Tube feeding nursing complication retrieval recommendation processing method and system

Abstract

The invention provides a tube feeding nursing complication retrieval recommendation processing method and system, and the method comprises the steps: collecting multi-modal nursing data including structured medical record data and unstructured text data, and storing the multi-modal nursing 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 a community semantic tag; expanding to form a refined retrieval area based on the matching degree of the user query intention and the community theme; and sorting the nodes in the refined retrieval area, and outputting a retrieval recommendation result. According to the method, a local refined retrieval strategy is provided, so that the system can only perform deep retrieval and correlation analysis on key complication information and simplify other information according to the query intention and nursing requirements of nursing personnel, the calculated amount of the system is remarkably reduced, and meanwhile, the retrieval efficiency is improved; and the key complication information can be quickly presented and updated in real time.

Description

Technical Field

[0001] This invention relates to the field of intensive care management technology, specifically to a method and system for retrieving and recommending treatments for complications in tube feeding care. Background Technology

[0002] Currently, nursing information systems are widely used in the management of complications in tube feeding care, playing a crucial role, especially in the daily health monitoring of elderly patients, critically ill patients, and long-term bedridden individuals. However, most existing systems still rely heavily on highly structured medical record data such as electronic health records (EHRs), including vital signs, medication records, and laboratory indicators, while neglecting a large amount of significant untextual and unstructured medical data, such as medical images, real-time sensor monitoring data, handwritten or dictated care notes from nursing staff, and records of patient complaints and behavioral manifestations. This incomplete data integration makes it difficult for the system to fully capture the multidimensional and multimodal characteristics of complications, thus limiting its effectiveness in early warning and precise decision support.

[0003] Specifically, while structured data is easy to standardize, it often fails to reflect real-time changes in a patient's condition and subtle manifestations. For example, a tube-fed patient may not exhibit abnormalities in temperature or blood pressure at a specific time, but may show decreased feeding tolerance, slight abdominal distension, agitation, or discomfort conveyed through informal expressions (such as groans, gestures, or facial expressions). This information is typically recorded in nursing handover reports, voice recordings, or temporary observation notes, which are typical examples of unstructured data. If the system relies solely on structured data for analysis, it is highly likely to miss such early signs, thereby delaying the identification and intervention of complications such as aspiration pneumonia, intestinal obstruction, or gastric retention.

[0004] Furthermore, existing information systems mostly employ keyword-based retrieval mechanisms, such as searching for related records by searching for terms like "vomiting" or "bloating." This method lacks natural language processing and semantic understanding capabilities, and cannot recognize diverse expressions such as "feeling nauseous" or "having a painfully bloated stomach," resulting in incomplete or inaccurate information retrieval.

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

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for retrieving and recommending treatments for complications in tube feeding care, thereby resolving the problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides a method for retrieving and recommending treatments for complications in tube feeding care, comprising 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 system monitors the care behavior sequence in real time. When user interaction is detected, the raw text data associated with the event is extracted as the first dynamic risk segment, and the corresponding second segment is matched from the historical database accordingly. The vocabulary of the two segments is mapped to a medical terminology semantic database. Static and dynamic semantic vectors are assigned to mapped concepts and unmapped words, 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: by initializing the score matrix and the path backtracking matrix, the maximum cumulative score of each unit in the score matrix based on word matching and insertion of gaps is iteratively calculated, and the similarity score reflecting the overall semantic relevance of the two segments is obtained by backtracking the path. Finally, a threshold is set to filter out highly matched semantic pairs. 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 medical feature set, and communities are optimized and divided. Based on the node weights of the knowledge graph, the top-level nodes are used as search keys to generate the final recommended content.

[0008] As a second aspect of the present invention, a tube feeding care complication retrieval and recommendation processing system includes a memory and a processor, wherein the memory includes a tube feeding care complication retrieval and recommendation processing program, and when the tube feeding care complication retrieval and recommendation processing program is executed by the processor, the tube feeding care complication retrieval and recommendation processing method is implemented.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves deep fusion and precise semantic alignment of multi-source heterogeneous nursing data by constructing a care behavior sequence and monitoring dynamic risks in real time. It effectively solves the problem of semantic fragmentation in existing medical texts by using dynamic programming algorithms to screen highly matched semantic pairs and combining them with clustering to generate medical feature points that represent complication patterns. Furthermore, it constructs a dynamic evolutionary knowledge graph through graph attention networks and community partitioning, which significantly improves the relevance of complication relationships and greatly enhances the accuracy, efficiency, and practicality of intelligent care recommendations for tube feeding complications. Attached Figure Description

[0010] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein: Figure 1 This is a schematic diagram of the overall processing flow of the tube feeding care complication retrieval and recommendation treatment method proposed in one embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of community C2 (complication cluster of glucose metabolism disorder) obtained during community segmentation in one embodiment of the present invention. Detailed Implementation

[0011] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0012] The present invention will be further described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of the invention.

