Multi-modal data fusion perioperative period risk prediction and intervention method and system

Abstract

The invention discloses a perioperative period risk prediction and intervention method and system based on multi-modal data fusion, and relates to the technical field of medical artificial intelligence, and the method comprises the following steps: S100, building a data collection mechanism, obtaining perioperative period related data from a plurality of medical information systems, the data comprises structured examination indexes, text medical history records, medical image report conclusions and pathological diagnosis results. According to the method, the high-dimensional feature vector is constructed by fusing the structured test data, the text medical history, the image conclusion and the pathological result, and the quantitative prediction of the postoperative complications is realized in combination with the trained risk prediction model; the system can automatically generate a natural language interpretation and personalized intervention plan, and an electronic medical record is embedded, so that closed-loop management of risk identification, cause interpretation and intervention execution is realized, the intelligent, standardized and personalized level of perioperative period management is improved, and the occurrence rate of complications is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to a method and system for perioperative risk prediction and intervention based on multimodal data fusion. Background Technology

[0002] "Perioperative Risk Prediction and Intervention Based on Multimodal Data Fusion" refers to the systematic integration of various types of data from different medical information systems—such as structured data (laboratory test indicators, vital signs, medication records, etc.), text data (admission records, doctor's notes, medical history descriptions, etc.), imaging data (CT and MRI imaging conclusions or quantitative features), and pathological data (pathology report text)—through standardized processing and unified modeling to form a high-dimensional fusion feature vector that comprehensively reflects the patient's clinical status. Subsequently, a trained machine learning model is used to predict and intervene in various postoperative complications (such as thrombosis, etc.). This method predicts the probability of developing risks such as infection, myocardial infarction, and kidney injury. Once high-risk items are identified, a Large Language Model (LLM) is used to provide natural language explanations of the causes. Based on authoritative clinical guidelines and expert knowledge, specific, personalized, and actionable intervention plans (such as recommended medications, nursing protocols, and monitoring strategies) are automatically generated and seamlessly integrated into the electronic medical record system. This assists physicians in making more accurate, standardized, and forward-looking medical decisions before surgery. This approach achieves intelligent closed-loop management throughout the entire process, from dispersed data to centralized modeling, from fuzzy judgment to quantitative assessment, and from risk identification to intervention actions, significantly improving perioperative patient safety and medical quality.

[0003] The existing technology has the following shortcomings: Currently, widely used perioperative risk assessment methods are mostly experience-based or rating scale-based, such as the ASA classification, Caprini thrombosis risk score, and Goldman Sachs cardiac risk index. While these methods have certain guiding value in specific fields, they generally suffer from the following shortcomings: First, the assessment heavily relies on the physician's personal experience and subjective judgment, easily overlooking potential risks in complex, rare, or atypical cases. Second, they cannot fully integrate multi-source, multimodal clinical information from electronic medical records, including structured (laboratory tests, vital signs), text (medical history, physician notes), imaging (CT, MRI reports), and pathology, leading to the easy neglect of key "weak signals." Third, the assessment models are mostly static and rule-driven, failing to dynamically reflect changes in patient status and lacking personalized adjustment capabilities. Fourth, even when high risk is identified, there is a lack of intervention plans based on standardized guidelines combined with specific patient characteristics, resulting in a lack of standardization and consistency in clinical responses. Therefore, constructing a closed-loop system that can integrate multimodal medical data and achieve quantitative prediction, automatic interpretation, and intelligent intervention has become an important development direction for medical artificial intelligence and clinical decision support.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for perioperative risk prediction and intervention based on multimodal data fusion, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for perioperative risk prediction and intervention based on multimodal data fusion, comprising the following steps: S100: Establish a data collection mechanism to obtain perioperative-related data from multiple medical information systems. The data includes structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results. S200 standardizes the acquired structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results, including normalizing structured numerical indicators, parsing medical entities in text medical history records, extracting quantitative feature information from imaging reports and pathology reports, and forming a high-dimensional fusion feature vector with unified structure. S300 integrates feature vectors into the trained complication risk prediction model and outputs the probability values ​​of various perioperative complications to achieve quantitative prediction of complication risk. S400 generates a graphical risk level distribution map based on the probability value of complication occurrence, and generates natural language descriptions to explain the causes of high risk based on key feature information in the fused feature vector. S500: When the risk probability value of a certain complication reaches a preset threshold, a corresponding intervention plan is generated based on the high-risk cause information and medical knowledge base content. The intervention plan includes medication use recommendations, monitoring items and frequencies, nursing pathways and rehabilitation guidance. S600 integrates intervention plans into the preoperative discussion template or postoperative medical order template of the electronic medical record system, realizing an automated closed loop for perioperative risk identification and intervention recommendations.

[0007] Preferably, step S100 includes: Obtain structured examination indicators, including blood routine test indicators, biochemical indicators, coagulation function indicators and vital signs parameters; Obtain textual medical history records, including admission records, present illness history, past medical history, personal history, family history, preoperative anesthesia visit records, preoperative discussion records, multidisciplinary consultation opinions, and progress notes, and identify medical entity information through terminology standardization processing; Obtain the conclusions of medical imaging reports, including qualitative descriptions of lesion location, nature, size, boundary, enhancement features, and surrounding relationships in imaging diagnoses, and extract structured fields; Obtain pathological diagnostic results, including tumor type, differentiation degree, depth of invasion, lymph node metastasis, resection margin status, and expression of immunohistochemical indicators, and construct a standardized set of structural fields.

[0008] Preferably, step S200 includes: The structured inspection indicators are normalized and outliers are removed, and the maximum, minimum, average and rate of change in the time series are extracted. The text medical history records are standardized in terms of terminology and parsed in terms of grammatical rules. Disease names, duration of disease, control status and treatment records are extracted and converted into structured fields. The conclusions of medical imaging reports are analyzed by keyword matching and sentence structure parsing to extract the lesion location, nature, size, boundary features, and preliminary diagnostic results, and then the results are transformed into a structured form. Key fields were extracted and the coding structure was transformed from the pathological diagnosis results. Pathological type, degree of differentiation, depth of invasion, lymph node status, resection margin, and expression of immunohistochemical markers were extracted to construct a structured feature set in a unified format.

[0009] Preferably, step S300 includes: The fields in the fused feature vector are filtered and matched for consistency to ensure that the units of each field are consistent, the format is standardized, and the values ​​are valid. Input the validated fusion feature vectors into the complication risk prediction model according to the preset field order, and output the risk probability values ​​of each perioperative complication. The probability values ​​of each risk are standardized and expressed, and corresponding risk level labels are generated based on preset classification thresholds; The complication name, risk probability value, risk level, prediction timestamp, and feature source information are compiled into a structured list to generate a comprehensive risk profile of the patient during the perioperative period.

[0010] Preferably, step S400 includes: Based on the risk probability value of each complication and the preset threshold standard, risk levels are classified and a risk level distribution map in the form of a radar chart is drawn. Identify key features from the fused feature vectors that influence the prediction results of high-risk complications; Based on high-risk events, key triggering characteristics, synergistic aggravating factors, and comprehensive risk warnings, a semantic structure is constructed to generate natural language description content. The graphical risk level distribution map and natural language description are simultaneously compiled into a risk assessment report and embedded into the electronic medical record system.

[0011] Preferably, in the step of identifying key feature information, the feature data in the structured examination indicators, parsed text medical history records, imaging report conclusions and pathological diagnosis results are compared with the standard risk feature value range of the corresponding complications to screen out variables with abnormal values ​​or high-risk semantic representations, and sort them according to their degree of influence on the prediction results of complication risks, which are then used for the generation of subsequent natural language description content.

[0012] Preferably, step S500 includes: Determine whether the risk probability value of a certain complication exceeds a preset threshold; if it does, trigger the intervention condition. Key features that influence the prediction of this complication risk are extracted as high-risk factors. Match medical knowledge base entries corresponding to the high-risk cause and screen for intervention measures that meet the adaptation criteria; The system is structured and generated into text based on four categories: medication use recommendations, monitoring items and frequencies, nursing pathways, and rehabilitation guidance. The output is readable and actionable intervention recommendations.

