Elderly patient perioperative anesthesia risk management method based on multi-modal knowledge graph

Abstract

The invention relates to the technical field of medical information, and discloses a perioperative anesthesia risk management method for an elderly patient based on a multi-modal knowledge graph, and the method comprises the following steps: S1, constructing a multi-modal knowledge graph of perioperative anesthesia risks of the elderly patient; s2, acquiring intraoperative multi-modal physiological time sequence data of the target elderly patient in real time; s3, performing dynamic risk monitoring based on the multi-modal knowledge graph and the data in the S2, and judging whether an abnormal risk event exists or not; s4, if yes, generating a personalized intervention decision based on a deep reinforcement learning DRL framework, and outputting an optimal intervention action; and S5, according to the clinical outcome feedback after the execution of the intervention action, dynamically adjusting the weight of the corresponding edge in the multi-modal knowledge graph, and realizing the closed-loop collaborative optimization of the knowledge graph and the decision model. According to the method, the accuracy of preoperative risk prediction for the elderly patients can be improved, the time for finding abnormal events in the operation is shortened, the matching degree of personalized intervention strategies is improved, and the decision load of doctors is reduced.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to a method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graphs. Background Technology

[0002] With the increasing aging of the population, perioperative anesthesia management for elderly patients (≥65 years old) faces unprecedented challenges. Due to declining physiological functions and the coexistence of multiple diseases, elderly patients have significantly reduced tolerance to anesthesia and surgery. Although the traditional perioperative anesthesia management system has formed standardized procedures, its technical content remains at the stage of experience-driven and single-modal data-centric. In the preoperative assessment stage, clinicians generally rely on the American Society of Anesthesiologists (ASA) classification standard, which classifies risk levels by simply listing comorbidities such as hypertension and diabetes. However, this discrete stratification model is difficult to quantify the synergistic pathogenic effects of chronic diseases, and even more so, it cannot dynamically track the potential correlation between preoperative medication history and intraoperative hemodynamics. Although intraoperative monitoring is equipped with vital sign monitors and vasoactive drug infusion systems, its "three lows" warning criteria (hypotension, low BIS value, and low MAC value) are based on fixed threshold settings and lack dynamic correlation with preoperative comorbidity characteristics, making it difficult for the warning system to trigger intervention measures based on individualized risk thresholds. Postoperative management often falls into a passive observation trap, with complication diagnosis heavily reliant on physician experience. Metabolic abnormalities such as hypokalemia are often detected 24-72 hours late, missing the optimal intervention window. This technological lag is particularly fatal in elderly patients, whose fragile physiological state combined with the complexity of multiple coexisting diseases creates a cumulative effect, necessitating a breakthrough from the limitations of the traditional system.

[0003] More seriously, the existing system suffers from fundamental flaws in data analysis. Existing risk assessment models (such as prediction algorithms based on logistic regression) are limited by linear assumptions and the sparsity of high-dimensional features, making them unable to analyze multi-factor nonlinear relationships and complex causal paths. Although some systems have introduced LSTM time-series models, their real-time processing of high-frequency physiological signals still suffers from significant delays (typically >5 seconds), making it difficult to meet the needs of dynamic intraoperative early warning. At the decision support level, Clinical Decision Support Systems (CDSS) generally adopt a static rule-based model, with knowledge update cycles as long as 3-6 months, making it difficult to incorporate the latest evidence-based medicine in real time, and lacking dynamic anomaly detection and feedback capabilities, thus failing to identify high-risk conditions such as hemodynamic decompensation in real time during surgery. A deeper contradiction lies in the fragmented design of the system architecture. There is a serious semantic gap between preoperative electronic medical records, intraoperative monitors, and postoperative PACS systems, resulting in long data retrieval times and delays in identifying critical conditions. The lack of real-time data stream processing capabilities further highlights the technological gap; multi-dimensional data collected by new sensors, such as respiratory rate and cough frequency, cannot be integrated into existing platforms, forming information silos. This technological lag means that perioperative management always faces three core challenges: the inability to deeply integrate multimodal data, the inability to quantify dynamic risks in real time, and the lack of closed-loop verification for individualized decisions.

[0004] Therefore, there is an urgent need to establish a dynamic technology system that integrates multimodal data modeling to solve the above problems. Summary of the Invention

[0005] The present invention aims to provide a perioperative anesthesia risk management method for elderly patients based on multimodal knowledge graphs, which can improve the accuracy of preoperative risk prediction for elderly patients, shorten the detection time of intraoperative abnormal events, improve the matching degree of personalized intervention strategies, and reduce the decision-making load of physicians.

[0006] The basic solution provided by this invention is: a method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graphs, including the following steps: S1. Construct a multimodal knowledge graph of perioperative anesthesia risk for elderly patients. The entities in the multimodal knowledge graph include at least preoperative comorbidities, intraoperative physiology and operation, and postoperative complications. The edges between entities are used to represent the medical association between entities, and the association strength is quantified by weight. S2. Real-time acquisition of intraoperative multimodal physiological time-series data of target elderly patients; S3. Based on the multimodal knowledge graph and the real-time data acquired in S2, perform dynamic risk monitoring: use a time-series prediction model to process the physiological time-series data to obtain predicted values ​​of physiological parameters; combine the entity association paths related to the target patient in the multimodal knowledge graph to dynamically determine whether there are any abnormal risk events. S4. When an abnormal risk event is determined to exist, a personalized intervention decision is generated based on the deep reinforcement learning (DRL) framework: the risk score and real-time physiological parameters, which are integrated with the association weights of the multimodal knowledge graph, are used as the state input of the DRL model, and the optimal intervention action is output from the predefined intervention action space. S5. Based on the clinical outcome feedback after the intervention, dynamically adjust the weights of the corresponding edges in the multimodal knowledge graph to achieve closed-loop collaborative optimization between the knowledge graph and the decision-making model.

