Gynecological general anesthesia postoperative complication prediction modeling method based on pathological data analysis

Abstract

The invention discloses a gynecological general anesthesia postoperative complication prediction modeling method based on pathological data analysis, and relates to the technical field of postoperative complication prediction, and the method comprises the following steps: collecting preoperative, intraoperative and postoperative multi-dimensional pathological and clinical data of gynecological general anesthesia; establishing a space-time incidence matrix of anxiety psychology, intraoperative pathology and complications; and constructing a gynecological general anesthesia postoperative complication prediction model based on a multi-modal deep learning prediction architecture based on the space-time incidence matrix. Through combination of multi-dimensional data fusion, dynamic correlation analysis and deep learning, transformation from an empirical decision to a data-driven decision is realized. The time-space incidence matrix provides standardized risk assessment logic for doctors; the prediction capability of the model on rare complications fills the prediction blank in a small sample scene. Finally, through real-time risk early warning and individualized intervention suggestions, the safety of the gynecological general anesthesia operation is remarkably improved, the postoperative recovery time of a patient is shortened, and the method has important clinical transformation value.

Description

Technical Field

[0001] The present invention relates to the technical field of postoperative complication prediction, and in particular to a modeling method for predicting gynecological postoperative complications under general anesthesia based on pathological data analysis. Background Art

[0002] Currently, research on the prediction of complications after gynecological general anesthesia mainly focuses on the correlation analysis between surgical types, such as hysterectomy and ovarian surgery, and traditional clinical indicators. Prediction models are mainly constructed based on routine pathological data such as patients' underlying diseases, such as hypertension, diabetes, operation duration, anesthetic drug dosage, and intraoperative blood loss. Commonly used methods include multivariate logistic regression and COX proportional hazard model.

[0003] Most of the focus is on physiological indicators, such as the impact of intraoperative hemodynamics and laboratory tests on complications, but insufficient attention is paid to patients' psychological factors, such as preoperative anxiety. Especially in specific gynecological surgeries such as female oophorectomy, research on the relationship between psychological factors and postoperative complications is almost blank.

[0004] Although psychological assessment tools such as the State Anxiety Inventory (SAI) have been proven to be feasible in thyroid surgery and elderly patients, their application in patients undergoing gynecological surgery under general anesthesia has not yet been widely used, and their synergistic predictive value with pathological indicators has not been fully explored.

[0005] Most existing prediction models are general models that lack individualized designs for gynecological surgeries, such as oophorectomy, and do not fully consider the physiological characteristics of gynecological patients, such as hormone levels, pelvic anatomy, and the specific relationship between complications.

[0006] Existing models mostly rely on basic indicators such as operation duration and anesthetic dosage, ignoring the interaction between psychological factors (such as preoperative anxiety) and pathological indicators (such as the cumulative impact of intraoperative blood pressure fluctuations caused by anxiety on complications), resulting in incomplete feature coverage.

[0007] Insufficient gynecological specificity: The model was not designed for the particularities of gynecological surgery (such as adnexectomy). For example, pathological data related to pelvic surgery (such as the degree of pelvic adhesions and the amount of bleeding during surgery) were not included, and there was insufficient exploration of the association between female patients' hormone levels, reproductive history, and complications.

[0008] Model performance limitations: Traditional statistical methods (such as univariate analysis and binary logistic regression) are mostly used, which have weak ability to capture nonlinear characteristics (such as the nonlinear relationship between anxiety score and postoperative pain threshold). The prediction accuracy (such as AUC value) and stability need to be improved.