[0013] Preferably, step S600 includes: Each suggestion in the intervention plan is broken down into fields and mapped into a structured format; Depending on the patient's perioperative stage, intervention recommendations will be written into either the preoperative discussion template or the postoperative medical order template. Nursing recommendations and rehabilitation recommendations will be updated in conjunction with the nursing record sheet and rehabilitation record sheet; Attach a generation description to each suggestion and record the doctor's response behavior in the intervention suggestion response log.

[0014] A multimodal data fusion-based perioperative risk prediction and intervention system includes a data acquisition module, a feature construction module, a risk prediction module, a risk interpretation module, an intervention generation module, and an intervention execution closed-loop module. The data acquisition module establishes a data acquisition mechanism to obtain perioperative-related data from multiple medical information systems. The data includes structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results. The feature construction module standardizes the acquired structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results. This includes normalizing structured numerical indicators, parsing medical entities in text medical history records, extracting quantitative feature information from imaging reports and pathology reports, and forming a high-dimensional fusion feature vector with a unified structure. The risk prediction module integrates the feature vector inputs into the complication risk prediction model after training, and outputs the probability values ​​of various perioperative complications to achieve quantitative prediction of complication risk. The risk interpretation module generates a graphical risk level distribution map based on the probability value of complication occurrence, and generates natural language descriptions to explain the causes of high risk based on key feature information in the fused feature vector. The intervention generation module generates a corresponding intervention plan based on high-risk cause information and medical knowledge base content when the risk probability value of a certain complication reaches a preset threshold. The intervention plan includes medication use recommendations, monitoring items and frequencies, nursing pathways and rehabilitation guidance. The intervention execution closed-loop module integrates the intervention plan into the preoperative discussion template or postoperative medical order template of the electronic medical record system, realizing the automated execution closed loop of perioperative risk identification and intervention recommendations.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a high-dimensional feature representation that integrates structured test indicators, free-text medical history, imaging conclusions, and pathological results. This feature representation is then input into a trained complication risk prediction model. The system can quantitatively predict the risk of various postoperative complications (such as thrombosis, infection, kidney injury, and myocardial infarction). Furthermore, by incorporating an interpretable mechanism, the system can automatically generate natural language explanations based on key influencing factors, effectively improving the clinical comprehensibility and acceptability of the prediction results. When the risk level reaches a preset threshold, the system automatically matches appropriate intervention measures from a medical knowledge base, constructing a personalized intervention plan covering medication recommendations, nursing procedures, monitoring frequency, and rehabilitation pathways. This plan is automatically embedded into key nodes of the electronic medical record system in a structured manner, achieving closed-loop management of the entire process from risk identification and causal interpretation to intervention execution. This invention significantly improves the intelligence, standardization, and personalization of perioperative risk management, helping to reduce the incidence of postoperative complications and improve patient safety and medical quality. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the perioperative risk prediction and intervention method based on multimodal data fusion of the present invention.

[0018] Figure 2 This is a schematic diagram of the modules of the perioperative risk prediction and intervention system based on multimodal data fusion of the present invention.

[0019] Figure 3 This is a schematic diagram of the data fusion and risk prediction process.

[0020] Figure 4 A schematic diagram of the process for generating intervention plans and integrating them with EMR. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0022] This invention provides, for example Figure 1 The perioperative risk prediction and intervention method based on multimodal data fusion, as shown, includes the following steps: S100: Establish a data collection mechanism to obtain perioperative-related data from multiple medical information systems. The data includes structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results. The first step in perioperative risk prediction and intervention methods is to establish a data collection mechanism to acquire clinical data related to surgery. This involves analyzing and integrating the data structures of multiple information systems within the hospital to collect key information generated by patients at various stages before, during, and after surgery. Specifically, this includes structured examination indicators, text-based medical history records, conclusive descriptions from medical imaging reports, and the final results of pathological diagnoses. This step, by clearly defining data fields, collection paths, and data processing rules, provides a complete and unified data input foundation for subsequent feature construction and risk prediction.

[0023] The structured examination indicators were collected. This type of data mainly comes from the hospital's laboratory information system, vital signs monitoring and recording system, and inpatient medical record electronic system. The specific collection content includes the following examination items: (1) blood routine test indicators, such as hemoglobin content (g / L), hematocrit (%), white blood cell count (×10). 9 (2) Biochemical indicators, including albumin concentration (g / L), total protein concentration (g / L), creatinine concentration (μmol / L), blood glucose (mmol / L), alanine aminotransferase (U / L), aspartate aminotransferase (U / L), alkaline phosphatase (U / L), and total bilirubin (μmol / L); (3) Coagulation function indicators, including prothrombin time (seconds), activated partial thromboplastin time (seconds), international normalized ratio (INR), and D-dimer concentration (mg / L); (4) Vital signs, such as body temperature (°C), pulse (beats / min), systolic and diastolic blood pressure (mmHg), respiratory rate (beats / min), and blood oxygen saturation (%). Each type of data is stored in the information system based on a fixed-format electronic form structure, and is extracted in batches by calling standard interfaces. It is also time-labeled according to the collection time point (24 hours before surgery, 1 hour after surgery, and 72 hours after surgery) to distinguish different stages of the state.

[0024] The collection of text-based medical history records was conducted. These records primarily originated from the electronic medical record system, including physician-written admission records, present illness, past medical history, personal history, family history, preoperative anesthesia visit records, preoperative discussion records, preoperative multidisciplinary consultation opinions, and daily progress notes. This text data was written in natural language, with inconsistent sentence structures and instances of synonyms and abbreviations being used interchangeably. During the collection process, a terminology standardization method based on a standard medical glossary was employed to analyze the free text word by word, identifying clinically significant disease names, disease progression descriptions, behavioral factors, and surgical history. For example, in the "present illness" section, "patient's weight loss exceeding 5 kg in the past three months" was extracted; in the "past medical history" section, "hypertension for over 10 years, poorly controlled" was identified; in the "admission record" section, "chief complaint: persistent right lower quadrant abdominal pain for 2 weeks" was extracted; and in the "preoperative discussion record" section, "planned laparoscopic right hemicolectomy, with a high probability of tumor" was identified. All extracted results are stored as structured data items, including fields such as disease name, time information, and clinical judgment conclusions, to ensure usability and consistency in subsequent analyses.

[0025] This invention collects the conclusions from medical imaging reports. Medical imaging reports are generated by radiologists after completing imaging examinations and typically include two parts: "Imaging Findings" and "Imaging Diagnosis." This invention focuses on collecting the clinical judgments of surgery-related lesions in the "Imaging Diagnosis" section, such as: "CT scan shows thickening of the sigmoid colon wall with uneven local enhancement and blurred adjacent fat spaces, suggesting a possible neoplastic lesion," "MRI shows multiple low-signal nodules in the left lobe of the liver, with partial enhancement, requiring investigation of metastases," and "Chest X-ray shows a leaf-shaped infiltrate in the right lower lung, suspected to be an infectious focus." To ensure the accuracy of the collected content, an imaging diagnostic descriptive vocabulary is developed, covering fields such as location, lesion nature, size, enhancement characteristics, and relationship with surrounding structures. The report content is parsed based on predefined sentence structure rules and converted into structured fields. For example, extracted fields might include: Lesion location = "sigmoid colon," Lesion type = "space-occupying lesion," Possible diagnosis = "tumor," Lesion size = "4.2cm × 3.8cm," Lesion boundary = "unclear," and Enhancement characteristics = "heterogeneous." Each extracted content is indexed with the original image report text to facilitate traceability and manual verification.

[0026] The pathological diagnosis results are collected. Pathological diagnostic data mainly comes from the pathology information system, including preoperative biopsy, intraoperative frozen section results, and postoperative formal pathology reports. This implementation prioritizes extracting the final diagnostic conclusion from the report, specifically including tumor type, differentiation degree, depth of invasion, lymph node metastasis, surgical margin status, and characteristic immunohistochemical indicators. For example, the report might state, "Moderately differentiated adenocarcinoma, invading extraserosa, mesenteric lymph node metastasis 5 / 18, tumor margin negative, CDX2 positive, CK20 positive, CK7 negative." The extracted results are converted into structured fields, such as: Pathological diagnosis = "Adenocarcinoma", Differentiation degree = "Moderately differentiated", Invasion degree = "Extraserosa", Lymph node metastasis = "Positive, 5 / 18", Surgical margin status = "Negative", Immunohistochemical expression = "CDX2 positive, CK20 positive, CK7 negative". In addition, the histological subtypes in the report are identified. If a special subtype such as "mucinous adenocarcinoma" or "signet ring cell carcinoma" exists, the feature is saved as a label so that it can be assigned the corresponding risk weight in subsequent feature construction.