[0007] The working principle and advantages of this invention are as follows: This invention presents a perioperative anesthesia risk management method for elderly patients based on multimodal knowledge graphs. It constructs a dynamic technical system integrating multimodal data modeling, which can improve the accuracy of preoperative risk prediction for elderly patients, shorten the detection time of intraoperative abnormal events, improve the matching degree of personalized intervention strategies, and reduce the decision-making burden on physicians. The key points are: First, the accuracy of preoperative risk prediction has improved.

[0008] This approach integrates patients' preoperative baseline data, medical history, and comorbidities using a multimodal knowledge graph, and combines this with deep learning algorithms to quantitatively assess risk. Compared to the traditional ASA grading method, this approach more comprehensively considers individual patient differences and potential risk factors, significantly improving the accuracy of preoperative risk prediction. This provides a more scientific basis for preoperative anesthesia planning and can reduce intraoperative and postoperative risks.

[0009] What's particularly noteworthy here is that this approach not only structurally integrates massive amounts of heterogeneous knowledge from guidelines, literature, and medical records, but more importantly, transforms it into computable and reasonable network relationships. This allows each preoperative comorbidity and intraoperative parameter fluctuation of a patient to be evaluated within a complex causal and relational network. Through the flexible definition of entities and relationships, this approach can characterize the synergistic risk effects of multiple co-existing diseases in elderly patients and the unique risk pathways brought about by physiological functional decline, providing a solid prior knowledge foundation and rich contextual information for subsequent real-time analysis.

[0010] Second, the time for detecting intraoperative abnormal events is shortened.

[0011] This protocol incorporates a real-time anomaly detection step, utilizing a time-series prediction model to monitor and analyze multi-parameter physiological signals during surgery in real time. Through a dynamic early warning system, alerts can be issued promptly at the initial stage of abnormal events, significantly shortening the detection time for intraoperative anomalies. This allows medical staff to take intervention measures earlier, reducing the incidence of serious complications and improving the safety and success rate of the surgery.

[0012] What's unique here is that this approach doesn't run a predictive model or decision engine in isolation. Instead, it drives the knowledge graph to perform instant retrieval and reasoning through real-time data streams, dynamically activating the risk path subgraphs most relevant to the current patient. Subsequently, the attention mechanism of the time-series predictive model (such as LSTM) is adjusted by the weights of these graph paths, allowing its predictive focus to be tailored to high-risk indicators for each individual, significantly improving the sensitivity and specificity of anomaly detection.

[0013] Third, the matching degree of personalized intervention strategies has improved.

[0014] This approach, based on multidimensional risk assessment and a deep reinforcement learning (DRL) framework, dynamically generates personalized intervention strategies according to the patient's specific risk level and real-time physiological state. This ensures that the final intervention recommendation is both a data-driven optimal response to the current emergency and a prudent choice guided by medical knowledge and in accordance with clinical guidelines. This strategy not only considers individual patient differences but also incorporates the latest medical guidelines and clinical feedback data, significantly improving the matching degree of intervention measures, effectively reducing the decision-making load of physicians, and enhancing the overall efficiency and quality of perioperative management.

[0015] Fourth, system performance can be continuously optimized.

[0016] This solution employs a closed-loop feedback mechanism, using clinical feedback data to update the knowledge graph weights and dynamically adjust system parameters. Utilizing reinforcement learning, it can optimize decision-making strategies in real time based on clinical outcomes, continuously improving system performance and reliability, and ensuring the system maintains efficient and accurate decision support capabilities over the long term.

[0017] Fifth, multimodal data fusion and dynamic updating This solution constructs a multimodal data fusion architecture, systematically integrating heterogeneous data sources throughout the perioperative process, including structured knowledge bases and clinical multimodal data. Through real-time data access and dynamic update mechanisms, the timeliness and accuracy of the knowledge graph are ensured, providing the latest data support for clinical decision-making, improving the system's adaptability and flexibility, and enabling it to meet the diverse needs of different patient groups. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of an embodiment of the perioperative anesthesia risk management method for elderly patients based on multimodal knowledge graph of the present invention. Detailed Implementation

[0019] The following detailed explanation illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 The following is an example of a perioperative anesthesia risk management method for elderly patients based on a multimodal knowledge graph, comprising the following steps: S1. Construct a multimodal knowledge graph of perioperative anesthesia risk in elderly patients. The entities in the multimodal knowledge graph include at least preoperative comorbidities, intraoperative physiology and procedures, and postoperative complications. Edges between entities represent the medical relationships between them, and the strength of these relationships is quantified by weights. Each entity is associated with standardized codes such as ICD-11 and RxNorm to ensure data consistency.

[0020] S1.1, Integrate and preprocess multimodal data. First, systematically integrate heterogeneous data sources throughout the perioperative period and adopt a hierarchical standardization strategy: The structured knowledge base integrates authoritative guidelines such as the "Guidelines for Perioperative Anesthesia Management of Elderly Patients in China (2020 Edition)," extracting key clinical pathways (e.g., "Management Procedure for Intraoperative Hypotension in Hypertensive Patients") and contraindications (e.g., "Contrast agents are contraindicated in patients with eGFR < 30") through a rule engine, mapping them to causal relationships and weights in a knowledge graph. Simultaneously, the latest literature is crawled in real-time via the PubMed API, and new associations between entities are extracted using a BERT model, incrementally updating the rule base after expert review.