[0013] As one embodiment of the present invention, such as Figure 1 As shown, a proposed method for retrieving and recommending treatments for complications in tube feeding care includes the following steps: S1. Multimodal nursing data acquisition and semantic processing: S1-1. Collect multimodal nursing data on tube feeding complications. First, deploy a data acquisition interface, focusing on collecting structured medical record data containing keywords such as "gastric retention," "diarrhea," "constipation," "aspiration," "reflux," "hyperglycemia," and "refeeding syndrome." Simultaneously acquire unstructured text data, including at least: nursing notes describing complications, textual descriptions of tube placement or lung infection in imaging reports, and relevant narrative data from consultation records. Next, collect interaction data from caregivers when querying tube feeding complications, specifically including: search terms entered in the search system (e.g., "what to do about abdominal distension after nasogastric feeding"), case IDs clicked to view, duration of stay on specific case pages, and behavior logs such as saving or annotation (e.g., annotating "gastric retention treatment plan"). Finally, send all collected multi-source heterogeneous data to a unified data buffer pool to complete the raw query data storage.

[0014] S1-2. Vectorize the original query data. First, preprocess the original query data (e.g., "severe diarrhea after nasogastric feeding") obtained from the data buffer pool. Use a Chinese word segmentation tool (e.g., Jieba segmenter or a BERT-based segmentation model) to segment the data, obtaining the word sequence {"nasogastric feeding", "after", "diarrhea", "severe"}. Next, call a custom nursing / medical stop word list to remove stop words without actual semantic meaning (e.g., "after"), and perform stemming or word form normalization on the remaining words (e.g., normalize "severe" to "heavy"), finally generating the purified word sequence {"nasogastric feeding", "diarrhea", "heavy"}. This purified word sequence will serve as the basis for the "effective vocabulary set" in subsequent steps.

[0015] Secondly, a pre-trained Clinical-BERT model is used to convert the processed word sequence into a fixed-dimensional query vector q (e.g., 768-dimensional) to numerically represent the core semantics of the query.

[0016] Finally, to quantify user behavior, a care behavior sequence M is constructed. int The sequence is set so that the rows represent session IDs and the columns represent different types of interactive events (such as keyword clicks, case viewing, and document retrieval). Whenever a user performs a specific action in a search session, the timestamp and frequency of the event are recorded at the corresponding position in the sequence.

[0017] S1-3. Generating medical feature points representing care needs through semantic alignment. This step aims to achieve semantic fusion and standardization of textual data on tube feeding complications (such as nursing notes and imaging reports) from different care needs and with varying expressions. By constructing a dynamic, iterative semantic mapping framework, enhanced weighted semantic similarity between heterogeneous text fragments is calculated to generate medical feature points that can represent specific complication patterns.

[0018] The specific implementation steps include: S1-31, System Real-time Monitoring Care Behavior Sequence M int Set a frequency change threshold or a recently triggered time window. When the frequency of the "keyword click" event type in a session (such as SessionID:S001) suddenly increases beyond the threshold, or its latest timestamp is within the time window, it is considered a newly triggered event (e.g., the user clicks the keywords "choking" and "SpO2 decrease"). It is determined that the user's interaction behavior is active. Then, the original text data description associated with the event (e.g., "The patient experienced severe choking after nasogastric feeding, and SpO2 dropped to 88%) is extracted and defined as the first fragment (i.e., the fragment containing dynamic unresolved care risk events).

[0019] Next, using the core medical terms from the first segment (such as "choking" and "decreased SpO2") as search keys, parallel matching queries are performed in the historical database to obtain the corresponding standardized information units. For example, retrieving "Responding to choking: Immediately stop tube feeding, place in right lateral decubitus position..." from the standard care protocol or matching "Chest X-ray shows patchy blurred shadows in the right lower lung field" from historical imaging reports, these statically structured knowledge units are defined as the second segment (i.e., the segment containing statically structured knowledge), providing a reference solution and care basis for the 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 care protocols and key information from historical case data).