[0027] Through the above operations, structured examination indicators, textual medical history records, medical imaging report conclusions, and pathological diagnosis results were collected categorized by type. The content covers key clinical parameters required for surgical risk assessment. The format was standardized, fields were clearly labeled, and semantics were cleaned and normalized, ultimately constructing a unified, information-rich clinical data set covering key perioperative conditions. This data collection method not only possesses information integrity and data universality but can also be rapidly deployed in different hospital environments, meeting the clinical practicality requirements for data preparation for patients with different surgical types, demonstrating significant technological advancement and feasibility.

[0028] This step serves to provide a data foundation and quality assurance for the entire perioperative risk prediction and intelligent intervention methods, ensuring that subsequent risk modeling and intervention recommendations are accurate, comprehensive, and clinically relevant. Perioperative patient conditions are often complex and variable, unlike general disease management which relies on a single data dimension or data from a specific stage. True perioperative management must cover three key stages: preoperative assessment, intraoperative monitoring, and postoperative recovery. By establishing a unified data collection mechanism, this step breaks down data barriers between multiple information systems within the hospital, enabling the collaborative acquisition of multi-dimensional, multimodal clinical information, from laboratory tests, vital signs, free-text records, imaging diagnoses to pathological analysis.

[0029] Specifically, this step not only automates the collection of data in different formats but also transforms the originally scattered, inconsistent, redundant, or noisy data into a unified, high-quality input dataset through standardization. This unified dataset effectively avoids problems such as "missing information," "inconsistent features," and "insufficient dimensionality" in traditional models, significantly improving the stability and generalization ability of subsequent fusion modeling. Simultaneously, fine-grained data labels and time-point markers ensure that the risk analysis results truly correspond to the patient's specific perioperative state, avoiding prediction errors caused by misaligned data timelines or semantic ambiguity.

[0030] More importantly, the panoramic data acquisition mechanism in this step allows the system to identify and utilize "weak signals" that are difficult to capture through traditional clinical experience. For example, a text message mentioning "intermittent fever one week before surgery" or a slight change in lung markings described in an imaging report can become an important predictor of postoperative infection or pulmonary complications. Only by establishing this solid, clearly structured, and semantically clear data collection system can we support subsequent steps in building clinically credible risk prediction and intervention strategies, thereby achieving truly intelligent and personalized perioperative management.

[0031] S200 standardizes the acquired structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results, including normalizing structured numerical indicators, parsing medical entities in text medical history records, extracting quantitative feature information from imaging reports and pathology reports, and forming a high-dimensional fusion feature vector with unified structure. To accurately predict the risk of perioperative complications and generate personalized intervention recommendations for high-risk patients, it is necessary to standardize the collected structured examination indicators, textual medical history records, medical imaging report conclusions, and pathological diagnosis results. This addresses issues such as inconsistent formats, non-standard units, and non-standardized expression methods among multi-source data. This processing transforms heterogeneous data into a unified, high-dimensional fusion feature vector by unifying fields, cleaning noise, extracting key features, and constructing a fusion data structure. This facilitates subsequent quantitative assessment and intervention strategy reasoning.

[0032] The structured examination indicators were normalized in terms of units and format. These structured examination indicators include laboratory test results and vital sign parameters. Specific fields include hemoglobin (g / L), white blood cell count (×10⁻⁶), etc. 9The following parameters were measured: neutrophil percentage (%), albumin (g / L), creatinine (μmol / L), fasting blood glucose (mmol / L), alanine aminotransferase (ALT) (U / L), aspartate aminotransferase (AST) (U / L), total bilirubin (μmol / L), D-dimer (mg / L), prothrombin time (seconds), international normalized ratio (INR), body temperature (°C), pulse (beats / minute), systolic blood pressure (mmHg), diastolic blood pressure (mmHg), respiratory rate (breaths / minute), and oxygen saturation (%). The units for each indicator were standardized, and the value range was limited to eliminate systematic errors or outliers. For example, D-dimer values ​​higher than 10 mg / L were uniformly set to 10 mg / L. For indicators collected at continuous time points, four statistical characteristics were extracted: maximum value, minimum value, average value, and rate of change within a fixed time window. For example, in the body temperature data measured within 48 hours before surgery, the maximum, minimum, average and hourly rate of change of body temperature (unit: ℃ / hour) can be calculated to capture the fever trend.

[0033] Medical entity recognition and structured representation were performed on text-based medical history records. These records included admission records, present illness, past medical history, personal history, family history, preoperative anesthesia visit records, preoperative multidisciplinary consultation opinions, and daily progress notes. The text content was written in natural language, resulting in inconsistencies in expression. The processing involved first establishing a standardized medical terminology table, covering standard disease names, drug names, surgical procedures, symptoms, familial genetic diseases, lifestyle descriptions, and commonly used clinical examinations. Non-standard terms, such as "hypertension," "high blood pressure," and "HTN," were uniformly mapped to the standard term "hypertension." Subsequently, sentence structures were identified based on grammatical rules, such as "having a medical history of ×× years," "having previously undergone ×× surgery," and "having ×× symptoms before surgery," and the disease name, duration of illness, onset time, treatment records, and control status were extracted. For example, if the statement "The patient has a 12-year history of type 2 diabetes, and their blood sugar is generally controlled with metformin" is identified, the following fields are extracted: Disease Name = "Diabetes", Type = "Type 2", Duration = "12 years", Control Status = "General", Treatment Method = "Oral hypoglycemic drugs". All extracted items are mapped to structured features and expanded into multi-valued fields.

[0034] Semantic analysis and quantitative feature extraction were performed on the conclusions of medical imaging reports. Medical imaging reports, issued by radiologists, generally include two parts: "Imaging Findings" and "Imaging Diagnosis," with a relatively standardized text structure. This invention selects the "Imaging Diagnosis" section as the main analysis object, extracting descriptive information highly relevant to perioperative complications through sentence structure rules and keyword matching. This includes the anatomical location of the lesion, the nature of the lesion, its size measurement, the clarity of its boundaries, enhancement characteristics, involvement of adjacent structures, and the preliminary radiological diagnosis. For example, in the diagnostic description "A soft tissue mass was seen in the sigmoid colon, approximately 4.2cm × 3.8cm in size, with heterogeneous enhancement, suggestive of adenocarcinoma," the extracted fields are: Lesion location = "sigmoid colon," Lesion type = "soft tissue mass," Maximum diameter = "4.2cm," Enhancement characteristics = "heterogeneous," Diagnostic judgment = "adenocarcinoma tendency." For size descriptions, a character extraction function was used to extract numerical values, and the units were uniformly converted to centimeters. The resulting fields were uniformly named and formatted into a structured format for merging with other features in calculations.

[0035] The pathological diagnosis results were processed by extracting diagnostic conclusions and converting the coding structure. Pathological diagnosis data came from preoperative biopsies, intraoperative frozen sections, and postoperative formal pathology reports. The reports included gross findings, microscopic structural descriptions, immunohistochemical staining results, and the final diagnosis. During processing, the final diagnosis segment was extracted first, obtaining pathological type, degree of differentiation, depth of invasion, lymph node metastasis, resection margin status, and expression of characteristic markers. For example, if the report stated "moderately differentiated adenocarcinoma, serosal invasion, lymph node metastasis 4 / 15, negative periintestinal resection margin, positive CDX2, negative CK7," then the extracted fields would be: Pathological type = "Adenocarcinoma", Differentiation grade = "Moderately differentiated", Depth of invasion = "Serosa", Number of positive lymph nodes = 4, Total number of lymph nodes = 15, Resection margin status = "Negative", CDX2 expression = "Positive", CK7 expression = "Negative". Immunohistochemical marker expression is represented by Boolean values ​​(positive = 1, negative = 0). If expression intensity is required, a three-level scoring system of "strong positive = 2, weak positive = 1, negative = 0" is introduced. All pathological fields are uniformly coded and named to ensure consistent integration with other data dimensions.