[0021] Clinical multimodal data: Structured data (EMR diagnosis, medication prescription): mapped to ICD-11 and RxNorm encoding through FHIR standard, semantically unified, with an error rate of <1%.

[0022] Time series data (physiological signals from monitors): aggregated using a sliding window (length 120 seconds, step size 30 seconds), and time series features extracted by an LSTM encoder. Dropout (rate=0.3) was used to prevent overfitting.

[0023] Unstructured data (medical images, text records): Image features are extracted using a pre-trained ResNet-50 convolutional neural network; text is processed using HanLP for named entity recognition and keyword extraction.

[0024] S1.2, precisely define entities and relationships.

[0025] The preoperative comorbidity entity encompasses common underlying diseases and preoperative conditions in elderly patients, including: Disease entities: For example, hypertension (ICD-11 code: 1-6A60), type 2 diabetes (ICD-11: E11), coronary atherosclerotic heart disease (ICD-11: BA42), chronic obstructive pulmonary disease (ICD-11: CB22), sequelae of stroke (ICD-11: ND9Z), chronic renal insufficiency (ICD-11: GB83), etc. Data is sourced from diagnostic records in the hospital's electronic medical record (EMR) system.

[0026] Physiological indicators: such as preoperative serum potassium (unit: mmol / L, normal range 3.5-5.5), preoperative hemoglobin (unit: g / L, ≥120 for elderly men, ≥110 for women), and left ventricular ejection fraction (LVEF) (unit: %, normal ≥50). Data are derived from preoperative laboratory test reports and echocardiogram reports.

[0027] Medication history entities: For example, long-term use of antihypertensive drugs such as amlodipine (RxNorm code: C09CA02), hypoglycemic drugs such as insulin (RxNorm: C10AA01), and anticoagulants such as warfarin (RxNorm: B01AA03). Data are derived from long-term medical orders and medication history records in EMR.

[0028] Assessment entity: American Society of Anesthesiologists (ASA) classification (Levels I-V). Data sourced from pre-anesthesia visit records.

[0029] Intraoperative parameters: Real-time monitoring of physiological and procedural data during the procedure using solid-focused imaging, including: Vital signs include: mean arterial pressure (MAP) (unit: mmHg, derived from arterial puncture pressure measurement or non-invasive blood pressure monitoring, sampling frequency 1 time / 30-60 seconds, normal range is 65-100), heart rate (HR) (unit: beats / min, derived from electrocardiogram monitoring, normal range is 60-100), end-tidal carbon dioxide partial pressure, and bispectral index (BIS) (derived from the BIS module).

[0030] Drug infusion entities: For example, propofol infusion rate (unit: mg / (kg·h, derived from target-controlled infusion pump or infusion pump data interface) and other anesthetic drug dosages, norepinephrine infusion rate (unit: Use of vasoactive drugs such as ( ).

[0031] Operational entities: such as anesthesia operation data such as crystalloid infusion volume (unit: ml) and operation duration (unit: minutes); and equipment parameters such as anesthesia machine oxygen concentration, ventilator tidal volume, and airway pressure.

[0032] Postoperative complications include adverse events and abnormal indicators that may occur within 72 hours after surgery, including: Acute complications include, for example, postoperative hypotension (defined as MAP < 65 mmHg lasting > 10 minutes), postoperative delirium (positive CAM-ICU assessment), postoperative bleeding (hemoglobin decrease ≥ 30 g / L compared to preoperative level), and postoperative infection (body temperature ≥ 38.5℃ and white blood cell count < 0.05%). ).

[0033] Metabolic abnormalities include, for example, postoperative hypokalemia (defined as serum potassium <3.5 mmol / L, derived from postoperative laboratory reports), postoperative hyperglycemia (random blood glucose >11.1 mmol / L), and postoperative hyponatremia (serum sodium <135 mmol / L).

[0034] Organ dysfunction entities: such as acute kidney injury (serum creatinine ≥26.5 μmol / L above baseline according to KDIGO criteria), postoperative myocardial injury (troponin I ≥0.04 ng / ml above baseline), postoperative respiratory failure (requiring mechanical ventilation for more than 48 hours), etc.

[0035] The medical relationships between the entities include causal relationships, intervention relationships, and contraindication relationships.

[0036] Specifically, medical associations are constructed based on medical guidelines and clinical evidence, and the strength of the association is quantified by weight (ranging from 0 to 1, with higher weight indicating a more significant association). They are divided into three core types of relationships: causal association, intervention association, and contraindication association.

[0037] Causal association indicates that the preceding entity directly causes the occurrence of the following entity. It is often found in pathways such as "preoperative comorbidities → intraoperative parameter abnormalities" and "intraoperative parameter abnormalities → postoperative complications". For example, hypertension (preoperative) → intraoperative hypotension (weight 0.8), diabetes (preoperative) → postoperative infection (weight 0.6), intraoperative MAP < 60 mmHg (lasting for 10 minutes) → postoperative acute kidney injury (weight 0.75), and postoperative hypokalemia → arrhythmia (weight 0.85), indicating that the strength of this causal relationship is relatively high.

[0038] Intervention association describes the improving or moderating effect of interventions on the target entity, mainly reflected in the pathway of "medication / operation → intraoperative parameters / postoperative complications", such as norepinephrine (0.05 μg / (kg)). min)) → Increase MAP (weight 0.9), insulin (intravenous infusion, 0.1U / (kg) h)) → Reduce intraoperative blood glucose (weight 0.88), recheck blood potassium 2 hours postoperatively → Reduce hypokalemia complications (weight 0.7), mechanical ventilation (PEEP 5cm) → Improve (Weight 0.82) reflects the effectiveness of drugs or procedures in managing intraoperative parameters and postoperative complications.

