CN-121983297-A - Clinical care data management method and system for clinical big data

CN121983297ACN 121983297 ACN121983297 ACN 121983297ACN-121983297-A

Abstract

The invention relates to the technical field of medical information, and discloses a clinical care data management method and system for clinical big data. The method comprises the steps of obtaining structured physiological sign and unstructured nursing text data, extracting weak semantic signals through a deep semantic mining model, converting the weak semantic signals into high-dimensional feature vectors, mapping features to medical knowledge maps, constructing a hidden nursing risk dynamic deduction network, predicting target risk event probability by utilizing a graph neural network and Bayesian reasoning, and generating hierarchical early warning and intervention suggestions based on probability distribution. The system comprises a data acquisition module, a deep semantic extraction engine, a risk deduction network construction module, a prediction analysis engine and an early warning intervention generation module. According to the invention, the prospective identification and active intervention of the hidden risk are realized by fusing the NLP and the knowledge graph, and the early warning timeliness and the nursing decision accuracy are improved.

Inventors

ZHANG YAN
SHEN QIANG

Assignees

铜川市人民医院

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. A method for clinical care data management for clinical big data, comprising the steps of: Step S1, acquiring multi-source clinical care data in real time through a data acquisition module, wherein the multi-source clinical care data comprises structured physiological sign data read by a medical equipment networking interface and unstructured clinical care text records synchronously acquired from an electronic care document system, adding time stamps to the structured physiological sign data and the unstructured clinical care text records, and executing alignment operation based on a time axis in a data cache area; Step S2, performing deep processing on the unstructured clinical care text record through a preset deep semantic mining model, specifically, filtering and cleaning an original text to remove redundant characters, performing professional word segmentation processing aiming at the medical field, calculating association weights among word segmentation units in the text by utilizing a coding base with a multi-layer self-attention mechanism, capturing long-distance semantic dependencies, extracting global semantic representation, mapping implicit behavior features in the text to a unified high-dimensional embedded space through a feature alignment algorithm, and converting the implicit behavior features into high-dimensional semantic feature vectors representing the nursing record semantics; Step S3, mapping the high-dimensional semantic feature vector into a pre-constructed medical knowledge graph, and identifying a logic corresponding relation between the implicit behavior feature and a known pathological state by searching clinical pathological entities of basic diseases, past medical history and current diagnosis of a patient to construct an implicit care risk dynamic deduction network taking the patient as a center, wherein the implicit care risk dynamic deduction network comprises patient entity nodes, behavior feature nodes, pathological state nodes and directed weight edges for representing correlation intensity among elements; Step S4, based on the hidden care risk dynamic deduction network, performing multi-round information transfer and aggregation operation by using a graph neural network, enabling each node in the network to sense risk signals in the neighborhood of the node, inputting aggregated global graph characteristics into a Bayesian probability inference model to perform state evolution deduction by combining the accumulated effect of the risk signals captured by using a time sequence algorithm, and predicting probability distribution of a target care risk event of a patient in a preset time period; And S5, determining a risk level according to the probability distribution and a preset risk gradient threshold value, automatically generating a real-time early warning instruction, and screening out clinical care intervention suggestions aiming at the current hidden risk point by matching the current risk evolution path with a standard care rule based on a preset clinical care path library.
2. The method for clinical care data management for clinical big data according to claim 1, wherein the step S2 specifically comprises: using a depth bidirectional encoder representation model or a large language model based on a transformer as the encoding base, and carrying out multi-layer self-attention mechanism calculation on the standardized text; extracting the global semantic representation from an output layer of the coding base, and ensuring the stability of feature expression by using a residual error connection technology and a layer normalization technology configured in the extraction process; The feature alignment algorithm adopts a mapping strategy based on comparison learning, draws the Euclidean distance in the high-dimensional embedded space of nursing records which have the same clinical meaning and different expressions in a training stage, and pushes away the nursing records which are similar in word but different in meaning in a feature space; Before executing the step S2, the method further comprises the step of performing incremental training on the deep semantic mining model, introducing understanding capability of a clinical care labeling corpus optimization model on medical vocabulary and spoken language expression, introducing an countermeasure training strategy in the training process, and adding random disturbance in an input text to enhance the feature extraction robustness of the model in the face of nonstandard input.
3. The method for clinical care data management for clinical big data according to claim 2, wherein said step S3 specifically comprises: Retrieving clinical pathology entities associated with a patient from the medical knowledge map, the medical knowledge map encompassing clinical protection, anatomy, pharmacology, and complication evolution logic and stored in a map database in the form of a resource description framework or attribute map; Identifying a logical correspondence between the implicit behavioral characteristics extracted from weak semantic signals in descriptive language, the weak semantic signals including abnormal turn-over frequency, language aggressiveness tendency, drug refusal behavior, mood swings characteristics and atypical conscious state descriptions of the patient in a time period; The multi-mode fusion technology is adopted to complement the structured physiological sign data as the attribute information of the nodes in the hidden care risk dynamic deduction network, so that each node contains static pathological features and dynamic physical sign features, and the multi-mode fusion technology adopts a fusion strategy guided by an attention mechanism, and determines the contribution degree of each data source under different risk scenes by calculating the mutual weight between the structured physiological sign data and the unstructured clinical care text record.