[0020] Based on the above technical concept, it should be noted that texts from different sources (such as real-time nursing records and standard guidelines) vary greatly in their expression, level of detail, and terminology. For example, a single medical concept can have multiple expressions ("bloated stomach," "abdominal distension," "all referring to," "abdominal distension"), and this same expression can refer to different concepts in different contexts. Therefore, steps S1-32 are required to perform lexical analysis on the two segments, "translating" or "aligning" the words in the two segments into a unified conceptual space, thereby providing a data foundation for the subsequent accurate calculation of the overall semantic similarity between the two segments.

[0021] S1-32. First, receive the original text data of the first and second segments output from step S1-31, and perform independent natural language processing on the text of these two segments respectively, denoted as the effective vocabulary set V of the first segment. first (e.g., {"nasal feeding", "violent", "choking", "SpO2", "decline"}) and the second segment's effective vocabulary set V second (e.g., {"chest X-ray", "right lower lung field", "patchy", "blurred", "density shadow", "inhalation"}), thereby transforming data describing care behaviors into a set of terms representing their core medical semantics; Secondly, traverse the first segment's effective vocabulary set V. first For V first Each word in w fi Calculate its relationship with each standard concept c in the medical terminology semantic database. j The cosine similarity between the embedding vectors is calculated using the following formula: In the formula, For vectors The specific numerical value in the k-th dimension, the vector For the vocabulary w fiThe d-dimensional semantic vector representation (i.e., the embedding vector), generated by a pre-trained model (such as BioBERT or Clinical-BERT), is used to numerically capture the semantic information of the word. For vectors The specific numerical value in the k-th dimension, the vector For standard concept c j The d-dimensional semantic vector representation (i.e., the embedding vector) is also obtained by encoding the name or description of a standard concept through a pre-trained model, where k is the dimension index of the element in the vector.

[0022] Understandably, a medical terminology semantic database is a pre-generated, readable knowledge base containing authoritative and standardized care medical terms and their semantic relationships. Its vectors are obtained by training deep learning models (such as BioBERT and Clinical-BERT) on massive amounts of medical texts (such as textbooks, care medical guidelines, and high-quality medical records), and typically include medical concepts that characterize care behaviors, such as symptoms, signs, diseases, procedures, and medications.

[0023] At this point, for a word w fi This requires using an approximate nearest neighbor search algorithm to calculate the vector of the word and its relation to each standard concept c in the medical terminology semantic database. j The cosine similarity between vectors is calculated, resulting in N similarity values. The maximum cosine similarity among these N values ​​is then determined. If this maximum cosine similarity exceeds a preset vocabulary concept mapping threshold θ, the result is considered invalid. map (e.g., 0.75), then a unique word w is established. fi To the standard concept c j Map(w) fi )=c jmax This indicates that the term represented by the word has been successfully anchored to a standard concept. jmax To achieve the maximum similarity Sim max If the standard concept is found, then it indicates that a sufficiently similar corresponding concept was not found in the concept base, and it is marked as an unmapped word. fiu .

[0024] Finally, the same applies to V. second The same operation is performed on the words in the second segment, which will not be repeated here, to obtain the effective vocabulary set V of the second segment. second Each word in the text sk To the standard concept c j Map(w) sk ) and unmapped words w skuThis ensures that the vocabulary of each segment is bound to at most one standard concept, ultimately forming a one-to-one or zero mapping relationship, ensuring the fairness and consistency of aligning the vocabulary of two segments within the same standard concept space.

[0025] Based on the above technical concept, it is understandable that the purpose of obtaining mapped concepts is to transform the aforementioned vocabulary into standard concepts, ensuring the accuracy and consistency of the system's core semantics. For example, ensuring that "bloating" and "abdominal distension" both point to the unique "abdominal distention" is fundamental to building reliable medical knowledge. The purpose of obtaining unmapped vocabulary is to identify "marginal" words that cannot be easily standardized. It should be noted that these words often contain crucial information rich in care details. Separating them is to mark objects requiring special processing, thus enabling this system to be more intelligent than traditional dictionary-based systems, capable of understanding free and diverse care descriptions.

[0026] S1-33. By calculating the optimal semantic alignment path of all word pairs (mapped standard concepts and unmapped words) between the first and second segments, the overall semantic relevance of the two segments is quantified, thereby determining whether the dynamic care risk events to be resolved and the static structured knowledge units describe the same care need problem.