[0036] After the above processing, the structured examination data, structured text data, structured image description data, and structured pathological diagnosis data are concatenated according to the field order to construct a unique high-dimensional fusion feature vector for each patient. This fusion vector is expressed in a standard tabular format, with each row representing a single patient and each column representing a uniformly named feature field. The total number of features can reach hundreds depending on the actual project, and it features consistent fields, unified value ranges, clear types, and accurate semantics. This provides a high-quality and structurally stable input data foundation for subsequent risk probability prediction and intervention suggestion generation.

[0037] This step aims to unify, standardize, and unify the format of medical data from different sources and in various formats, thereby constructing a high-dimensional fusion feature vector that is complete, readable, and usable for computational analysis. This vector serves as the foundational data input for subsequent perioperative risk prediction and intelligent intervention generation. In clinical practice, various perioperative information of patients is usually scattered across multiple information systems. For example, laboratory test results are stored in the laboratory system, text records are stored in the electronic medical record system, imaging information is managed by the radiology department, and pathology reports are archived in the pathology information system. These data exhibit significant heterogeneity, including different data formats (e.g., numerical vs. text), inconsistent units (e.g., g / L vs. g / dL), non-standard language descriptions (e.g., mixed terminology), and inconsistent field definitions (e.g., differences in test names). Directly inputting this data into the risk assessment process will lead to a series of problems such as misaligned feature dimensions, variable conflicts, and semantic ambiguity, ultimately affecting the accuracy of model predictions and the credibility of intervention recommendations. Therefore, the goal of this step is to transform the raw data into a structured input set with consistent format, clear content, and stable fields through methods such as unit normalization, field alignment, semantic parsing, keyword recognition, feature extraction, and structured coding.

[0038] This fusion feature set encompasses the dynamic trends of structured numerical indicators, clinical elements in text records, diagnostic judgments in imaging conclusions, and entity expressions in pathology reports. It fully preserves the patient's multi-dimensional, multi-time-point, and cross-modal clinical status information, providing data-level assurance for subsequent prediction of different types of postoperative complications (such as thrombosis, infection, kidney injury, and myocardial infarction). More importantly, the feature vectors constructed in this step not only meet the needs of machine processing but also possess clinical semantic interpretability. This supports subsequent visualization of risk level distribution maps and automatic output of risk interpretations and personalized intervention suggestions by a large language model, thus constructing the core foundation of a closed-loop "data-prediction-intervention" path. Therefore, this step is a crucial link in this invention, connecting front-end data acquisition and supporting subsequent model evaluation and clinical decision output. It plays an irreplaceable foundational role, demonstrating the practicality, inventiveness, and non-obviousness of this invention in fusion processing technology.

[0039] S300 integrates feature vectors into the trained complication risk prediction model and outputs the probability values ​​of various perioperative complications to achieve quantitative prediction of complication risk. To achieve quantitative prediction of perioperative complication risks, previously standardized data needs to be integrated into a high-dimensional fusion feature vector, which is then input into an established complication risk prediction model. This model outputs the probability values ​​of multiple complications, ultimately constructing a personalized risk profile for each patient. This process comprises four stages: field consistency matching, risk prediction workflow input, standardized expression of probability values, and unified processing of prediction results.

[0040] Field filtering and consistency matching were performed on the fused feature vector. The fused feature vector is a unified data structure formed by standardizing and splicing structured examination indicators, textual medical history information, medical imaging report content and pathological diagnosis data. It contains multiple clinical fields from different sources, such as D-dimer value (mg / L) within 24 hours before surgery, intraoperative blood loss (mL), preoperative albumin concentration (g / L), age (years), body mass index (kg / m²), history of diabetes (yes / no), preoperative heart rate variation range (bpm), duration of preoperative fever (hours), presence or absence of history of venous thrombosis (yes / no), surgical procedure type (open / laparoscopic), tumor differentiation degree (poor / moderate / highly differentiated), and whether the preoperative CT report indicated a tendency for lung infection (yes / no), etc.

[0041] For each perioperative complication, such as postoperative venous thrombosis, postoperative incision infection, postoperative acute kidney injury, postoperative myocardial infarction, and postoperative pulmonary infection, a field list is established to clearly define the required input fields, units, and data format requirements. For example, the field list for predicting postoperative venous thrombosis includes: preoperative D-dimer value (mg / L), age (years), surgical procedure type (whether it is a high-risk procedure), history of thrombosis (yes / no), preoperative activity level score (1-5 points), albumin concentration (g / L), and presence of preoperative lower extremity edema. The system automatically verifies whether the input vector contains all necessary fields, checks whether the units match, the format is compliant, and the values ​​exceed limits, and removes data records that do not meet the requirements to ensure the consistency and accuracy of subsequent inputs.

[0042] The validated fusion feature vectors are input into the complication risk prediction model. Each type of complication has an independent prediction process, with its input order and field definitions pre-defined and aligned with the risk assessment needs in hospital clinical practice. For example, when predicting the risk of postoperative pulmonary infection, input fields include preoperative maximum body temperature (°C), preoperative C-reactive protein concentration (mg / L), whether imaging reports indicate increased lung markings (yes / no), preoperative smoking history (yes / no), intraoperative mechanical ventilation time (minutes), surgical procedure type (intrathoracic / intraperitoneal), and preoperative nutritional status score (1-5 points).

[0043] In this step, the "complication risk prediction model" refers to a mathematical modeling structure built using machine learning methods based on large historical clinical sample data. This model can quantitatively predict the probability of specific perioperative complications (such as postoperative infection, venous thrombosis, pulmonary infection, acute kidney injury, myocardial infarction, etc.) in patients. The model takes a standardized high-dimensional fusion feature vector as input, which integrates structured examination indicators (such as D-dimer, albumin, creatinine, etc.), medical history information (such as diabetes, history of thrombosis), imaging report conclusions (such as increased lung markings, presence of consolidation), and pathological diagnosis content (such as differentiation degree, lymph node metastasis). It outputs a specific risk probability value for each complication, such as "the probability of postoperative pulmonary infection is 23.7%" and "the probability of postoperative venous thrombosis is 15.2%".

[0044] The core role of this model in this invention is to transform complex, multimodal, and multi-source raw patient data into quantifiable and comparable risk assessment results, serving as the decision-making foundation for the entire intelligent risk identification and intervention process. The risk probability values ​​output by the model not only determine the patient's risk level for various postoperative complications but also provide crucial input for subsequent steps such as graphical risk level distribution display, natural language generation of explanations for high-risk causes, and the formulation of personalized intervention recommendations. Therefore, the complication risk prediction model constitutes the central engine of the "from data to decision" closed-loop structure of this invention, achieving a leap from traditional experience-based judgment to data-driven precise prediction and intervention management, and is one of the key technological foundations for the innovation and practicality of this invention.

[0045] Each field is filled in strictly according to the field name, unit, and value range during the input phase. The system does not accept input vectors with missing fields or incorrect field order. Taking a patient as an example, if the preoperative imaging report indicates increased lung markings, a persistently elevated preoperative body temperature for more than 48 hours, and intraoperative ventilation time reaches 180 minutes, the prediction process will uniformly process these input values ​​and output the probability value of postoperative lung infection. For example, the final output result might be 23.7%, which serves as the basis for the next step of risk stratification.

[0046] The probability values ​​for each complication are graded and standardized. Different types of complications vary significantly in severity and intervention timeliness; therefore, this invention pre-defines multiple risk level threshold systems to convert continuous probability results into clinically interpretable risk levels. For example, for postoperative venous thrombosis risk, 0%-5% is defined as "low risk," 5%-15% as "medium risk," 15%-25% as "high risk," and above 25% as "very high risk." For postoperative infection risk, 0%-10% is "low risk," 10%-20% as "medium risk," 20%-30% as "high risk," and above 30% as "very high risk." All grading standards are set before deployment based on real clinical data distribution and expert consensus and are stored in the system parameter configuration table. Each probability output is accompanied by a risk level label, while the original probability value is retained for subsequent visualization and clinical review.