[0039] Contraindication associations refer to situations where the coexistence of one entity and another entity increases the risk and their combined occurrence should be avoided. These are commonly seen in combinations of "drug → comorbidity" or "drug → complications," such as ibuprofen (nonsteroidal anti-inflammatory drug) → history of gastrointestinal bleeding (weight 0.95), beta-blockers (e.g., metoprolol) → severe bradycardia (HR < 50 bpm, weight 0.9), and contrast agents → chronic renal insufficiency. (weight 0.85), high glucose solution infusion → history of diabetic ketoacidosis (weight 0.92), highlighting the risk of drug-induced complications.

[0040] Weighting the relationship edges between entities The weights The calculation integrates the mutual information value between entities, the causal orientation verification results, and the evidence confidence level based on medical guidelines. The specific calculation formula is as follows: .

[0041] in, The corresponding mutual information value is calculated by extracting entities from a historical de-identified medical record database (e.g., a database containing 100,000 surgical records). and Co-occurrence data were collected. Continuous variables (such as MAP and serum potassium) were binned and discretized. The k-nearest neighbor algorithm (k=5) was used to estimate the nonlinear correlation strength between the two variables, and the calculation results were normalized to the [0, 1] interval. For example, the mutual information value between "preoperative serum potassium" and "postoperative hypokalemia" was calculated.

[0042] The corresponding causal direction verification result is calculated using a Python-based causal discovery library (such as DoWhy or Causal Discovery Toolbox). Based on historical data, a PC algorithm or FCI algorithm is used to learn the causal structure, and domain knowledge (such as temporal order: preoperative cause, postoperative consequence) is incorporated as constraints to obtain directed edges. The existence check returns a value of 1 (exists) or 0 (does not exist).

[0043] This corresponds to the confidence level of evidence, and its calculation process includes assigning values ​​based on authoritative medical guidelines and high-quality evidence-based medicine research. For example: Level A evidence (supported by a meta-analysis of multiple randomized controlled trials): assigned a value of 1.2. Level B evidence (supported by a single high-quality randomized controlled trial): assigned a value of 1.0. Level C evidence (supported by observational studies or expert consensus): assigned a value of 0.8. Evidence can be derived from guidelines such as the "Guidelines for Perioperative Anesthesia Management of Elderly Patients in China (2020 Edition)" and the "American Heart Association Guidelines for Perioperative Assessment of Non-Cardiac Surgeries," and must be reviewed and confirmed by a clinical expert committee. For example, the relationship of "hypertension → intraoperative hypotension" is assigned Level A evidence based on guidelines and multiple studies, with a confidence level of 1.2.

[0044] S1.3, Knowledge Graph Construction and Dynamic Network Topology Analysis.

[0045] The TransE model is used to map entities and relations to a low-dimensional vector space (e.g., 128-dimensional). A confidence-weighted voting strategy is employed to determine the final edge weights. ExpertRule comes from the guidelines in S1.1, and ModelPrediction comes from the prediction model trained on historical data.

[0046] The constructed knowledge graph is stored using a graph database (such as Neo4j or Nebula Graph). Entities are treated as nodes, and relationships as weighted edges. During initialization, approximately 500-1000 core medical relationships are imported in batches based on the above rules, forming the basic multimodal knowledge graph.

[0047] After constructing the multimodal knowledge graph, the process includes a dynamic network topology analysis: The Louvain algorithm, incorporating a resolution parameter, is used to partition entities into risk communities. Specifically, community discovery is performed by optimizing the modularity function Q, grouping closely related comorbidities and complications into the same risk community. A node importance ranking algorithm incorporating a time decay factor is then used to assign higher Rank values ​​to recently occurring clinical event nodes, thus identifying core risk transmission paths. For example, calculations can pinpoint core high-risk transmission paths such as diabetes → intraoperative hyperglycemia → postoperative infection. The partitioning formula is as follows: ; in, The elements of the adjacency matrix represent the connection weights between node i and node j (0 / 1 for a bipartite graph). , Let be the degree (total number of connecting edges) of nodes i and j. m is half the total number of edges in the network (undirected graph) or the total number of edges (directed graph). This is the Kronecker delta, which is 1 when nodes i and j belong to the same community, and 0 otherwise. Let i and j be the community labels to which they belong. A spatiotemporally weighted PageRank algorithm is designed, incorporating a time decay factor. Core factors are located using the following formula: ; in, Let be the PageRank value of node u at time step t. α is the damping coefficient (usually taken as 0.85), representing the probability of a random jump. Neighboring nodes PageRank value. For nodes The number of outgoing chains (i.e., the number of edges pointing to other nodes). For smoothing terms (usually taken) ), to prevent division by zero errors. For all predecessor nodes of node u Sum.

[0048] S1.4, Modeling of specific risk factors in elderly patients.

[0049] Building upon a foundational multimodal knowledge graph, we add exclusive entities and relationships to quantify the characteristics of elderly patients: "coexistence of multiple diseases and physiological decline." Collaborative Association: Create virtual nodes, such as hypertension + diabetes synergy, where the weight pointing to intraoperative hypotension is the weighted sum of the weights of individual diseases (e.g., The synergy coefficient of 1.2 is indicated, representing a 20% increase in risk.

[0050] Deep correlation of medication history: Establish the relationship between long-term amlodipine use (≥5 years) and increased sensitivity to intraoperative hypotension, with a weight of 0.85 and an additional elderly accumulation coefficient of 1.1.

