate the attention coefficients between the transformed nodes , the formula is as follows (4), where a and W are both trainable / learnable parameters inside the GAT network, represents vector concatenation, are the original feature vectors (i.e. centroid vectors) of node i and node j, respectively.

[0071] Secondly, the attention score on all incident edges of each node is normalized by softmax to obtain the normalized attention weight αij;

[0072] Finally, the neighbor node features are weighted and summed according to these weights, and the new feature representation of the node is output through the ELU activation function. It can be understood that the entire optimization process is based on the standard forward propagation and parameter update mechanism of the existing graph attention network (GAT) and does not need to be coded from zero. The optimized edge weight is derived from the attention coefficient, and can be sparsified according to the set threshold to remove the weak connection edge, and finally obtain the optimized graph G op , which significantly improves the care interpretability and structural quality of the graph.

[0073] S2-3: After obtaining the optimized graph G op , identify the complication community with close internal connections and sparse external connections.

[0074] The specific implementation steps include: using the Louvain algorithm based on modularity optimization to iteratively divide the community by maximizing the modularity Q value of the entire graph attention network (GAT).

[0075] It can be understood that the modularity (Modularity) Q is a core indicator for measuring the quality of community division, and its calculation formula is: , wherein, is the edge weight between nodes i and j in the optimized graph G op , ki is the weight of all edges connected to node i and is half of the total weight of all edges in the optimized graph G op , ci is the community to which node i belongs, is the Kronecker function, which is 1 when the two nodes belong to the same community, i.e. ci=cj, otherwise 0.

[0076] In this embodiment, when dividing the community, first, each node in the optimized graph G op is regarded as an independent community, then each node i is traversed, the modularity gain ΔQ brought by moving it to each neighbor community is calculated, the community that can bring the maximum positive gain ΔQ is selected, and the node i is moved to the community, if all movements cannot produce positive gain, the node stays in the original community; then the nodes belonging to the same community are aggregated into new nodes, and the above process is repeated on the new graph until convergence.

[0077] At this point, as Figure 2As shown, a stable community partition is output, each community being a care-guiding meaningful complication community consisting of GAT-optimized semantic relations (e.g., "reflux and aspiration cluster"), thereby realizing automatic discovery and quantification of complication complex semantic relations from data.

[0078] For example, community C1: {vomiting, reflux, bed head elevation, aspiration, aspiration pneumonia, tachypnea, SpO2 drop}; the care interpretation is that this community reveals that "reflux" and "aspiration" may be caused by "insufficient bed head elevation" during tube feeding, which in turn leads to "aspiration pneumonia", and the care shows a high correlation care path of "tachypnea" and "SpO2 drop". At this time, this community is named "reflux and aspiration and related complications cluster".

[0079] Community C2: {hyperglycemia, tube feeding rate too fast, insulin use, polyuria, dehydration}; the care interpretation is that this community reflects that "hyperglycemia" is caused by "tube feeding rate too fast", which in turn requires "insulin use", accompanied by complications such as "polyuria" and "dehydration". At this time, this community is named "disorder of glucose metabolism complication cluster".

[0080] Finally, the above community information (i.e., which community each node belongs to) is used as a new attribute and is fed back and updated to the optimized graph G op , forming a final knowledge graph G opfinal containing community semantic labels.

[0081] S3, generate a refined search area:

[0082] It should be noted that, in general, in the construction of a complication knowledge graph, each complication theme community is usually composed of dozens of highly correlated medical nodes, covering a complete care recommendation path from etiology, symptoms, signs to intervention measures. If direct search is performed in the full graph range, not only is the calculation efficiency low, but also a large amount of irrelevant information may be returned, affecting the accuracy and response speed of care decisions.

[0083] For ease of understanding, based on the community partition result of step S3, taking the "reflux and aspiration and related complications cluster" (community C1) as an example, in addition to the main path mentioned above, there are several high-correlation care paths with different focuses as shown in Table 2. These paths share the same core complication theme, but are triggered by different initial factors or exhibit different symptom combinations, ultimately leading to similar serious complications.

[0084] Table 2:

[0085]

[0086] Based on this, the present step S3 is proposed, aiming to avoid inefficient full-graph search on the entire huge knowledge graph.