[0047] Finally, all prediction results were compiled into a structured checklist to generate a comprehensive perioperative risk profile for the patient. The checklist format is a standardized two-dimensional table, with fields including complication name, risk probability value (%), risk level, prediction timestamp, and a summary of the source of the characteristic data. For example, the checklist records the following: Postoperative venous thrombosis risk = 21.3%, level = high risk; Postoperative incision infection risk = 14.2%, level = medium risk; Postoperative acute kidney injury risk = 6.9%, level = medium risk; Postoperative pulmonary infection risk = 23.7%, level = high risk; Postoperative myocardial infarction risk = 3.5%, level = low risk. This risk profile not only provides a basis for subsequent causal explanation and intervention recommendations but also can be displayed in real-time on the doctor's workstation interface for preoperative discussions, patient informed consent communication, and postoperative monitoring decisions, improving the standardization and foresight of perioperative management.

[0048] The core function of this step is to transform structured and fused multi-source, multimodal clinical data into quantifiable probabilities of perioperative complication risk, enabling comprehensive, quantitative, and interpretable predictions of individual patient risk. This provides a data-driven scientific basis for the development of subsequent personalized intervention strategies. In traditional clinical pathways, physicians primarily rely on experience or simple scoring systems, such as the Caprini score and ASA classification, to assess postoperative complication risk. While these methods are effective in some routine scenarios, they often fail to accurately identify potential high-risk points in complex or atypical cases due to their inability to comprehensively analyze large-scale, multivariate, and dynamically changing patient data. Furthermore, they lack the ability to characterize individualized risks.

[0049] This step involves encoding standardized structured examination indicators (such as D-dimer, albumin, body temperature, and heart rate), textual medical history (such as the duration of diabetes and past thrombosis), imaging diagnostic information (such as CT findings of increased lung markings and lesion description), and pathological diagnostic data (such as the degree of differentiation of adenocarcinoma and lymph node metastasis) into a high-dimensional feature vector. This vector is then input into a pre-defined prediction process focusing on specific complications (such as postoperative infection, thrombosis, myocardial infarction, pneumonia, and kidney injury) to derive the probability value for each type of complication, accurately quantifying the patient's risk level for that complication. For example, the model might output a patient's postoperative thrombosis risk of 21.3%, infection risk of 14.2%, and lung infection risk of 23.7%, assigning corresponding risk levels (such as high risk or medium risk). This mechanism not only significantly improves the objectivity and accuracy of risk assessment but also makes the assessment results clinically applicable, allowing them to be directly used for preoperative discussions, informed consent, and monitoring decisions.

[0050] Furthermore, this step converts continuous probability values ​​into discrete risk levels by setting grading standards, facilitating subsequent graphical display and document template integration. Ultimately, the risk prediction results not only provide doctors with clear reference information but also serve as input for the intelligent system to generate explanatory content and intervention recommendations, achieving an intelligent closed loop from "risk identification" to "intervention response." Therefore, this step plays a crucial role in this invention, serving as a core bridge connecting multimodal data processing and clinical decision support, possessing significant practical technical value and innovation.

[0051] S400 generates a graphical risk level distribution map based on the probability value of complication occurrence, and generates natural language descriptions to explain the causes of high risk based on key feature information in the fused feature vector. After obtaining the risk probability values ​​for various perioperative complications, to improve the clinical visibility and interpretability of the risk assessment results, the risk level of each complication needs to be displayed graphically. Based on key feature information affecting the prediction results, a natural language explanation paragraph with clinical semantic integrity is generated to assist physicians in accurately identifying high-risk causes and developing contingency plans. This process mainly includes the following four steps: risk level graphical generation, key feature identification, semantic structure construction, and natural language generation and output.

[0052] Based on the predicted probability value of each complication, a graphical risk level distribution map is generated. Specifically, after predicting the major perioperative complications (including postoperative incision infection, postoperative venous thrombosis, postoperative pulmonary infection, postoperative myocardial infarction, and postoperative acute kidney injury), the system classifies the probability values ​​of each complication according to a preset risk level classification standard. For example, the classification standard for postoperative pulmonary infection is set as follows: a predicted probability in the range of 0% to 10% is marked as "low risk," 10% to 20% as "medium risk," 20% to 30% as "high risk," and greater than 30% as "very high risk." The aforementioned risk levels are visualized using a radar chart to construct an overall risk map. The map uses complication type as the dimension and standard risk level conversion values ​​(e.g., low risk = 1, medium risk = 2, high risk = 3, very high risk = 4) as the radius scale to form the outer contour of the graph. Different colors are used to distinguish the degree of risk (e.g., green for low risk, yellow for medium risk, orange for high risk, and red for very high risk), thus intuitively reflecting the patient's risk distribution across different complication directions. This graph can be integrated into preoperative discussion platforms and physician terminal pages for rapid assessment during multidisciplinary consultations.

[0053] From the fusion feature vector of the input prediction process, the key feature variables with the greatest impact on the prediction results are identified. This fusion feature vector consists of multiple structurally unified fields, sourced from structured numerical indicators (such as D-dimer, albumin, creatinine, and systolic blood pressure), parsed text information (such as history of diabetes, preoperative fever, and history of thrombosis), medical imaging report results (such as CT scan showing increased lung markings and consolidation), and pathological diagnosis report content (such as differentiation degree and whether there is lymph node metastasis). After determining that a complication is at a high or very high risk level, the system will identify variables with significant contributions from the feature fields corresponding to the prediction process for that complication. For example, when postoperative lung infection is assessed as "high risk," key features may include a continuous increase in body temperature for more than 48 hours before surgery (temperature range of 37.8~38.9℃), a C-reactive protein concentration of 86 mg / L (above the upper limit of normal), a CT report clearly indicating "increased lower lobe markings with a few patchy shadows," and intraoperative ventilation time exceeding 3 hours. These features are all compared with the current value using standard value range judgment rules or clinical guideline recommended value ranges to determine whether they are abnormal and to serve as candidates for high-risk factors.

[0054] The semantic structure required for natural language description is constructed to achieve causal logical expression between key features and high-risk outcomes. The generation process of natural language description adopts a structure template-driven approach, setting a fixed sentence structure and filling in the following in sequence: ① high-risk event ontology, ② key triggering features, ③ co-aggravating factors, and ④ comprehensive risk warning. For example, for the case of a high risk of "postoperative pulmonary infection", the constructed semantic structure is as follows: "The patient has [key triggering factors]: such as a persistently elevated body temperature before surgery (maximum temperature 38.9℃, lasting for 48 hours), accompanied by [co-aggravating factors]: preoperative C-reactive protein level increased to 86 mg / L, CT scan showed abnormal lung structure (increased texture and patchy shadows in the lower lobes of both lungs), and the surgical ventilation time was as long as 185 minutes. Based on the above factors, it is suggested that the patient has a high risk of postoperative pulmonary infection." This structure ensures grammatical standardization and clinical semantic accuracy through a fixed sentence structure, while also ensuring that different content can be generated for different patients, making it feasible for batch implementation.

[0055] The graphical risk distribution map and natural language interpretation results are compiled and output together as a risk assessment report integrated into the electronic medical record system for clinicians' reference. Each output report contains two parts: the first part is a risk radar chart, showing the risk level of each complication, with specific probability values ​​and level labels; the second part is a text description, listing the causes of each high-risk or very high-risk complication, covering information such as abnormal numerical indicators, imaging findings, preoperative medical history, and intraoperative conditions, and expressing the causes of risk and logical reasoning in complete sentences. For example, the report content includes: "The predicted probability of postoperative pulmonary infection is 26.4%, with a risk level of high risk. The main reasons include persistent preoperative fever, prolonged intraoperative ventilation time, and abnormal imaging findings in the CT report. It is recommended to strengthen postoperative pulmonary function monitoring and infection prevention intervention." This information can be written into preoperative discussion records, informed consent forms, or postoperative management plans through an interface to assist doctors in making scientific and reasonable clinical decisions.