4. The method for clinical care data management for clinical big data according to claim 3, wherein said step S4 specifically comprises: Executing the multi-round information transfer and aggregation operation on the hidden care risk dynamic deduction network, so that each node performs weighted summation according to the weight of the directed weight edge by collecting risk signals transmitted by the neighborhood nodes to update the state characteristics of the node; capturing the accumulated effect of the risk signals along with time by using a cyclic neural network or a time convolution network, and identifying a deterioration signal with trend by comprehensively modeling the sign fluctuation and the text record in the history period; Inputting the aggregated global graph characteristics into the Bayesian probabilistic reasoning model, and obtaining the possibility of occurrence of the target nursing risk event by utilizing a posterior probability calculation mode, wherein the target nursing risk event comprises an unplanned tube drawing, a falling bed, pressure injury, an injury event caused by hidden surging behaviors and sudden disease deterioration caused by consciousness disturbance.
5. The method for clinical care data management for clinical big data according to claim 4, wherein said step S5 specifically comprises: Comparing the calculated probability distribution with a plurality of preset risk gradient thresholds to determine the current risk level, triggering alarm signals with corresponding intensities according to the risk level, and pushing the alarm signals to a mobile nursing terminal or a central monitoring system in real time; The generation of the clinical care intervention advice is based on multi-criterion decision logic in the clinical care path library, and the optimal intervention measure combination is screened out from the drug intervention advice, the physical constraint advice, the environment optimization advice and the psychological dispersion key advice through a matching engine according to key incentive nodes in the current risk evolution path; The clinical care path library is provided with an external knowledge source verification interface which is used for automatically and synchronously updating weight distribution in the library according to release of external care guidelines, supporting long-term tracking of intervention effects, and automatically optimizing the sorting priority of the clinical care intervention suggestions by comparing the risk probability reduction amplitude after the intervention suggestions are executed.
6. The method of clinical care data management for clinical big data according to claim 5, wherein the unstructured clinical care text records include nurse shift records, care diaries, patient complaint records, and care observation descriptions; The structured physiological sign data includes heart rate, respiratory rate, blood oxygen saturation, blood pressure, body temperature, and laboratory examination indicators.
7. The method for clinical care data management for clinical big data according to claim 6, wherein the implicit care risk dynamic deduction network is configured with a self-correction function, specifically: When the risk result predicted by the system is inconsistent with the clinical result actually occurring, triggering a feedback mechanism and generating a negative feedback signal to store in a database; After accumulation of a preset period, the system automatically triggers a weight adjustment program, and dynamic balance of missing report and false report is realized by adjusting weight coefficients of corresponding edges in the hidden nursing risk dynamic deduction network.
8. The clinical care data management method for clinical big data according to claim 7, comprising: The data acquisition module is configured to synchronously acquire structured physiological sign data and unstructured clinical care text records in a clinical big data environment, and comprises a distributed data acquisition unit and a data protocol analyzer, and is used for converting original code streams output by different monitoring devices into standard format data and executing data alignment based on a time axis; The deep semantic extraction engine is configured to analyze the latent semantic features from the unstructured clinical care text records by using a natural language processing algorithm, integrates a pre-trained language model and an instruction fine adjustment layer aiming at the medical field, and realizes high concurrence processing of large-scale texts by using a distributed reasoning technology; The risk deduction network construction module is configured to combine the latent semantic features with the medical knowledge graph to generate a dynamic risk evolution topological structure, and comprises a real-time graph database interface for supporting topology inquiry and dynamic node update of the medical knowledge graph comprising nodes and edges; The prediction analysis engine is configured to deduce the risk evolution topological structure by utilizing graph calculation and a probability model to obtain risk probability distribution, adopts a streaming calculation frame and is configured with a task priority scheduling algorithm, and the prediction analysis engine is used for preferentially ensuring a risk deduction task of a high-risk state patient when calculation resources are limited; The early warning and intervention generation module is configured to output multidimensional risk early warning information and clinical intervention schemes, and comprises an interactive interface generation unit which is used for displaying a risk evolution path diagram on terminal equipment in a visual mode.
9. The method for clinical care data management for clinical big data according to claim 8, wherein the coding base in the deep semantic extraction engine adopts a deep neural network with a residual structure, and the instruction fine tuning layer learns semantic relations under medical context by self-supervised learning of real care records; The real-time graph database interface supports distributed storage and fragmented query so as to meet the requirement of concurrent access of a large-scale disease area.
10. A clinical care data management system for clinical big data, characterized in that the clinical care data management is realized using the clinical care data management method for clinical big data according to any one of claims 1 to 9.