[0027] Specifically, firstly, based on the processing results of steps S1-32, a list of valid units L for the first segment is formed. first (Includes the standard concept Map (w) that has been mapped) fi ) and unmapped words w fiu ) and the list of valid units L of the second segment second (Includes Map(w) sk ) and unmapped words w sku ), load the predefined part-of-speech weighting strategy (noun weight ω) noun =1.0, adjective ω adj =0.8, verb ω verb =0.6), assigning words with different parts of speech to the above list of effective units to characterize their different importance in medical description.

[0028] Secondly, a unified high-dimensional semantic vector is generated for all valid unit lists to serve as the basis for computation. For mapped standard concepts, their pre-generated static semantic vectors are directly read from the medical terminology semantic database. For unmapped words, they are placed back into the context of their original segments (e.g., "the patient experienced severe choking after nasogastric feeding") and input into the pre-trained context-aware model (such as Clinical-BERT) in steps S1-2. The hidden state corresponding to the word position in the model's output layer is taken as its dynamic semantic vector. .

[0029] Next, construct an m x n semantic similarity matrix M between segments. sim Set the list of valid units L corresponding to the first segment for each row. first The list of valid units corresponding to the second segment L second For each element (u) in the matrix i ,v j ), by the first segment effective unit u i With the second segment effective unit v j The weighted cosine similarity between them is calculated using the following formula: In the formula, Used to calculate vectors and The standard cosine similarity between them, with a value range of [-1, 1]. They represent the elements based on unit u. i , and v j The weight coefficients assigned to the parts of speech are used to ensure that the matching of key medical entities (nouns) representing care needs contributes the most to the overall similarity, while the matching contribution of modifying words (adjectives, verbs) is relatively small. For example, the matching weight of two nouns is (1.0+1.0) / 2=1.0, while the matching weight of a noun and an adjective is (1.0+0.8) / 2=0.9.

[0030] Therefore, a dynamic programming algorithm is used in the semantic similarity matrix M. sim Find the optimal semantic alignment path with the highest overall alignment score.

[0031] It is understandable that the logic of the dynamic programming algorithm proposed in this embodiment is to decompose the optimal alignment into a series of interrelated sub-problems, which exist in the semantic similarity matrix M. sim The optimal score for each position [i][j] is calculated, and this is addressed step-by-step by filling a table, i.e., Malign, to avoid redundant calculations. The goal is to calculate the overall semantic relevance score S of the two segments. segment The specific process is as follows: The first step is to initialize a score matrix M. align With path backtracking matrix M path Meanwhile, its dimensions and semantic similarity matrix M are set. simConsistent, the number of rows equals the list of valid units L of the first segment. first The length m and the number of columns are equal to the list of valid units L of the second segment. second The length n.

[0032] Understandably, existing technologies based on keyword matching or local similarity calculation can only capture isolated, point-to-point word associations (e.g., only finding similarities between the words "choking" and "inhalation"), and cannot determine from a holistic perspective whether a dynamic, unresolved care risk event description such as "Patient experiences severe choking after nasogastric feeding, SpO2 decreases" and a static, structured knowledge unit description such as "Respond to choking: Immediately stop tube feeding...Be alert for aspiration pneumonia" express the same care risk (i.e., aspiration risk).

[0033] Based on this, the initialization score matrix M proposed in this invention align The goal is to use a dynamic programming algorithm to progressively fill in and record the cumulative optimal score among all possible alignment schemes from the beginning of a fragment to any position (i,j), thereby quantifying the overall semantic relevance from the start to the end.

[0034] Furthermore, the proposed initialization path backtracking matrix M path This is to accurately record the score matrix M align The optimal source direction for each cell's score (e.g., diagonal matching or inserting a gap) is determined so that, after calculation, the specific word alignment path that maximizes the overall alignment score can be constructed backtrackingly (e.g., the path includes the high-scoring match "cough-inhale," and any potential gap adjustments). The final score is based on the total path score S. segment Make judgments to achieve accurate semantic alignment and identification of care needs issues.

[0035] It should be noted that, in this embodiment, during the dynamic programming process, any position (i,j) refers to a position in the score matrix M. align and semantic similarity matrix M sim The same position in the matrix. In specific implementation, the score matrix M is initialized during the initialization phase. align [0][0] is set to 0, and the first row and first column are respectively accumulated with the gap penalty value (usually set to a fixed negative value, generally gap=-0.5, to handle the case of cells not being aligned).