[0009] Insufficient clinical interpretability: Although some machine learning models (such as black box models) can improve prediction accuracy, it is difficult to clearly define the impact of each pathological feature on complications and cannot provide specific targets for clinical intervention. Summary of the Invention

[0010] In order to solve the above technical problems, the present invention provides a prediction modeling method for gynecological postoperative complications after general anesthesia based on pathological data analysis. The following technical solutions are adopted:

[0011] The prediction modeling method for complications after gynecological general anesthesia based on pathological data analysis includes the following steps:

[0012] Step 1: Collect multidimensional pathological and clinical data before, during, and after gynecological general anesthesia;

[0013] Step 2: Establish a spatiotemporal correlation matrix among anxiety, intraoperative pathology, and complications;

[0014] Step 3: Based on the spatiotemporal correlation matrix, a gynecological general anesthesia postoperative complication prediction model based on a multimodal deep learning prediction architecture was constructed;

[0015] In step 4, the complication prediction model inputs the multi-dimensional pathological and clinical data of the current patient before, during, and after surgery, and outputs the type of complication and the probability of its occurrence.

[0016] Optionally, the multi-dimensional pathological and clinical data before gynecological general anesthesia in step 1 include basic clinical index data, pathological molecular marker detection data, and psychological status quantitative data;

[0017] Multi-dimensional pathological and clinical data during gynecological general anesthesia include real-time physiological index data, intraoperative pathological index data, and surgical operation parameters;

[0018] Multidimensional pathological and clinical data after gynecological general anesthesia surgery include complication records and repair and infection marker detection data.

[0019] Optionally, in step 2, the time is divided into several key time nodes;

[0020] The spatial dimension is defined as several types of associated subjects, and several groups of core associated pairs are formed based on the combination of representative indicators of several types of associated subjects.

[0021] Optionally, in step 2, the spatiotemporal correlation matrix uses several key time nodes as row labels and several groups of core correlation pairs as column labels. The initial values ​​of the matrix are all 0, and the correlation coefficients obtained through subsequent statistical calculations are used to fill the matrix cells to form a complete spatiotemporal correlation matrix.

[0022] Optional, several key time nodes are: 24 hours before surgery, after anesthesia induction, 1 hour after surgery, at the end of surgery, and 24 hours after surgery;

[0023] Several categories of related subjects are: three categories of related subjects, namely psychological state, pathological indicators and physiological reactions;

[0024] Representative indicators of psychological status are anxiety assessment, heart rate variability, and galvanic skin response;

[0025] Representative pathological indicators include preoperative inflammatory factors, inflammation grade, and high-mobility group protein B1 detection value;

[0026] Representative indicators of physiological responses are blood pressure fluctuation amplitude, heart rate variability and BIS value.

[0027] Optional, specific statistical calculation process of the correlation coefficient:

[0028] Step a: extracting the associated pair data of the corresponding time period according to the time node from the standardized time series database;

[0029] Step b: If the correlation pair is a continuous variable and a continuous variable, a normality test is performed first. If it conforms to the normal distribution, the Pearson correlation coefficient is used; if it is a continuous variable and a rank variable, the Spearman rank correlation coefficient is used, and the rank correlation is calculated by sorting the data;

[0030] Step c: perform a t-test on the calculated correlation coefficient and calculate the P value. Calculate the above steps for each of the five key time nodes.

[0031] Step d: Perform clinical logic verification on the statistical results. If any results conflict with medical common sense, they will be revised based on the opinions of clinical experts.

[0032] Optionally, in step 3, the model architecture of the complication prediction model includes an input layer, a fusion layer, and an output layer;

[0033] The input layer includes static data channel, time series data channel and image data channel;

[0034] The fusion layer introduces an attention mechanism based on the spatiotemporal correlation matrix: dynamic attention weights are assigned to the indicators involved in the core correlation pairs;

[0035] The output layer is designed based on multi-task learning to simultaneously predict multiple types of complications and risk probabilities;

[0036] The Sigmoid function was used to predict the probability of a single complication, and the Softmax function was used to rank multiple complications.

[0037] Optionally, step 5 is also included, generating clinical intervention recommendations based on the complication prediction results.