Description

Clinical care data management method and system for clinical big data Technical Field The invention belongs to the technical field of medical information, and particularly relates to a clinical care data management method and system for clinical big data. Background With the rapid development of the smart medical and medical big data industry, deep mining and refined management of clinical data have become core driving forces for improving medical service quality and guaranteeing patient safety. The clinical big data not only covers basic physiological sign data of a patient, but also comprises massive medical operation records and nursing behavior information, and scientific support can be provided for clinical decision and medical resource allocation can be optimized through the integration and utilization of the complex data. Clinical care data management is a core link in medical informatization construction, and aims to collect, store and analyze various information in a care process through informatization means so as to realize dynamic monitoring of care quality. The nursing management system can reflect the evolution trend of the patient's illness state in real time through the digital processing of the patient sign data and the clinical nursing records, and provides accurate risk early warning and intervention guidance advice for clinical nursing staff. The traditional clinical care data management technology still faces a plurality of technical problems in practical application, namely, data processing capability is limited to structured information, threshold monitoring on vital sign values is mainly relied on, semantic mining capability of a large amount of unstructured text information such as nurse shift records and nursing diaries is lost, so that hidden key information in a clinical care process is in a deep sleep state and cannot be converted into decision basis, a risk early warning mechanism presents post-reaction characteristics, an alarm is usually triggered after a clear abnormality occurs in vital signs of patients, deep deduction on weak semantic signals in nursing records is lacked, potential medical risks in a sprouting state are difficult to capture, and deep fusion of knowledge maps in a clinical pathology field is lacked due to single data analysis dimension, so that a system cannot construct an entity association network aiming at complex nursing behaviors and pathology signs, and accuracy and breadth of risk identification are limited. The problems above together lead to serious hysteresis of nursing risk early warning and high rate of missing report, and cannot meet the core requirements of modern intelligent medical treatment on high-precision and prospective nursing safety management. Disclosure of Invention The invention aims to provide a clinical care data management method and a system for clinical big data, which can solve the problems in the background technology. The invention aims to solve the problems of deep sleep, decision basis deficiency of clinical information caused by incapability of processing unstructured text data and high risk early warning lag and risk missing report rate caused by the fact that a traditional early warning mechanism only depends on structured sign data in the traditional clinical care management process, and realize the technical crossing from simple data recording to deep risk pre-judgment. In order to achieve the purpose, the technical scheme adopted by the invention is that the clinical care data management method for clinical big data comprises the following steps: S1, acquiring multi-source clinical care data in real time, wherein the multi-source clinical care data comprises structured physiological sign data and unstructured clinical care text records; s2, processing unstructured clinical care text records through a preset deep semantic mining model, extracting weak semantic signals in descriptive language and converting the weak semantic signals into high-dimensional semantic feature vectors; S3, mapping the high-dimensional semantic feature vector into a pre-constructed medical knowledge graph, and constructing a hidden nursing risk dynamic deduction network through entity identification and relation extraction; S4, based on the hidden nursing risk dynamic deduction network, carrying out state evolution deduction by using a graph neural network and a Bayesian probability inference model, and predicting probability distribution of a target nursing risk event of a patient in a preset time period; And S5, automatically generating a real-time early warning instruction and a targeted clinical nursing intervention suggestion according to the probability distribution and a preset risk threshold. Preferably, the step S2 specifically includes the following steps: S21, performing text cleaning and standardization processing on unstructured clinical care text records, removing redundant characters irrelevant to care behaviors, and executin