[0056] The main purpose of this step is to transform the raw quantitative risk prediction results into more clinically interpretable and operational graphical information and natural language content, thereby improving the readability, acceptability, and decision-making support capabilities of the prediction results in clinical practice. In traditional risk assessment methods, even if a certain risk score or probability value can be obtained, clinicians often still need to spend a lot of time interpreting the causes, judging the influencing factors, and inferring possible interventions based on experience. This increases the cognitive burden and also carries the risk of judgment bias due to differences in individual experience. Therefore, relying solely on predicted probability values ​​as the basis for clinical decision-making is insufficient; it is necessary to use structured and semantic methods to deeply interpret and express the results.

[0057] This step in the invention constructs a standardized risk grading system, mapping numerical risk probability results to a unified grading framework (e.g., low risk, medium risk, high risk, very high risk), and displays the risk level of each type of postoperative complication graphically (e.g., radar charts, risk bar charts), allowing doctors to intuitively perceive the risk differences of patients across various dimensions. Simultaneously, by identifying and extracting key influencing variables from the high-dimensional fusion feature vector, such as abnormally elevated preoperative D-dimer levels, a history of thrombosis, persistently elevated preoperative body temperature, and imaging reports indicating lung lesions, and combining the causal logical relationships between these variables and specific complications, a semantically clear and structurally complete natural language description paragraph is constructed. This interpretation method not only covers the values, sources, and clinical significance of specific variables but also clearly indicates their contribution path to increased risk, enabling doctors to "understand the causes" rather than merely "see the results," thus improving clinical trust and acceptance.

[0058] Furthermore, the graphic and textual explanations generated in this step can be directly embedded into key documents such as preoperative discussion records, patient informed consent forms, postoperative follow-up plans, and nursing documentation, achieving a natural connection and systematic expression between predicted results and clinical interventions. In other words, this step is not only a transformation of information presentation format, but also a "bridge mechanism" for transitioning from data analysis to clinical application. It concretizes, visualizes, and semanticizes abstract risk prediction results, truly realizing "interpretability," "operability," and "implementability" in AI-assisted clinical decision-making. It is one of the key links in realizing the intelligent closed-loop decision-making process of this invention, demonstrating high technical practicality and non-obvious innovative points.

[0059] S500: When the risk probability value of a certain complication reaches a preset threshold, a corresponding intervention plan is generated based on the high-risk cause information and medical knowledge base content. The intervention plan includes medication use recommendations, monitoring items and frequencies, nursing pathways and rehabilitation guidance. In the perioperative patient risk prediction process, when the predicted probability of a postoperative complication exceeds a set intervention threshold, the system will automatically generate a targeted, operable, and personalized intervention plan based on the aforementioned high-risk cause information and a pre-built structured medical knowledge base. This intervention plan aims to provide clear and standardized risk management recommendations in clinical practice, covering medication use recommendations, key monitoring items and frequencies, nursing pathway arrangements, and postoperative rehabilitation guidance. The specific implementation process includes the following steps: Based on the model prediction results from the previous stage, high-risk complications involving the current patient are identified, and it is confirmed whether the intervention trigger conditions are met. Judgment is made based on the set intervention thresholds. For example, if the intervention threshold for postoperative pulmonary infection risk is set at 20%, and a patient's predicted probability for this complication is 23.5%, then the intervention trigger condition is met. Subsequently, several fusion feature vector elements that contribute the most to the prediction result of this complication are automatically extracted as high-risk causal information. These features include, but are not limited to: continuous fever for more than 48 hours (maximum value reaching 38.9℃), preoperative C-reactive protein of 84 mg / L (10 mg / L above the upper limit of normal), chest CT report indicating "increased lung texture and scattered patchy shadows in the lower lobes of both lungs", and surgical procedure of general anesthesia + open abdominal surgery (operation time expected to be greater than 180 minutes), etc. All features are clearly extracted and structurally represented through the aforementioned standardized steps.

[0060] Based on the identified high-risk causes, interventions highly relevant to the complication were retrieved from a structured medical knowledge base. The knowledge base is indexed by disease type and risk factors, and includes intervention strategies for different clinical scenarios. Each entry includes the intervention name, specific implementation method, appropriate conditions, timing of implementation, duration, and precautions regarding adverse reactions. For example, for high-risk cases of pulmonary infection, the knowledge base includes the following interventions: ① Preoperative administration of prophylactic antibiotic cefuroxime sodium 1.5g intravenous drip, repeated every 12 hours postoperatively for a total of three times; ② Daily monitoring of body temperature, C-reactive protein, and white blood cell count for three days postoperatively; ③ Encouraging patients to perform independent deep breathing exercises in bed for 6 hours postoperatively, 10 times per hour, and starting bedside sitting-up exercises 12 hours postoperatively; ④ Encouraging patients to get out of bed and move around for 48 hours postoperatively, with each walk ≥10 meters, ≥2 times per day. The patient's current causal characteristics were compared with the appropriate conditions of the above intervention entries, and only intervention recommendations that met the conditions were retained.

[0061] The selected intervention items were summarized, structured, and translated into language to form intervention recommendation texts that are readable and actionable for clinical staff. This text is organized according to four intervention types: medication recommendations, monitoring recommendations, nursing recommendations, and rehabilitation recommendations. Within each type of intervention, the text includes the intervention objective, recommended measures, start time, frequency of implementation, duration, and necessary precautions. For example: "

Drug Recommendation

Monitoring Recommendation

Nursing Recommendation

Rehabilitation Recommendation

[0062] The aforementioned intervention recommendations were transformed into a structured output format and integrated into relevant document templates within the hospital information system. Specific integration locations include preoperative discussion record templates, postoperative medical order recommendation templates, nursing pathway plans, and rehabilitation plan record forms. The system automatically identifies the target role based on the implementing entity; for example, medication recommendations are assigned to the attending physician, nursing recommendations to the responsible nurse, and rehabilitation recommendations to the rehabilitation therapist. The recommendation text will be inserted as a suggestion box in the corresponding document node, accompanied by prompts such as "Source of Generation: Triggered by Complication Risk Prediction" and "Related Event: Postoperative Lung Infection Risk is 23.5%", allowing clinical staff to retain traceability information during reading, review, and modification. Clinical staff can choose to adopt, modify, or replace the recommendation content according to the actual situation, and all manual modification operations are recorded for auditing purposes.

[0063] The purpose of this step is to automatically generate individualized, structured, and actionable clinical intervention recommendations based on the identification of patients with a high risk of postoperative complications. This transforms risk prediction results into intervention behaviors with direct guiding significance, bridging the key link from "risk identification" to "risk control" and constructing a truly closed-loop intelligent auxiliary decision-making process. In traditional medical practice, even if some high-risk patients can be identified, doctors often still need to rely on personal experience or consult a large number of literature guidelines to manually formulate intervention measures. This not only increases the clinical workload but may also lead to missing, delayed, or non-standard implementation of interventions due to knowledge lag, experience differences, or inconsistent judgments, seriously affecting the quality of perioperative management and patient safety.

[0064] This invention automatically filters out information such as medication recommendations, monitoring items and frequencies, nursing pathways, and postoperative rehabilitation arrangements corresponding to the current patient's risk status by matching high-risk causal information (such as abnormal test values, medical history elements, and imaging conclusions) with structured medical knowledge base content (such as guideline recommendations, consensus opinions, and real clinical pathways). This information is then organized into a complete intervention recommendation text. These recommendations include elements such as time points, dosage, frequency, objectives, and target population, and can be directly integrated into preoperative discussion records, postoperative medical order systems, or nursing documents for doctors to quickly access or modify, greatly improving the efficiency and accuracy of intervention formulation.

[0065] More importantly, this step breaks through the limitations of the traditional disconnect between "identification and suggestion," achieving automatic connection from prediction to action. It not only improves the standardization and normalization of perioperative management but also strengthens the application of artificial intelligence systems in real-world medical scenarios, representing a key path for realizing the deeper clinical value of smart healthcare. Therefore, this step has extremely high practical value, innovation, and non-obviousness in this invention, and is one of the core technical components for achieving intelligent closed-loop control of postoperative complications.