[0038] Optionally, the generation of clinical intervention recommendations in step 5 includes the following steps:

[0039] The complication prediction model outputs the complication risk probability and the contribution of each feature, sets the contribution threshold, screens out the core risk factors, and sorts them in descending order of contribution to form a risk factor priority list;

[0040] Establish a mapping library of risk factors and intervention measures, including: pathological indicator intervention, psychological state intervention, and physiological indicator intervention;

[0041] The corresponding intervention measures are called from the mapping library according to the risk factor type.

[0042] Optionally, genetic polymorphism correction factors are introduced to adjust intervention measures in combination with the patient's basic characteristics.

[0043] In summary, the present invention has the following beneficial technical effects:

[0044] This invention can provide a predictive modeling method for gynecological postoperative complications under general anesthesia based on pathological data analysis. By combining multi-dimensional data fusion, dynamic correlation analysis, and deep learning, it achieves a shift from empirical decision-making to data-driven decision-making. On the one hand, the spatiotemporal correlation matrix provides doctors with a standardized risk assessment logic; on the other hand, the model's ability to predict rare complications fills the prediction gap in small sample scenarios. Ultimately, through real-time risk warnings and personalized intervention recommendations, the safety of gynecological general anesthesia surgery has been significantly improved, and the patient's postoperative recovery time has been shortened, which has important clinical translational value. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of the modeling method for predicting complications after gynecological general anesthesia based on pathological data analysis of the present invention. DETAILED DESCRIPTION

[0046] The present invention will be further described in detail below with reference to the accompanying drawings.

[0047] The embodiment of the present invention discloses a modeling method for predicting complications after gynecological general anesthesia based on pathological data analysis.

[0048] Reference Figure 1 Example 1: A method for predicting complications after gynecological general anesthesia based on pathological data analysis, comprising the following steps:

[0049] Step 1: Collect multidimensional pathological and clinical data before, during, and after gynecological general anesthesia;

[0050] Step 2: Establish a spatiotemporal correlation matrix among anxiety, intraoperative pathology, and complications;

[0051] Step 3: Based on the spatiotemporal correlation matrix, a gynecological general anesthesia postoperative complication prediction model based on a multimodal deep learning prediction architecture was constructed;

[0052] In step 4, the complication prediction model inputs the multi-dimensional pathological and clinical data of the current patient before, during, and after surgery, and outputs the type of complication and the probability of its occurrence.

[0053] By employing this technical solution, Step 1 overcomes the limitations of traditional reliance on a single clinical indicator by collecting multidimensional pathological data from pre-, intra-, and postoperative periods, such as inflammatory factors, genotyping, and frozen section grading, along with clinical data such as SAI scores, physiological indicators, and complication records. The integration of full-cycle time-series data not only provides a comprehensive picture of the patient's psychological state, pathological characteristics, and physiological responses, but also ensures data consistency through timestamps and standardization. This provides high-quality, multidimensional foundational material for subsequent correlation analysis and model training, addressing the problem of biased predictions caused by traditional data fragmentation.

[0054] In step 1, the standardized processing of multi-dimensional pathological and clinical data before, during, and after gynecological general anesthesia includes the following steps:

[0055] Data type and source:

[0056] Preoperative data included: basic clinical indicators, age, BMI, medical history, etc., which were entered into the electronic medical record system through structured data; pathological molecular marker detection data, such as IL-6 and CRP, which were detected using electrochemiluminescence / PCR fluorescent probe methods, with the results synchronized to the laboratory information system; psychological status quantitative data (SAI scores were digitally collected through the WeChat mini-program, and HRV indicators were recorded by wearing a wrist heart rate variability monitor simultaneously).