[0051] Physiological decline association: Establish the relationship between LVEF < 40% (older) and insufficient surgical center output, with a weight of 0.9, and set age correction coefficients (1.0 for 65-74 years, 1.2 for 75-84 years, and 1.4 for ≥85 years).

[0052] S2. Real-time acquisition of intraoperative multimodal physiological time-series data of target elderly patients.

[0053] The real-time acquisition is achieved through a medical edge computing platform.

[0054] The medical edge computing platform includes: The data acquisition layer is used to acquire multi-source heterogeneous data from the hospital information system and monitoring equipment interfaces. Specifically, by configuring an HL7 FHIR gateway to interface with the hospital's EMR system, it automatically retrieves the patient's diagnosis, medication history, and laboratory test results before the patient enters the operating room and maps them to entities defined in S1. Through RS-232, USB, or network API interfaces, it connects to monitors in the operating room (such as Philips IntelliVue) to acquire waveform or numerical data such as MAP, HR, SpO2, and BIS in real time, with a sampling frequency of 0.5-1 Hz. Through serial port or analog acquisition modules, it connects to target-controlled infusion pumps and vasoactive drug pumps to read drug names, concentrations, and infusion rates in real time. Through the anesthesia machine data port, it acquires end-tidal carbon dioxide partial pressure, oxygen concentration, etc.

[0055] The real-time processing layer processes data using a stream processing engine, with event latency controlled to within 200 milliseconds. Specifically, it is built on the Apache Flink stream processing engine and performs the following processing operations: (1) Data standardization: Map all collected values ​​to a uniform range (e.g., using Z-score standardization), and map text descriptions (e.g., drug names) to standard codes (e.g., RxNorm).

[0056] (2) Window aggregation: Perform sliding window aggregation on high-frequency physiological signals (such as blood pressure at 1 Hz) (window length 120 seconds, step size 30 seconds), calculate the mean, variance, trend and other features within the window, and form a feature vector for the model to use.

[0057] (3) Event encapsulation and distribution: The processed data is encapsulated into events in a unified format (such as JSON) and published to a message queue (such as Kafka). The system ensures that the total latency from data collection to event publication is less than 200 milliseconds.

[0058] Furthermore, and notably, considering the slower changes in physiological parameters in elderly patients, the sliding window length was adjusted from the usual 60 seconds to 120 seconds, and the step size was adjusted from 15 seconds to 30 seconds. Data streams containing high-risk parameters (such as MAP and derived parameters related to LVEF) were given higher priority for acquisition and processing.

[0059] S3. Based on the multimodal knowledge graph and the real-time data acquired in S2, perform dynamic risk monitoring: process the physiological time-series data using a time-series prediction model to obtain predicted values ​​of physiological parameters; and dynamically determine whether abnormal risk events exist by combining the entity association paths related to the target patient in the multimodal knowledge graph. Specifically, this includes the following operations: S3.1 When the target patient enters the operating room, based on their preoperative entities (such as "hypertension" or "diabetes"), a SPARQL-like query is executed in the graph database to retrieve all connected paths, forming a "patient current risk subgraph". For example, the retrieved paths are: "hypertension → intraoperative hypotension (weight 0.8)" and "diabetes → postoperative infection (weight 0.6)".

[0060] S3.2, Time Series Prediction.

[0061] The temporal prediction model is an LSTM-Attention model; the attention weight allocation of the LSTM-Attention model is adjusted by the prior knowledge of the multimodal knowledge graph: based on the abnormal path weights of comorbidities-intraoperative parameters extracted from the multimodal knowledge graph that are associated with the preoperative comorbidities of the target patient, the attention weights of the corresponding physiological parameters are biased and adjusted.

[0062] The dynamic determination includes: comparing the predicted values ​​of physiological parameters with a safety threshold range dynamically generated based on the patient's individual baseline and the information in the multimodal knowledge graph; and / or triggering a contraindication association rule in the multimodal knowledge graph that matches the current physiological state (e.g., the patient is using a β-blocker and HR < 50 beats / min).

[0063] Specifically, real-time physiological time-series data processed by S2 (such as MAP and HR sequences from the past 2 minutes) is used as model input. This data includes structured features (such as comorbidity vectors), which are converted into feature vectors through a multilayer perceptron (MLP). Temporal features (such as MAP sequences) are converted into feature vectors using an LSTM encoder (128-dimensional hidden layer). .

[0064] The two are spliced ​​together By setting it to (Time-series weight) is usually greater than 0.6 to ensure that dynamic data dominates.

[0065] ; ; in, Indicates the weight of time series data. Represents the weights of structured data, hyperparameters Set it to 0.4-0.6 (preferably 0.5). This represents the Euclidean norm of the eigenvectors. This weighting mechanism can dynamically adjust the relative importance of the two data types during the fusion process, ensuring the robustness and accuracy of the fusion result.

[0066] Will A single-layer LSTM network with a hidden layer dimension of 128 is used, followed by an attention layer for feature processing. The attention layer assigns weights to the hidden states at each time step. The dropout rate is 0.3, and the loss function is the mean squared error (MSE).

[0067] During model initialization or inference, adjustments are made based on the risk subgraph retrieved in S3.1 to bias the attention mechanism. For example, if the weight of "hypertension → intraoperative hypotension" is high (0.8) in the subgraph, the base attention weight of the MAP feature (usually 0.3) is increased by 60%, i.e., adjusted to 0.48, while the weight of the HR feature is reduced accordingly, making the model more focused on the fluctuation of MAP.

[0068] The model outputs predicted values ​​of physiological parameters for the next 30 seconds (or the next time step). Simultaneously, based on the patient's preoperative baseline (such as the preoperative mean MAP) and safety rules in the knowledge graph, an individualized safety threshold range is dynamically calculated. .For example, .like If the value consistently exceeds this range (e.g., for three consecutive prediction cycles), an anomaly detection is triggered. Furthermore, if real-time physiological parameter values ​​exceed this range, an anomaly detection is also triggered.