[0066] S600 integrates intervention plans into the preoperative discussion template or postoperative medical order template of the electronic medical record system, realizing an automated closed loop for perioperative risk identification and intervention recommendations. To achieve automated closed-loop linkage between perioperative risk prediction results and actual intervention implementation, the aforementioned automatically generated intervention plan content needs to be formatted according to the field structure requirements of the electronic medical record system and integrated into the preoperative discussion record template or postoperative medical advice template. The entire integration process includes the following steps: Each recommendation in the intervention plan is broken down and structured to meet the interface specifications and field entry rules for different types of clinical documents in the hospital's electronic medical record system. The intervention plan mainly includes information such as medication administration recommendations, vital sign monitoring frequency, nursing operation guidelines, and postoperative rehabilitation activity plans. Taking a patient at high risk of postoperative venous thrombosis as an example, the intervention recommendation "Administer 40 mg of enoxaparin sodium subcutaneously once daily for 5 consecutive days postoperatively" is broken down as follows: intervention category field: "Medication Use"; specific medication field: "enoxaparin sodium"; dosage field: "40 mg"; route of administration field: "subcutaneous injection"; frequency field: "once daily"; start time field: "immediately after surgery"; course of treatment field: "5 days". All fields have clear definitions and value ranges, facilitating entry into the predefined data tables of the electronic medical record system (EMR).

[0067] Based on the patient's current perioperative stage, the system determines whether the targeted intervention recommendation should be embedded in the preoperative discussion template or the postoperative medical order template. For patients still in the preoperative assessment stage, the intervention recommendation will be written into the "Risk Assessment and Management Recommendations" section of the preoperative discussion template, displayed as a recommended statement in the system-generated template, such as: "Based on the predicted risk of postoperative thrombosis (predicted value 28.7%), it is recommended to administer enoxaparin sodium 40 mg subcutaneously once daily for 5 days, starting 12 hours preoperatively." When the patient enters the postoperative management stage, the system automatically identifies that the item should be written into the "Pre-set Medication Orders" area of ​​the postoperative medical order template. The template will display it as a standardized medical order: "Postoperative day 0, 06:00, subcutaneous injection of enoxaparin sodium 40 mg, once daily, for 5 days." The above content will automatically enter the hospital's medical order signing system, where doctors can directly click "Accept" or "Adjust and Accept."

[0068] To ensure the traceability of intervention recommendations throughout the entire process and the compliance with clinical implementation, the system will be linked and updated with the nursing pathway execution table and rehabilitation treatment record table. Specifically, if some pre-recommendations contain nursing-related content, such as "wear medical compression stockings within the first 6 hours after surgery, and perform leg raises in bed 3 times a day for 15 minutes each time during bed rest," a clear execution instruction field will appear in the nursing record table, with "completed / incomplete / partially completed" options in the "Nursing Plan Execution Status" column. Simultaneously, for postoperative rehabilitation recommendations, such as "start assisted walking training on the second day after surgery, with a cumulative time of no less than 20 minutes per day," a patient schedule reminder will be generated in the rehabilitation record table, along with recording the rehabilitation therapist's execution feedback. All content is written to the database through standard fields to ensure consistent data format and searchable content.

[0069] Below each intervention suggestion embedded in the electronic medical record, an additional generation description is provided, including the source path, prediction basis, and triggering conditions. This clearly informs clinicians that the suggestion originates from the intelligent prediction process, enhancing the transparency and credibility of the intervention suggestions. An example prompt reads: "This suggestion was generated by the perioperative intelligent risk identification and intervention system. Based on a postoperative venous thrombosis prediction probability of 28.7%, exceeding the preset intervention threshold (25%), combined with high-risk factors such as the patient's D-dimer level of 5.2 mg / L, preoperative bed rest duration of 5 days, and history of thrombosis." Furthermore, all suggestions can be "adopted," "modified," or "rejected" by physicians or relevant personnel. The actions are automatically recorded in the "Intervention Suggestion Response Log" of the electronic medical record, documenting the operator, time, modification content, and reason, ensuring that the entire medical process is auditable, verifiable, and archived.

[0070] This step serves to automate the closed-loop transformation of perioperative risk prediction results into actual clinical interventions, and is the key final link in this invention's "intelligent identification - personalized intervention - implementation" process. Currently, many perioperative risk assessment systems or models face the problem of "identifying but not executing" during clinical implementation, failing to effectively translate prediction results into clinical operational instructions. Consequently, prediction results remain merely at the reference level, and intervention measures still rely on doctors' manual judgment and prescription writing, resulting in low efficiency, poor standardization, and a high risk of overlooking high-risk individuals. This invention, however, structures the personalized intervention plan content generated by the system (such as drug names, administration methods, monitoring indicators, nursing measures, and rehabilitation suggestions) and automatically writes it into the preoperative discussion template or postoperative prescription template of the electronic medical record system according to the patient's stage, thus completely opening up the data channel and behavioral collaboration between the intelligent system and the medical information system.

[0071] This integrated approach allows doctors to directly access system-provided suggestions within the native context of medical record writing or prescription signing, offering the convenience of "direct adoption and one-click signing." All suggestions include source information, triggering criteria, and risk factor warnings, ensuring doctors maintain autonomy in medical decision-making while enabling rapid decision-making based on structured evidence. Furthermore, this process supports automatic linkage between nursing record sheets and rehabilitation plan sheets, ensuring that intervention suggestions not only exist in medical records but are also effectively implemented by healthcare professionals in different roles, improving team collaboration efficiency and the quality of the medical closed-loop system.

[0072] Therefore, the value of implementing this step lies not only in "presenting" intelligent results to doctors, but more importantly, in building an automatic flow mechanism from data prediction to intervention execution. It deeply embeds intelligent prediction capabilities into the medical workflow, enabling the entire perioperative risk management system to have self-driven, self-closed-loop, and self-tracking capabilities. It is the basic support and practical guarantee for realizing the closed-loop logic of the technical solution of this invention, and has significant innovation and practical value.

[0073] The multimodal data fusion perioperative risk prediction and intervention method and system provided by this invention have significant beneficial effects: This method introduces for the first time a systematic integration and standardized modeling mechanism of multi-source heterogeneous medical data in perioperative management, breaking the dependence of existing assessment methods on a small amount of structured data and human experience, and can comprehensively, dynamically, and accurately depict the patient's clinical status; By constructing a high-dimensional feature expression that integrates structured test indicators, free text medical history, imaging conclusions, and pathological results, and inputting it into a trained complication risk prediction model, the system can achieve quantitative prediction of the risk of various postoperative complications (such as thrombosis, infection, kidney injury, myocardial infarction, etc.); Further combined with an interpretable mechanism, the system can automatically generate natural language explanations based on key influencing factors, effectively improving the clinical understanding and acceptance of the prediction results; When the risk level reaches a preset threshold, the system automatically matches appropriate intervention measures from the medical knowledge base, constructs a personalized intervention plan, covering drug recommendations, nursing operations, monitoring frequency, and rehabilitation pathways, and automatically embeds it into key nodes of the electronic medical record system in a structured manner, realizing a closed-loop management of the entire process from risk identification, cause interpretation to intervention execution. This invention significantly improves the intelligence, standardization, and personalization of perioperative risk management, helping to reduce the incidence of postoperative complications and improve patient safety and medical quality.

[0074] This invention provides, for example Figure 2 The multimodal data fusion perioperative risk prediction and intervention system shown includes a data acquisition module, a feature construction module, a risk prediction module, a risk interpretation module, an intervention generation module, and an intervention execution closed-loop module. The data acquisition module establishes a data acquisition mechanism to obtain perioperative-related data from multiple medical information systems. The data includes structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results. The feature construction module standardizes the acquired structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results. This includes normalizing structured numerical indicators, parsing medical entities in text medical history records, extracting quantitative feature information from imaging reports and pathology reports, and forming a high-dimensional fusion feature vector with a unified structure. The risk prediction module integrates the feature vector inputs into the complication risk prediction model after training, and outputs the probability values ​​of various perioperative complications to achieve quantitative prediction of complication risk. The risk interpretation module generates a graphical risk level distribution map based on the probability value of complication occurrence, and generates natural language descriptions to explain the causes of high risk based on key feature information in the fused feature vector. The intervention generation module generates a corresponding intervention plan based on high-risk cause information and medical knowledge base content when the risk probability value of a certain complication reaches a preset threshold. The intervention plan includes medication use recommendations, monitoring items and frequencies, nursing pathways and rehabilitation guidance. The intervention execution closed-loop module integrates the intervention plan into the preoperative discussion template or postoperative medical order template of the electronic medical record system, realizing the automated execution closed loop of perioperative risk identification and intervention recommendations.