[0057] Intraoperative data include: real-time physiological indicator data, blood pressure and heart rate are recorded once every 5 minutes, BIS values ​​are continuously sampled at 1Hz and automatically uploaded to the time series database through the anesthesia monitor, intraoperative pathological indicator data (frozen section inflammation grading is blindly reviewed by two pathologists and quantitatively analyzed by a digital pathology system; HMGB1 is detected using a portable microfluidic chip, and the results are output within 15 minutes and synchronized to the intraoperative database), and surgical operation parameters (operation type, duration, and blood loss are recorded through surgical procedure coding and dual measurement methods).

[0058] Postoperative data include: complication records, using the Clavien-Dindo classification, joint evaluation 24 / 48 / 72 hours after surgery and entry into the system, repair and infection marker detection data, VEGF, PCT, etc., collected at time nodes and dynamic ratios calculated.

[0059] Data unification method:

[0060] Time alignment: Add timestamps to all data, anchoring them to key time nodes, such as 24 hours before surgery and after anesthesia induction, to ensure temporal consistency;

[0061] Format conversion: Unstructured data, such as pathological images, are converted into DICOM format; text data, such as medical history, are converted into structured labels using natural language processing; and physiological signals are converted into time series arrays;

[0062] Standardized database construction: Build a unified time-series database to integrate data from various systems, including electronic medical records, anesthesia monitors, laboratory systems, etc., achieve real-time synchronization through API interfaces, and unify the data storage format into timestamp + data type + value + unit, supporting fast retrieval by time node or data type.

[0063] The spatiotemporal correlation matrix constructed in step 2 quantifies the strength of associations between anxiety, such as SAI scores, intraoperative pathological indicators such as HMGB1, and inflammation grade, and complication risk at different time points. This systematically reveals for the first time the cascade mechanism of psychological state, pathological changes, physiological responses, and complications. For example, the synergistic effect of a high preoperative SAI score and an intraoperative inflammation grade of 2 or higher on postoperative agitation is clearly identified, providing interpretable association rules for the predictive model. This eliminates the model output from a black box result and instead allows for causal inference based on clinical logic, enhancing the credibility of the prediction results.

[0064] Step 2: The specific steps for establishing the spatiotemporal correlation matrix of anxiety, intraoperative pathology, and complications are as follows:

[0065] Definition of space-time dimension:

[0066] Time dimension:

[0067] It is divided into 5 key time nodes, namely 24 hours before surgery, after anesthesia induction, 1 hour after surgery, at the end of surgery, and 24 hours after surgery. It is based on the pathophysiological laws of gynecological surgery, covering the preoperative basic state, intraoperative stress peak, and early postoperative reaction stage.

[0068] Spatial Dimensions:

[0069] Three types of related subjects are defined: psychological state, pathological indicators, and physiological reactions. Their representative indicators are as follows:

[0070] Psychological status: anxiety assessment value (SAI score), heart rate variability (LF / HF ratio), skin galvanic response;

[0071] Pathological indicators: preoperative inflammatory factors (IL-6, CRP), inflammation grade (inflammatory cell density in frozen sections), high mobility group protein B1 (HMGB1) detection value;

[0072] Physiological reactions: blood pressure fluctuation amplitude, heart rate variability, BIS value (depth of anesthesia).

[0073] Core association pairs: Based on the pairwise combination of representative indicators of the three types of subjects, 9 core association pairs are formed (such as "SAI score-IL-6", "inflammation grade-blood pressure fluctuation amplitude", "HMGB1-BIS value", etc.), covering the psychological-pathological, pathological-physiological, and psychological-physiological interactions.

[0074] Matrix construction process:

[0075] Matrix structure: 5 key time nodes are used as row labels, 9 core association pairs are used as column labels, and the initial values ​​are all 0;

[0076] Correlation coefficient calculation:

[0077] Step a: Extracting association pair data from the standardized time series database by time node, such as SAI score and IL-6 value 24 hours before surgery;

[0078] Step b: Select the calculation method based on the data type. For continuous variables, such as SAI score and IL-6, the Pearson correlation coefficient is used when they conform to the normal distribution. The Spearman rank correlation coefficient is used for continuous variables and ordinal variables.