[0087] The specific implementation steps include:

[0088] S3-1, knowledge graph G containing community semantic tags opfinal Perform aggregation analysis to identify the complication topic community that the user is most likely to focus on. In the present embodiment, it is preferred that the analysis considers the following two aspects:

[0089] Quantify the internal importance of the community (graph structure level) by calculating the average attention weight of the nodes in the community to evaluate the semantic correlation tightness and overall importance within the community. The higher the weight, the stronger and more critical the internal connection of the care concept represented by the community.

[0090] Quantify the relevance of the community to the user (user intention level) by comparing the community topic with the user's historical query vector and interaction behavior log to determine the matching degree of the community content and the user's real care demand, and ensure that the recommended results are highly relevant to the user's interest.

[0091] Subsequently, the comprehensive attention degree of each community C k is calculated, which is used as the quantitative basis for sorting and screening the core community. The comprehensive attention degree calculation formula is: (6), wherein, is the size of the community divided from the knowledge graph by the Louvain algorithm, ni is the i-th node in the community, is a medical feature point (such as "aspiration pneumonia" and "gastric retention") generated by the S1 step, which is the most basic semantic unit in the knowledge graph, and the function is calculated based on the similarity between the semantic of the query vector q in step S1-2 and the community topic, which is usually an indicator function, and is 1 when the community topic matches the user's historical query intention, is used to calculate the average value of the attention weights of all nodes in the community. The higher the average value, the tighter the internal association of the community and the more important the overall, is the attention weight of the node ni, which is derived from the optimization output of the graph attention network (GAT) in step S2-2, and is used to numerically represent the influence and importance of the node in the context of its neighbor nodes. The higher the weight, the more critical the node, is a smoothing reward term for large-scale communities, which is used to avoid the absolute dominance of large-scale communities. Only when the community topic matches the user's query intention, f(q)=1, this term takes effect, and gives greater reward (taking the natural logarithm) to large-scale communities, so as to preferentially recommend those high-value communities that are both large-scale and relevant in the ranking.

[0092] Example: If the user history query intention contains keywords such as "gastric retention" and "abdominal distension", and the query vector q generated based on step S1-2 matches the theme of community C3 (the theme of the community is "gastric retention and related complications"), then =1, if community C3 contains 10 nodes, and the average attention weight is 0.8, then the comprehensive attention degree of this community C3 is Cl(C3)=0.8+ln(10)×1=0.8+2.3=3.1.

[0093] S3-2, select the top M=3 communities with the highest Cl scores, which represent the complication theme communities that the system believes the user is most interested in. Extract all nodes in these communities to form a basic node set, and then calculate the convex hull boundary of the basic node set: , where is the convex hull function, and the function output is the convex hull H corresponding to the point set, represents the coordinates of node ni in a certain vector space. Generally, this space is usually a low-dimensional semantic space (for example, the 128-dimensional space projected by the GAT network in step S2-2, which is used to visualize the calculation of the convex hull in the embodiment), is the union operation of the set.

[0094] S3-3, aggregate all nodes ni in the top 3 communities with the highest comprehensive attention degree by the above formula, and calculate the convex hull H of the coordinates points corresponding to these nodes in the semantic space. Form the convex hull region Rc within the range defined by the convex hull. It can be understood that, since the comprehensive attention degree Cl(C k ) is the aggregation of all nodes in the top M complication theme communities (such as "reflux aspiration cluster" and "gastrointestinal symptom cluster"), it has fused the internal importance of the community in the graph structure layer (based on the attention weight optimized by GAT) and the query relevance in the user intention layer, thereby ensuring that the nodes in Rc are both key data in the graph and highly matched with the user's real-time care needs. Therefore, the convex hull region Rc has fully covered the core medical feature points most relevant to the current search session, and can be determined as the final refined search region, covering all nodes in the three communities.

[0095] S4, output the final search result:

[0096] For each medical feature point node ni in the convex hull region Rc, an attention weight thereof in the knowledge graph is ranked in descending order, and after the ranking is completed, a standardized medical concept (such as "gastric retention") corresponding to the node with the highest ranking / top ranking is taken as a core search key, a plurality of source knowledge bases associated therewith are queried through a preset data interface, and specific recommended content is generated and output. For example, when "abdominal distension after nasogastric feeding" is queried, the output "recommend referring to the 'gastric retention management' chapter in the 'tube feeding complication prevention and treatment guidelines' and the abdominal distension treatment process in the related nursing scheme" is taken as a search result.

[0097] As a second aspect of the present application, a tube feeding care complication search recommendation processing system is provided, including a memory and a processor, wherein the memory includes a tube feeding care complication search recommendation processing program, and the tube feeding care complication search recommendation processing program is executed by the processor to implement the tube feeding care complication search recommendation processing method.