[0075] The perioperative risk prediction and intervention method based on multimodal data fusion provided in this embodiment of the invention is implemented through the aforementioned perioperative risk prediction and intervention system based on multimodal data fusion. For details of the specific methods and procedures of the perioperative risk prediction and intervention system based on multimodal data fusion, please refer to the embodiments of the aforementioned perioperative risk prediction and intervention method based on multimodal data fusion, which will not be repeated here.

[0076] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A perioperative risk prediction and intervention method based on multimodal data fusion, characterized in that, Includes the following steps: S100: Establish a data collection mechanism to obtain perioperative-related data from multiple medical information systems. The data includes structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results. S200 standardizes the acquired structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results, including normalizing structured numerical indicators, parsing medical entities in text medical history records, extracting quantitative feature information from imaging reports and pathology reports, and forming a high-dimensional fusion feature vector with unified structure. S300 integrates feature vectors into the trained complication risk prediction model and outputs the probability values ​​of various perioperative complications to achieve quantitative prediction of complication risk. S400 generates a graphical risk level distribution map based on the probability value of complication occurrence, and generates natural language descriptions to explain the causes of high risk based on key feature information in the fused feature vector. S500: When the risk probability value of a certain complication reaches a preset threshold, a corresponding intervention plan is generated based on the high-risk cause information and medical knowledge base content. The intervention plan includes medication use recommendations, monitoring items and frequencies, nursing pathways and rehabilitation guidance. S600 integrates intervention plans into the preoperative discussion template or postoperative medical order template of the electronic medical record system, realizing an automated closed loop for perioperative risk identification and intervention recommendations.

2. The perioperative risk prediction and intervention method based on multimodal data fusion according to claim 1, characterized in that, Step S100 includes: Obtain structured examination indicators, including blood routine test indicators, biochemical indicators, coagulation function indicators and vital signs parameters; Obtain textual medical history records, including admission records, present illness history, past medical history, personal history, family history, preoperative anesthesia visit records, preoperative discussion records, multidisciplinary consultation opinions, and progress notes, and identify medical entity information through terminology standardization processing; Obtain the conclusions of medical imaging reports, including qualitative descriptions of lesion location, nature, size, boundary, enhancement features, and surrounding relationships in imaging diagnoses, and extract structured fields; Obtain pathological diagnostic results, including tumor type, differentiation degree, depth of invasion, lymph node metastasis, resection margin status, and expression of immunohistochemical indicators, and construct a standardized set of structural fields.

3. The perioperative risk prediction and intervention method based on multimodal data fusion according to claim 1, characterized in that, Step S200 includes: The structured inspection indicators are normalized and outliers are removed, and the maximum, minimum, average and rate of change in the time series are extracted. The text medical history records are standardized in terms of terminology and parsed in terms of grammatical rules. Disease names, duration of disease, control status and treatment records are extracted and converted into structured fields. The conclusions of medical imaging reports are analyzed by keyword matching and sentence structure parsing to extract the lesion location, nature, size, boundary features, and preliminary diagnostic results, and then the results are transformed into a structured form. Key fields were extracted and the coding structure was transformed from the pathological diagnosis results. Pathological type, degree of differentiation, depth of invasion, lymph node status, resection margin, and expression of immunohistochemical markers were extracted to construct a structured feature set in a unified format.

4. The perioperative risk prediction and intervention method based on multimodal data fusion according to claim 1, characterized in that, Step S300 includes: The fields in the fused feature vector are filtered and matched for consistency to ensure that the units of each field are consistent, the format is standardized, and the values ​​are valid. Input the validated fusion feature vectors into the complication risk prediction model according to the preset field order, and output the risk probability values ​​of each perioperative complication. The probability values ​​of each risk are standardized and expressed, and corresponding risk level labels are generated based on preset classification thresholds; The complication name, risk probability value, risk level, prediction timestamp, and feature source information are compiled into a structured list to generate a comprehensive risk profile of the patient during the perioperative period.

5. The perioperative risk prediction and intervention method based on multimodal data fusion according to claim 1, characterized in that, Step S400 includes: Based on the risk probability value of each complication and the preset threshold standard, risk levels are classified and a risk level distribution map in the form of a radar chart is drawn. Identify key features from the fused feature vectors that influence the prediction results of high-risk complications; Based on high-risk events, key triggering characteristics, synergistic aggravating factors, and comprehensive risk warnings, a semantic structure is constructed to generate natural language description content. The graphical risk level distribution map and natural language description are simultaneously compiled into a risk assessment report and embedded into the electronic medical record system.

6. The perioperative risk prediction and intervention method based on multimodal data fusion according to claim 5, characterized in that, In the step of identifying key feature information, the feature data in the structured examination indicators, parsed text medical history records, imaging report conclusions and pathological diagnosis results are compared with the standard risk feature value range of the corresponding complications. Variables with abnormal values ​​or high-risk semantic representations are screened out and sorted according to their degree of influence on the prediction results of complication risk, which are then used for the generation of subsequent natural language description content.

7. The perioperative risk prediction and intervention method based on multimodal data fusion according to claim 1, characterized in that, Step S500 includes: Determine whether the risk probability value of a certain complication exceeds a preset threshold; if it does, trigger the intervention condition. Key features that influence the prediction of this complication risk are extracted as high-risk factors. Match medical knowledge base entries corresponding to the high-risk cause and screen for intervention measures that meet the adaptation criteria; The system is structured and generated into text based on four categories: medication use recommendations, monitoring items and frequencies, nursing pathways, and rehabilitation guidance. The output is readable and actionable intervention recommendations.

8. The perioperative risk prediction and intervention method based on multimodal data fusion according to claim 1, characterized in that, Step S600 includes: Each suggestion in the intervention plan is broken down into fields and mapped into a structured format; Depending on the patient's perioperative stage, intervention recommendations will be written into either the preoperative discussion template or the postoperative medical order template. Nursing recommendations and rehabilitation recommendations will be updated in conjunction with the nursing record sheet and rehabilitation record sheet; Attach a generation description to each suggestion and record the doctor's response behavior in the intervention suggestion response log.

9. A perioperative risk prediction and intervention system based on multimodal data fusion, used to implement the perioperative risk prediction and intervention method based on multimodal data fusion as described in any one of claims 1-8, characterized in that, It includes a data acquisition module, a feature construction module, a risk prediction module, a risk interpretation module, an intervention generation module, and an intervention execution closed-loop module: The data acquisition module establishes a data acquisition mechanism to obtain perioperative-related data from multiple medical information systems. The data includes structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results. The feature construction module standardizes the acquired structured examination indicators, text medical history records, medical imaging report conclusions, and pathological diagnosis results. This includes normalizing structured numerical indicators, parsing medical entities in text medical history records, extracting quantitative feature information from imaging reports and pathology reports, and forming a high-dimensional fusion feature vector with a unified structure. The risk prediction module integrates the feature vector inputs into the complication risk prediction model after training, and outputs the probability values ​​of various perioperative complications to achieve quantitative prediction of complication risk. The risk interpretation module generates a graphical risk level distribution map based on the probability value of complication occurrence, and generates natural language descriptions to explain the causes of high risk based on key feature information in the fused feature vector. The intervention generation module generates a corresponding intervention plan based on high-risk cause information and medical knowledge base content when the risk probability value of a certain complication reaches a preset threshold. The intervention plan includes medication use recommendations, monitoring items and frequencies, nursing pathways and rehabilitation guidance. The intervention execution closed-loop module integrates the intervention plan into the preoperative discussion template or postoperative medical order template of the electronic medical record system, realizing the automated execution closed loop of perioperative risk identification and intervention recommendations.

Images