[0069] The corresponding determination formula is: .

[0070] These are real-time physiological parameter values.

[0071] S3.3, Comprehensive Risk Scoring and Classification.

[0072] During dynamic risk monitoring, patients are also subject to comprehensive risk stratification: By integrating static risk assessment based on the aforementioned multimodal knowledge graph, dynamic risk scoring based on real-time data, and intraoperative event urgency factors, a comprehensive risk score is calculated and categorized into different risk levels. Specifically, the following three levels of scores are combined: .like: .

[0073] .

[0074]

[0075] Event layer coefficient Adjust dynamically according to intraoperative emergencies (such as massive bleeding), within a range of 0.5-1.5.

[0076] For example, the calculated score is 22.

[0077] Based on risk scores, patients were categorized into three risk levels: low risk (≤15 points), intermediate risk (16-25 points), and high risk (≥26 points). Low-risk patients require no intervention; intermediate-risk patients are advised to have vasoactive drugs on hand; and high-risk patients require multidisciplinary consultation.

[0078] S4. When an abnormal risk event is determined to exist, a personalized intervention decision is generated based on the deep reinforcement learning (DRL) framework: the risk score and real-time physiological parameters, which are integrated with the association weights of the multimodal knowledge graph, are used as the state input of the DRL model, and the optimal intervention action is output from the predefined intervention action space.

[0079] The DRL model is a deep Q-network (DQN); the state input also includes the patient's preoperative comorbidity characteristics; the action options in the intervention action space are pre-screened according to the intervention association and contraindication association rules in the multimodal knowledge graph.

[0080] Specifically, the state space of the DRL model : .

[0081] The comorbidity vector is a multidimensional binary vector generated based on the S3.1 search results (e.g., [1,0,1] represents hypertension, no coronary heart disease, and diabetes).

[0082] Action Space (A): A discrete set of actions, such as {"Increase norepinephrine by 0.05 μg / (kg·min)", "Decrease propofol infusion rate by 10%", "Rapidly infuse 200ml of crystalloid solution", "No action"}. This space is filtered using contraindication associations from the knowledge graph during initialization. For example, if a patient has "chronic renal insufficiency," and "contrast agent → chronic renal insufficiency" is a contraindication association in the knowledge graph (weight 0.85), then DRL will exclude the action of "using contrast agent"; if "norepinephrine → increasing MAP" has a weight of 0.9 (higher than dopamine's 0.4), then norepinephrine dose adjustment will be prioritized as an action option.

[0083] The reward function of the DRL model is designed as follows: ; in, In order to take action Then, the risk score is estimated at the next moment. As an indicator function for complications, This is an indicator function for guideline compliance.

[0084] The rule for determining the value is: if no new complications occur during the current decision-making cycle, the indicator function is 1.

[0085] The determination of guideline compliance is accomplished by querying the evidence confidence level of the corresponding intervention action in the multimodal knowledge graph; specifically... The rule for determining the value is: Does it align with Level A evidence in the knowledge graph regarding the intervention association? For example, if "norepinephrine (0.05 μg / kg / min) → increases MAP" is Level A evidence in the graph, and the DRL recommends this action, then I (guideline compliance) = 1, and an additional 2 points are awarded. This differentiated reward mechanism ensures that DRL decisions are both tailored to individual circumstances and comply with clinical guidelines.

[0086] Furthermore, real-time policy generation is based on the following formula: ; in, Using discount factors, we generate real-time optimal strategies by maximizing future reward values, providing dynamic support for clinical decision-making.

[0087] Furthermore, when S3 triggers a medium- or high-risk warning, the DRL decision engine is activated.

[0088] The engine receives the current state. The Q-values ​​of all possible actions a are calculated using a trained Deep Q-Network (DQN). .

[0089] Choose the action with the highest Q value. This is used as a recommendation decision output. For example, the output decision might be: "Current MAP is trending downward, risk score 22 (intermediate risk), recommendation: intravenous bolus injection of norepinephrine." "and recheck MAP 5 minutes later."

[0090] Decision recommendations are pushed to anesthesiologists in a prominent manner through smart dashboards in the operating room or screens on the anesthesia workstation.

[0091] S5. Based on the clinical outcome feedback after the intervention, dynamically adjust the weights of the corresponding edges in the multimodal knowledge graph to achieve closed-loop collaborative optimization between the knowledge graph and the decision-making model.

[0092] Specifically, postoperatively, outcome data were collected from EMR and nursing records: whether complications such as intraoperative hypotension and postoperative hypokalemia occurred, and the actual interventions taken by the physician that differed from the system recommendations and their effects.

[0093] Then, the knowledge graph weights are dynamically adjusted. This dynamic adjustment uses a weight update formula based on a reinforcement learning mechanism: ; in, For edge weights, For learning rate, This refers to the real-time reward value (i.e., the actual reward value obtained by the DRL model in this case in S4). This is the average reward value for recent cases involving this relationship; The weight gradient, i.e., the weight For the final reward The contribution gradient is approximately estimated through the backpropagation process of the DRL model.

[0094] For example, in a case linked to a pathway of "hypertension → intraoperative hypotension," this protocol provided an early warning and recommended measures, ultimately preventing hypotension and resulting in a high reward. The system will then slightly increase the weight of that path. This will make it more influential in similar cases in the future.