[0036] The second step involves iterative calculation, row by row and column by column: filling the score matrix M row by row and column by column. align For each position (i,j), by calculating the scores in its three directions, we begin iterative calculation row by row (i from 1 to m) and column by column (j from 1 to n) to fill M. align Each cell M of the matrix align [i][j]: The diagonal score starting from position (i-1, j-1) can be understood as derived from aligning the i-th unit of the first segment with the j-th unit of the second segment, and its value is the upper diagonal position score M. align [i -1][j -1] plus the weighted cosine similarity M between the two. sim (i,j); The upward score starting from position (i-1,j) (where the unit of the first segment is aligned with an empty space) is derived from the fact that the i-th unit of the first segment is not aligned with any unit of the second segment, but rather with an empty space. Its value is the score M for the position directly above. align [i - 1][j] is increased by a gap penalty value; The score for the leftward movement starting from position (i,j-1), (where the unit of the second segment is aligned with an empty space), arises from the fact that the j-th unit of the second segment is not aligned with any unit of the first segment, but rather with an empty space. Its value is the left-side position score M. align [i][j-1] is added to the same empty space penalty value gap; Then, the maximum value of the three is taken as M. align The value of [i][j], that is, And in the path backtracking matrix M path Record the source direction in the middle; The third step, after completing the filling, is to analyze the scoring matrix M. align Starting from the bottom right corner [m][n], backtrack to the top left corner [0][0], collect all the unit pairs along the path with the maximum score, form the optimal semantic alignment path, and obtain the final score of this path: S segment =M align [m][n] represents the overall semantic similarity score between the two segments. It can be understood that its value directly depends on the semantic similarity matrix M. sim The accumulation and combination of all weighted cosine similarities in the path quantifies the overall semantic relevance of two segments. This is used to determine whether a dynamic, unresolved care risk event and a static, structured knowledge unit description describe the same care problem. For example, if the path contains high-scoring matching pairs such as "choking" and "inhalation" and the overall score is high, it indicates that the two are highly correlated. In this case, the word pairs recorded on the path (such as "choking" in the first segment and "inhalation" in the second segment) are the key semantic alignment pairs that contribute the most.

[0037] The fourth step is to set a high segment similarity threshold θ. high (e.g., 0.85), iterate through all unprocessed (first segment, second segment) pairs, only if S segment ≥θ highOnly then is the key semantic alignment pair identified as a "high-match semantic pair" and filtered out. Generally speaking, if the core medical entities (nouns) of two segments are successfully matched and have high scores (weighted at 1.0), the segment similarity threshold θ is easily reached. high .

[0038] For text data pairs that do not reach this threshold, the following feedback strategy is used: analyze the keyword pairs in their optimal semantic alignment path. If there are high-frequency co-occurrence (co-occurrence times ≥ 5 times) but unmapped word pairs (e.g., "gastric residual amount" and "gastric emptying disorder" are not mapped, but they co-occur in multiple text pairs and have high alignment scores (their average alignment scores exceed the set threshold (≥ 0.7))), then they are considered as new concept candidates. These new concept candidates are returned to the construction and updating process of the medical terminology semantic base to achieve continuous iterative optimization of the knowledge base.

[0039] The fifth step, to aggregate the scattered "highly matched semantic pairs" text fragments selected in the fourth step (e.g., "patient coughing with decreased SpO2" and "chest X-ray suggests aspiration pneumonia") into a unified concept with clear nursing significance, involves generating and labeling medical feature points: First, all text fragments (whether the first or second fragment) in the selected "highly matched semantic pairs" are converted into fixed-dimensional sentence-level semantic vectors by the sentence encoder (Sentence-BERT) to form a high-quality semantic vector pool.

[0040] Secondly, the DBSCAN density clustering algorithm is applied (assuming the neighborhood radius parameter eps = 0.6 and the minimum number of samples min). samples =5) Perform cluster analysis on the vector pool to automatically discover semantically dense clusters from the data's own distribution. Each identified dense cluster is defined as a medical feature point representing a specific complication pattern. Its core vectorized representation is determined by the centroid (mean vector) of all sentence vectors within the cluster. It should be noted that the centroid vector proposed in this embodiment aims to provide a precise and computable numerical definition for abstract "medical concepts." For example, when a patient queries "What to do if breathing is poor after nasogastric feeding," the system can also convert the query statement into a vector and then quickly find the medical feature point (such as "aspiration pneumonia" or "atelectasis") that is "closest" to it in the semantic space, thereby achieving accurate recommendations.