[0098] The technical scope of the present application is not limited to the content in the above description, and those skilled in the art can make various modifications and changes to the above embodiments without departing from the technical idea of the present application, and these modifications and changes should all belong to the protection scope of the present application.

Claims

1. A method for retrieving and recommending treatments for complications in tube feeding care, characterized in that: Including the following steps: By collecting nursing data that characterizes patients’ tube feeding complications, and after natural language processing and semantic vectorization, a care behavior sequence is constructed with session ID as the row and interaction event type as the column to characterize patients’ care needs. The care behavior sequence is monitored in real time. When user interaction behavior is detected, the original text data associated with the event is extracted as the first segment of dynamic risk, and the corresponding second segment is matched from the historical database accordingly. The vocabulary of two-segment texts is mapped to a medical terminology semantic database. Mapped concepts and unmapped words are assigned static and dynamic semantic vectors, respectively, and a weighted semantic similarity matrix between segments is constructed. A dynamic programming algorithm is used to calculate the optimal semantic alignment path between the two segments: an initialization matrix M is created. sim Score matrices M of the same dimensions align and path backtracking matrix M path And in the initialization phase, the score matrix M is... align [0][0] is set to 0, and its first row and first column are initialized with the accumulated empty space penalty value; the score matrix M is calculated iteratively. align For each cell (i,j) in M, where i and j start from 1, the maximum cumulative score is taken as M based on the word matching and the maximum cumulative score under the case of inserting gaps. align The value of [i][j], and in the path backtracking matrix M path Record the source direction of the score in the middle and complete the matrix filling; starting from the bottom right position [m][n], backtrack the matrix M according to the path. path Backtracking to the top left corner [0][0], find the optimal semantic alignment path. At this point, determine the endpoint value M of this path. align [m][n] represents the overall semantic similarity score S between the two segments. segment Finally, S segment Similarity threshold θ between the segments high Compare, if S segment ≥θ high If the first segment and the second segment form a highly matched semantic pair, then filter and output them. All highly matched semantic pairs are converted into sentence vectors, and medical feature points representing different complication patterns are generated by clustering algorithms. The main label and auxiliary descriptive label are assigned to them by calculating their similarity with the medical terminology semantic database, thus obtaining a medical feature set. A knowledge graph is constructed using the aforementioned set of medical features, and communities are optimized and segmented. Based on the node weights of the knowledge graph, the top-level nodes are used as search keys to generate the final recommended content.

2. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 1, characterized in that: The medical terminology semantic database is a pre-generated, readable knowledge base containing standardized care medical terms and their semantic relationships; The historical database is a knowledge base that includes structured knowledge units related to complications associated with tube feeding in patients. During real-time monitoring of the care behavior sequence, by setting a frequency change threshold or a recently triggered time window, when the frequency of an interaction event type increases beyond the threshold within a unit of time or its latest timestamp falls within the time window, it is determined that the user's interaction behavior is active. At this time, the original text data associated with the active interaction event is extracted and defined as a first segment containing dynamic unresolved care risk events. Simultaneously, using the core medical terms in the first segment as search keys, a parallel matching query is performed in the historical database to obtain the corresponding standardized information, which is defined as a second segment containing static structured knowledge, providing a reference for the real-time processing of the first segment.

3. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 1, characterized in that: The specific steps for mapping bi-segment text vocabulary to a medical terminology semantic database and constructing a weighted inter-segment semantic similarity matrix are as follows: The original text data in the first and second segments are converted into effective vocabulary sets, respectively. For each vocabulary, its cosine similarity to all standard concept vectors in the medical terminology semantic database is calculated. If the maximum similarity exceeds a preset vocabulary-concept mapping threshold, a mapping relationship is established between the vocabulary and its corresponding standard concept; otherwise, it is marked as an unmapped vocabulary. Based on the mapping results, effective unit lists for the first and second segments are formed, respectively. These lists contain mapped standard concepts and unmapped vocabulary, and predefined part-of-speech weights are loaded to characterize their different importance in medical descriptions. A unified semantic vector is generated for all effective unit lists: for mapped standard concepts, their pre-generated static semantic vectors are directly read from the medical terminology semantic database. For the unmapped words, they are placed back into the original segment context, and the hidden state of their position is obtained as a dynamic semantic vector using a pre-trained model. Construct an m×n semantic similarity matrix M between segments. sim Set its rows to correspond to the list of valid units in the first segment, and its columns to correspond to the list of valid units in the second segment; For each element (u) in the matrix i ,v j Its value is calculated using the following formula: In the formula, Used to calculate vectors and Standard cosine similarity between them These represent the weight coefficients assigned based on the part-of-speech tag of each element in the matrix.

4. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 1, characterized in that: It also includes steps for processing highly matched semantic pairs that do not reach the segment similarity threshold: For S segment <θ high The text data pairs are analyzed, and the word pairs in their optimal semantic alignment path are analyzed. If there are word pairs that frequently co-occur but are not successfully mapped, they are used as candidate new concepts and fed back to the system to optimize the medical terminology semantic database.

5. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 1, characterized in that: The specific steps for generating a medical feature set are as follows: All text fragments in the selected high-match semantic pairs are converted into sentence-level semantic vectors using a sentence encoder. The DBSCAN clustering algorithm is then applied to cluster these sentence-level semantic vectors, defining each identified dense cluster as a medical feature point representing a specific complication pattern. Finally, each medical feature point is assigned a semantic description: by calculating the similarity between its centroid vector and a medical terminology semantic database, the most similar standard concept is set as the primary label, and high-frequency terms within the cluster are statistically analyzed to form auxiliary descriptive labels. Ultimately, all feature points constitute a medical feature set.

6. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 1, characterized in that: The nursing data includes structured medical record data acquired based on the deployed data acquisition interface, text used to describe complications, and unstructured text data including narrative data in consultation records. After the nursing data is sent to the system data buffer pool to complete the original query data storage, the specific steps for natural language processing and semantic vectorization processing of the original query data are as follows: First, the original query data is segmented, stop words are removed, and word form normalization is performed to obtain a cleaned word sequence. Second, the cleaned word sequence is converted into a fixed-dimensional query vector using a pre-trained Clinical-BERT model. Finally, a care behavior sequence is constructed with session ID as the row and interaction event type as the column, and the timestamp and frequency of each event in the sequence are recorded.

7. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 5, characterized in that: The specific steps for constructing a knowledge graph based on the aforementioned medical feature set are as follows: Each feature point in the medical feature set is used as a graph node, and the node attributes are set to include its main label and auxiliary descriptive label. The cosine similarity score of each feature point in the semantic space is calculated. Based on the preset initial edge construction threshold, it is determined whether to establish a connection and the similarity value is used as the edge weight to form the initial graph G. in ; Introducing a graph attention network to optimize the structure of the initial graph: The initial graph G... in The semantic features of each node are input into the GAT network, and linear transformation and feature mapping are performed using the GATConv layer to obtain the transformed node feature representation. Simultaneously, the LeakyReLU activation function is used within this layer to calculate the attention coefficients between the transformed nodes. The formula is as follows: In the formula, a and W are both trainable / learnable parameters within the GAT network. This represents vector concatenation. These are the original feature vectors of node i and node j, respectively; Perform softmax normalization on the attention scores of all incident edges of each node to obtain normalized attention weights; Based on these weights, the features of neighboring nodes are weighted and summed, and then the new feature representations of the nodes are output through the ELU activation function to obtain the optimized graph G. op ; The Louvain algorithm is used to perform community partitioning on the optimized graph, and the community partitioning results are integrated back into the optimized graph G as new attributes of the nodes. o In the process, the final knowledge graph G is formed. opfinal .

8. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 1, characterized in that: After optimizing and dividing communities based on the knowledge graph, and before generating the final recommended content by sorting the nodes according to the knowledge graph's node weights, it is also necessary to calculate C for each community. k The overall attention received is used to filter core communities in the knowledge graph and generate refined search areas to avoid inefficient full-graph searches; The formula for calculating overall attention is: In the formula, Let ni represent the community size segmented from the knowledge graph using the Louvain algorithm. The i-th node in the knowledge graph represents the most basic semantic unit, and the function... For indicator functions, Used to calculate the average of the attention weights of all nodes within the community. Let be the attention weight of node ni. For smooth rewards for large-scale communities.

9. A system for retrieving and recommending treatments for complications in tube feeding care, comprising a memory and a processor, wherein, The memory includes a tube feeding care complication retrieval and recommendation processing program, which, when executed by the processor, implements the tube feeding care complication retrieval and recommendation processing method as described in any one of claims 1 to 8.

Images