[0095] Furthermore, accumulated new case data are periodically (e.g., weekly) added to the experience replay pool to retrain the DQN network offline, enabling the decision-making strategy to adapt to the latest knowledge graph weights and clinical practice models.

[0096] To facilitate understanding, the following two application examples illustrate the process of this solution: Application Case 1: Intraoperative hypotension early warning in hypertensive patients This application case demonstrates a complete closed-loop process for an elderly patient with essential hypertension, from preoperative risk assessment and real-time intraoperative alerts to personalized decision support. The patient, a 72-year-old male, was preoperatively diagnosed with essential hypertension (ICD-11: BA00) and had been taking amlodipine (10 mg / day, RxNorm: 329526) on a long-term oral medication. His baseline preoperative blood pressure was 160 / 95 mmHg.

[0097] Data Input and Processing: The system automatically extracts the structured data from the hospital's electronic medical record (EMR) via an HL7 FHIR gateway. After the patient enters the operating room, the monitor transmits physiological signals such as mean arterial pressure (MAP), heart rate (HR), and blood oxygen saturation in real time at a frequency of 30 seconds per transmission. All data is fed into a stream processing platform built on Apache Flink. The structured data is mapped to entity node attributes such as "hypertension" and "amlodipine" in a knowledge graph; the real-time physiological signals are aggregated through a sliding window (window length 120 seconds, step size 30 seconds) to form a standardized temporal feature vector, ready to be input into the prediction model.

[0098] Knowledge-Driven Risk Assessment and Model Preparation: The system performs intelligent retrieval within the initially constructed multimodal knowledge graph. First, based on the patient entity, the core causal path of "hypertension → intraoperative hypotension" is identified, with its initial weight determined to be 0.8 by a combination of guideline evidence (confidence level A) and historical data mutual information. Next, complex network analysis is performed, utilizing an improved Louvain algorithm to identify the patient's hypertension-cardiovascular risk community, and confirming "intraoperative hypotension" as a core risk node within this community using a spatiotemporal PageRank algorithm. Based on the intervention associations identified in the graph, the system pre-screens "norepinephrine" and "dopamine" as candidate intervention drugs, with norepinephrine showing a higher association weight (0.9) and evidence level (level A). This knowledge is used to configure the subsequent real-time analysis model: the attention mechanism of the LSTM-Attention model is pre-biased, assigning a higher base attention weight to the MAP parameter (increasing it from 0.3 to 0.5) to ensure sensitivity to abnormal blood pressure fluctuations.

[0099] Multimodal fusion real-time early warning and decision generation: During surgery, real-time physiological time-series data is semantically aligned and fused with the patient's structured features (history of hypertension, medication history). The fused features are input into a biased LSTM-Attention model, which not only analyzes historical trends but also focuses on the MAP parameter, predicting a MAP value of 85 mmHg for the next 5 minutes. Simultaneously, the system dynamically calculates the individualized safety threshold lower limit based on the patient's preoperative baseline blood pressure (MAP≈107 mmHg), which is max(107*0.8,60)=85.6 mmHg. If the predicted value remains below the threshold, an anomaly detection is triggered. Simultaneously, the system calculates a comprehensive risk score, integrating static hypertension weights, dynamic MAP deviation, and intraoperative status. A score of 22 triggers a "high-risk" red alert. At this point, the deep reinforcement learning (DRL) decision engine is activated. The engine receives real-time MAP, HR, and a risk score incorporating graph weights, and outputs the optimal action from a pre-selected action space: "It is recommended to adjust the norepinephrine infusion rate to 0.05 μg / (kg·min)." This recommendation is highly consistent with the high-weighted interventions of Level A evidence in the knowledge graph, and therefore will receive a higher expected reward for "guideline compliance" in the DRL reward function.

[0100] Closed-loop feedback optimization: The anesthesiologist adopted the suggestion and implemented the intervention, and the patient's blood pressure remained stable during the operation, without persistent hypotension. Postoperatively, the system obtained this outcome and determined that the DRL decision received a positive reward. According to the reinforcement learning weight update mechanism, the system used this reward value to fine-tune the weight of the "hypertension → intraoperative hypotension" edge in the knowledge graph with a small learning rate (η=0.05), slightly increasing it from 0.800 to 0.805. This will allow for more precise reinforcement of the risk influence of this path in subsequent similar cases.

[0101] Application Case 2: Prevention of Postoperative Hypokalemia in Diabetic Patients This application case demonstrates how the system can identify cross-cycle risk paths using a knowledge graph and provide preventative decision support. The patient, a 68-year-old female, was preoperatively diagnosed with type 2 diabetes (ICD-11: 5A11).

[0102] Data Input and Complex Network Analysis: The system retrieved diabetes diagnosis information from the EMR. During the operation, in addition to routine vital signs, the system also captured the key operational data of "intraoperative potassium supplementation of 20 mmol" through the infusion pump interface. In the background, the system continuously performed dynamic network topology analysis on the knowledge graph. For this patient, the system identified risk communities centered on "diabetes" and "electrolyte imbalance" using a community segmentation algorithm. Furthermore, using the spatiotemporal PageRank algorithm, combined with real-time intraoperative data (potassium supplementation <30 mmol), the system automatically activated and highlighted the high-risk transmission path of "insufficient intraoperative potassium supplementation + diabetes → postoperative hypokalemia". This path was discovered through historical data mining, with a weight of 0.7 and an evidence level of B.