[0041] Furthermore, rich semantic descriptions are generated for each medical feature point: First, the cosine similarity between its centroid vector and all concept vectors in the medical terminology semantic database is calculated, and the standard concept with the highest similarity (such as "aspiration pneumonia") is determined as the main label; Second, all texts within the cluster are segmented and filtered by part of speech, nouns and gerunds are extracted and their frequencies are counted, and the most frequent terms (such as "choking" and "aspiration") are used as auxiliary descriptive labels.

[0042] Finally, all these medical feature points are combined to form the medical feature set FP. i In this embodiment, the following feature point data is included: {FP_001:{"Main Label":"Aspiration Pneumonia","Auxiliary Description Label":["Coughing","Aspiration","Nasal Feeding","Decrease in SpO2"],"Centroid Vector":[0.12,-0.05,...,0.78] / / 768-dimensional floating-point array}; FP_002:{"Main Label":"Gastric Retention","Auxiliary Description Label":["Abdominal Distension","Gastric Residue","Nasal Feeding","Discontinuation of Feeding"],"Centroid Vector":[-0.23,0.15,...,-0.42] / / 768-dimensional floating-point array}}.

[0043] For ease of understanding, as shown in Table 1 below, the medical feature set FP can be intuitively obtained. i Provide a detailed description of each generated medical feature point.

[0044] Table 1:

[0045] S2. Construct a knowledge graph of complications to integrate the medical feature set FP generated in stage S1. i A multi-layered, dynamically evolving knowledge graph is constructed, enabling the graph to possess care reasoning capabilities.

[0046] The specific implementation steps include: S2-1, Based on semantic similarity matrix M sim An initial atlas is constructed, thereby transforming discrete medical features into an atlas with a basic topological structure.

[0047] The specific process is as follows: using the generated medical feature set FP i As graph nodes, each node represents a care complication feature (such as "aspiration pneumonia" or "gastric retention"). Node attributes include its primary label, auxiliary descriptive labels, and semantic information. The cosine similarity score of each feature point in the semantic space is calculated. Based on a preset initial edge threshold, a connection is established. For node pairs exceeding the threshold, undirected edges are created, and the similarity value is used as the initial weight of the edge, thus forming an initial, undirected, weighted graph G. inThis initially reveals a clear association between complications.

[0048] S2-2, Introduce a multi-head graph attention network (GAT) to process the initial graph G. in Structural optimization is performed to form an optimized graph Gop with clearer structure and richer semantics. The specific process utilizes the GATConv layer provided by existing graph neural network libraries (such as PyTorchGeometric). First, the initial map G in The semantic features of each node (composed of the centroid vectors of each feature point; understandably, each medical feature point (e.g., FP_001) becomes a node in the knowledge graph, and its centroid vector is the core attribute of this node) are input into the GAT network. A ready-made GATConv layer (with configurable heads=8) is used for linear transformation and feature mapping, projecting the (e.g., 768-dimensional) features onto a low-dimensional semantic subspace (e.g., 128-dimensional), resulting in the transformed node feature representation. Simultaneously, within this layer, the LeakyReLU activation function (with a default negative slope of 0.2) is used 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, and || denotes vector concatenation. These are the original feature vectors (i.e., centroid vectors) of nodes i and j, respectively.

[0049] Secondly, softmax normalization is performed on the attention scores on all incident edges of each node to obtain the normalized attention weights αij; Finally, the features of neighboring nodes are weighted and summed according to these weights, and the new feature representation of the nodes is output through the ELU activation function. It is understandable that the entire optimization process is implemented based on the standard forward propagation and parameter update mechanism of existing Graph Attention Networks (GAT), without the need for zero-based coding. The optimized edge weights are derived from the attention coefficients and can be sparsified according to a set threshold to remove weakly connected edges, ultimately yielding the optimized graph G. op This significantly improves the care interpretability and structural quality of the map.

[0050] S2-3: Obtaining the optimized map G op Subsequently, complication communities with tightly connected internal structures and sparse external structures were identified.

[0051] The specific implementation steps include: using the Louvain algorithm based on modularity optimization to iteratively divide communities by maximizing the modularity Q-value of the entire graph attention network (GAT).