[0103] Preventative Early Warning and Decision Recommendation: Based on this analysis, the system does not wait for abnormalities to occur but proactively predicts risks. It uses an LSTM model to analyze the patient's physiological homeostasis trends and, combined with the potential impact weight of the "diabetes" entity on metabolic disorders, predicts a significantly increased risk of hypokalemia 2 hours post-surgery. Therefore, the system triggers a preventative yellow alert intraoperatively and generates a clear intervention recommendation on the decision support interface: "The patient has diabetes; intraoperative potassium supplementation (20 mmol) may be insufficient. It is recommended to recheck serum potassium 2 hours post-surgery." This recommendation is directly derived from the pre-defined intervention strategy library in the knowledge graph for identified risk pathways.

[0104] Closed-loop learning: Postoperatively, medical staff followed the advice to check serum potassium levels, which were 3.8 mmol / L (normal), and hypokalemia did not occur. This positive clinical outcome was captured by the system as positive feedback. The system then initiated an incremental update process, appropriately increasing the weight of the associated pathway "insufficient intraoperative potassium supplementation + diabetes → postoperative hypokalemia" (from 0.700 to 0.710) based on this successful prevention case. At the same time, the confidence of "checking serum potassium 2 hours postoperatively" as an effective intervention in this pathway was strengthened, making the system's future prevention recommendations more accurate and reliable.

[0105] This embodiment provides a perioperative anesthesia risk management method for elderly patients based on multimodal knowledge graphs, which can improve the accuracy of preoperative risk prediction for elderly patients, shorten the detection time of intraoperative abnormal events, improve the matching degree of personalized intervention strategies, and reduce the decision-making load of doctors.

[0106] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for perioperative anesthesia risk management in elderly patients based on multimodal knowledge graphs, characterized in that, Includes the following steps: S1. Construct a multimodal knowledge graph of perioperative anesthesia risk for elderly patients. The entities in the multimodal knowledge graph include at least preoperative comorbidities, intraoperative physiology and operation, and postoperative complications. The edges between entities are used to represent the medical association between entities, and the association strength is quantified by weight. S2. Real-time acquisition of intraoperative multimodal physiological time-series data of target elderly patients; S3. Based on the multimodal knowledge graph and the real-time data acquired in S2, perform dynamic risk monitoring: use a time-series prediction model to process the physiological time-series data to obtain predicted values ​​of physiological parameters; By combining the entity association paths related to the target patient in the multimodal knowledge graph, the existence of abnormal risk events can be dynamically determined; S4. When an abnormal risk event is determined to exist, a personalized intervention decision is generated based on the deep reinforcement learning (DRL) framework: the risk score and real-time physiological parameters, which are integrated with the association weights of the multimodal knowledge graph, are used as the state input of the DRL model, and the optimal intervention action is output from the predefined intervention action space. S5. Based on the clinical outcome feedback after the intervention, dynamically adjust the weights of the corresponding edges in the multimodal knowledge graph to achieve closed-loop collaborative optimization between the knowledge graph and the decision-making model.

2. The method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graph according to claim 1, characterized in that, In S1, the medical associations between entities include causal associations, intervention associations, and contraindication associations; the calculation of the weights integrates the mutual information values ​​between entities, the causal direction verification results, and the evidence confidence level based on medical guidelines.

3. The method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graphs according to claim 1, characterized in that, In S1, after constructing the multimodal knowledge graph, the step of performing dynamic network topology analysis on the graph is also included: using the Louvain algorithm with a resolution parameter to divide entities into risk communities, and using a node importance ranking algorithm with a time decay factor to locate the core risk transmission path.

4. The method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graph according to claim 1, characterized in that, In S3, the time-series prediction model is an LSTM-Attention model; The dynamic determination includes comparing the predicted values ​​of physiological parameters with a safety threshold range dynamically generated based on the patient's individual baseline and the multimodal knowledge graph information; And / or, trigger taboo association rules in the multimodal knowledge graph that match the current physiological state.

5. The method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graph according to claim 4, characterized in that, The attention weight allocation of the LSTM-Attention model is adjusted by the prior knowledge of the multimodal knowledge graph: the attention weights of the corresponding physiological parameters are biased and adjusted according to the abnormal path weights of the comorbidities-intraoperative parameters extracted from the multimodal knowledge graph that are associated with the preoperative comorbidities of the target patient.

6. The method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graph according to claim 1, characterized in that, In S4, the DRL model is a deep Q-network (DQN); the state input also includes the patient's preoperative comorbidity characteristics; the action options in the intervention action space are pre-screened according to the intervention association and contraindication association rules in the multimodal knowledge graph.

7. The method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graph according to claim 6, characterized in that, The reward function of the DRL model is designed as follows: ； in, This is a risk score estimated at the next moment after an action is taken; As an indicator function for complications, The indicator function for guideline compliance is used, and the determination of guideline compliance is accomplished by querying the evidence confidence of the corresponding intervention action in the multimodal knowledge graph.

8. The method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graph according to claim 1, characterized in that, In S3, during dynamic risk monitoring, patients are also subject to comprehensive risk stratification: By integrating static risk assessment based on the multimodal knowledge graph, dynamic risk scoring based on real-time data, and intraoperative event urgency factors, a comprehensive risk score is calculated and classified into different risk levels.

9. The method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graph according to claim 1, characterized in that, In S2, the real-time acquisition is achieved through a medical edge computing platform; The medical edge computing platform includes: a data acquisition layer, used to acquire multi-source heterogeneous data from the hospital information system and monitoring equipment interface; The real-time processing layer processes data based on a stream processing engine, with event latency controlled within 200 milliseconds.

10. The method for perioperative anesthesia risk management of elderly patients based on multimodal knowledge graph according to claim 1, characterized in that, In S5, the dynamic adjustment adopts a weight update formula based on a reinforcement learning mechanism: ； in, For edge weights, For learning rate, For real-time reward value, The average reward value, This represents the weight gradient.

Images