[0052] Understandably, modularity (Q) is a core indicator for measuring the quality of community segmentation, and its calculation formula is as follows: In the formula, To optimize the map G op Let ki be the weight of the edge between nodes i and j, and ki be the sum of the weights of all edges connected to node i. To optimize the map G op The sum of the weights of all edges in the equation is half of the total weights, where ci is the community to which node i belongs. The function is the Kronecker function, which is 1 when two nodes belong to the same community (i.e., ci=cj) and 0 otherwise.

[0053] In this embodiment, when performing community segmentation, the optimized graph G is first... op Each node is treated as an independent community. Then, each node i is traversed, and the modularity gain ΔQ brought by moving it to each neighboring community is calculated. The community that can bring the maximum positive gain ΔQ is selected, and node i is moved to that community. If no movement can produce a positive gain, the node remains in the original community. Then, nodes belonging to the same community are aggregated into a new node, and the above process is repeated on the new graph until convergence.

[0054] Thus, as Figure 2 As shown, the output provides a stable community partitioning, where each community represents a complication community with care guidance significance, composed of semantic relationships optimized by GAT (such as "reflux aspiration cluster"). This enables the automated discovery and quantification of complex semantic relationships of complications from the data.

[0055] For example, community C1: {vomiting, regurgitation, head-of-bed elevation, aspiration, aspiration pneumonia, tachypnea, decreased SpO2}; the care interpretation of this community is that during tube feeding, "insufficient head-of-bed elevation" may lead to "regurgitation" and "aspiration," which in turn causes "aspiration pneumonia." The care manifestation is a highly correlated care pathway of "tachypnea" and "decreased SpO2." In this case, the community is named "Reflux, Aspiration, and Related Complications Cluster."

[0056] Community C2: {Hyperglycemia, rapid tube feeding rate, insulin use, polyuria, dehydration}; Care interpretation: This community reflects hyperglycemia caused by "rapid tube feeding rate," which necessitates "insulin use," accompanied by complications such as "polyuria" and "dehydration." This community is then named the "Glucose Metabolism Disorder Complication Cluster."

[0057] Finally, the aforementioned community information (i.e., which community each node belongs to) is used as a new attribute to feed back into and update the optimized graph G. op In this process, a knowledge graph G is ultimately formed that contains semantic tags for the community. opfinal .

[0058] S3. Generate a refined search area: It should be noted that in the construction of complication knowledge graphs, each complication topic community usually consists of dozens of highly related medical nodes, covering a complete care recommendation path from etiology, symptoms, signs to intervention measures. If a search is performed directly across the entire graph, not only will the computational efficiency be low, but it may also return a large amount of irrelevant information, affecting the accuracy and response speed of care decisions.

[0059] For ease of understanding, based on the community segmentation results of step S3, taking the "reflux aspiration and related complications cluster" (community C1) as an example, in addition to the main pathways already mentioned in this community, there are several highly related care pathways with different focuses as shown in Table 2 below. These pathways share the same core complication theme, but are triggered by different initial factors or exhibit different combinations of symptoms, ultimately leading to similar serious complications.

[0060] Table 2:

[0061] Based on this, step S3 is proposed to avoid inefficient full-graph search across the entire massive knowledge graph.

[0062] The specific implementation steps include: S3-1, Knowledge graph G containing community semantic tags opfinal Aggregate analysis is performed to identify communities on the topic of complications that users are most likely to be interested in. In this embodiment, the analysis preferably considers the following two aspects: The intrinsic importance of communities (graph structure level) is quantified by calculating the average attention weight of nodes within a community to assess the tightness of semantic connections and overall importance within the community. The higher the weight, the stronger and more critical the internal connections of the care concept represented by the community. Quantify the relevance between the community and the user (at the user intent level). By comparing the community topic with the user's historical query vector and interaction behavior logs, determine the degree of matching between the community content and the user's real care needs, and ensure that the recommendation results are highly relevant to the user's interests.

[0063] Subsequently, by calculating each community C k The overall level of attention is used as a quantitative basis for ranking and selecting core communities. The formula for calculating the overall level of 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 graph represents a medical feature point (such as "aspiration pneumonia" or "gastric retention") generated after step S1. It is the most basic semantic unit in the knowledge graph. The similarity between the semantics of the query vector q in steps S1-2 and the community topic is calculated, typically using an indicator function. This function returns 1 when the community topic matches the user's historical query intent. This is used to calculate the average attention weight of all nodes within a community. The higher the average value, the stronger the connections within the community and the more important it is as a whole. denoted as the attention weight of node ni, which originates from the optimized output of the Graph Attention Network (GAT) in step S2-2. It is used to numerically represent the node's influence and importance within the context of its neighboring nodes; the higher the weight, the more critical the node. This is a smoothing reward for large-scale communities, designed to prevent excessively large communities from having an absolute advantage. This item only takes effect when the community topic matches the user's query intent, and gives a larger reward to large-scale communities (taking the natural logarithm). This allows for the priority recommendation of high-value communities that are both large-scale and relevant in the ranking.

[0064] Example: If a user's historical query intent contains keywords such as "gastric retention" and "bloating," and the query vector q generated in step S1-2 matches the topic of community C3 (community topic is "gastric retention and related complications"), then... If community C3 contains 10 nodes and its average attention weight is 0.8, then the overall attention of community C3 is Cl(C3)=0.8+ln(10)×1=0.8+2.3=3.1.

[0065] S3-2. Select the top M=3 communities with the highest Cl scores. These communities represent the topic communities that the system believes users are most likely to be interested in regarding complications. Extract all nodes from these communities to form a basic node set, and then calculate the convex hull boundary of this basic node set: In the formula, This is a convex hull function, and its output is the convex hull H corresponding to the point set. This represents the coordinates of node ni in a certain vector space. Generally, this space is usually a low-dimensional semantic space (e.g., a 128-dimensional space to which the node features are projected by the GAT network in step S2-2, the purpose of which in this embodiment is to visualize the computation of the convex hull). This is the union operation for sets.

[0066] S3-3. Using the above formula, aggregate all nodes ni from the top 3 communities with the highest overall attention, and calculate the coordinates of these nodes in the semantic space. The convex hull H is defined by the convex hull, and the convex hull region Rc is formed by the area defined by the convex hull. It is understandable that, due to the overall attention level Cl(C... kThe convex hull region (Rc) is an aggregation of all nodes from the top M complication-related topic communities (such as the "reflux aspiration cluster" and the "gastrointestinal symptoms cluster"). It integrates the inherent importance of the communities at the graph structure level (attention weights optimized based on GAT) with the query relevance at the user intent level, ensuring that the nodes within Rc are both key data in the graph and highly matched to the user's real-time care needs. Therefore, the convex hull region Rc fully covers the core medical features most relevant to the current search session and can be identified as the final refined search area, encompassing all nodes from the three communities.

[0067] S4. Output the final search results: For each medical feature point node ni within the convex hull region Rc, it is sorted in descending order based on its attention weight in the knowledge graph. After sorting, the standardized medical concept (such as "gastric retention") corresponding to the highest-ranked / top-level node is used as the core search key. The associated multi-source knowledge base is queried through a preset data interface to generate and output specific recommended content. For example, when querying "management of abdominal distension after nasogastric feeding", the search result is "Recommended to refer to the 'Gastric Retention Management' chapter and related nursing care procedures in the 'Guidelines for Prevention and Treatment of Complications of Tube Feeding'".

[0068] As a second aspect of the present invention, a retrieval and recommendation processing system for tube feeding care complications is proposed, comprising a memory and a processor, wherein the memory includes a retrieval and recommendation processing program for tube feeding care complications, and when the processor executes the retrieval and recommendation processing program for tube feeding care complications, it implements a method for retrieval and recommendation processing of tube feeding care complications.

[0069] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

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: by initializing the score matrix and the path backtracking matrix, the maximum cumulative score of each unit in the score matrix based on word matching and gap insertion is iteratively calculated, and the similarity score reflecting the overall semantic relevance of the two segments is obtained by backtracking the path. Finally, a threshold is set to filter out highly matched semantic pairs. 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 medical feature set, and communities are optimized and divided. 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: The specific steps for using dynamic programming to calculate the optimal semantic alignment path between two segments and to select highly matching semantic pairs are as follows: Initialize a semantic similarity matrix M. 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 to the accumulated empty space penalty value; Iteratively calculate each cell (i,j) in the score matrix Malign, where i and j start from 1. Based on the maximum cumulative score under the word matching and gap insertion cases, take the maximum value as Malign. 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.

5. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 4, 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.

6. 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.

7. 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.

8. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 6, 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, and || denotes 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 .

9. The method for retrieving and recommending treatments for complications in tube feeding care according to claim 1, characterized in that: After optimizing and segmenting communities based on the constructed knowledge graph, and before generating the final recommended content by sorting according to the node weights of the knowledge graph, 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 search. 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.

10. 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 9.

Images