[0079] Step c: Perform a t-test on the correlation coefficient and calculate it separately at 5 time points;

[0080] Step d: Clinical logic verification. If the results conflict with medical common sense, they will be revised based on the expert opinions of three gynecologists with the title of deputy director or above.

[0081] Matrix filling: Fill the verified correlation coefficients into the corresponding cells to form a complete spatiotemporal correlation matrix. For example, the correlation coefficient of "inflammation grade-blood pressure fluctuation amplitude" within 1 hour of surgery is 0.62, indicating that the two are significantly positively correlated at this node.

[0082] Step 3: A gynecological postoperative complication prediction model for general anesthesia was constructed based on the spatiotemporal association matrix. This model integrates ResNet50 pathological image features, ST-GCN spatiotemporal physiological signals, and a fully connected layer of static clinical data. By embedding an attention mechanism through association rules, the model accurately captures the dynamic impact of key risk factors. Compared with traditional methods such as logistic regression or Apfel scoring, this complication prediction model is highly sensitive to complication warnings and can update risk probabilities in real time every 30 minutes during surgery, addressing the limitation of traditional static predictions that cannot adapt to the dynamic changes of the surgical process.

[0083] The specific architecture and construction process of the gynecological postoperative complications prediction model are as follows:

[0084] Model architecture: including input layer, fusion layer and output layer;

[0085] Input layer: contains three parallel channels;

[0086] Static data channel: Input basic clinical indicators, age, BMI, gene polymorphisms (HTR2A, COMT genotype) and other non-time series data, and convert them into feature vectors through the fully connected layer;

[0087] Time series data channel: Input intraoperative physiological signals, blood pressure, BIS value, dynamic pathological indicators, HMGB1 time series changes, and use the spatiotemporal graph convolutional network (ST-GCN) to extract time series features;

[0088] Image data channel: Input pathological slice images and extract texture features such as inflammatory cell distribution through the ResNet50 network.

[0089] Fusion layer: introduces an attention mechanism based on the spatiotemporal correlation matrix;

[0090] Based on the correlation coefficient of the spatiotemporal correlation matrix, dynamic weights are assigned to the indicators involved in the core correlation pairs.

[0091] The features of the three channels are fused into a unified feature vector through the attention layer.

[0092] Output layer: designed based on multi-task learning;

[0093] The Sigmoid function was used to independently predict single-category complications, such as agitation and nausea and vomiting, with the probability of occurrence ranging from 0 to 1;

[0094] The Softmax function is used to sort the risks of multiple types of complications and output the risk priority, such as the risk priority of nausea and vomiting is higher than that of pain and agitation.

[0095] Model training and optimization:

[0096] Training data: Multi-dimensional data of 1,000 gynecological general anesthesia surgery cases were used, 80% for training and 20% for validation;

[0097] Loss function: Cross entropy loss function is used, combined with clinical constraints of spatiotemporal correlation matrix;

[0098] Optimizer: Use the Adam optimizer with a dynamically adjusted learning rate, initially at 0.001, and decaying by 10% every 10 rounds until the validation set AUC stabilizes above 0.92.

[0099] Step 4 outputs specific complication types (such as agitation, nausea and vomiting) and their corresponding risk probabilities, providing clear targets for clinical decision-making. Combining the association rules from step 2 with the feature contribution analysis from step 3, doctors can directly identify high-risk causes and develop targeted intervention plans. This closed-loop process of risk identification, cause analysis, and action matching not only improves the timeliness of clinical intervention but also reduces the risk of overtreatment or underintervention by tailoring individualized characteristics such as genetic polymorphisms, ultimately lowering the incidence of complications.

[0100] This method, through the integration of multidimensional data fusion, dynamic correlation analysis, and deep learning, achieves a shift from empirical to data-driven decision-making. On the one hand, the spatiotemporal correlation matrix provides physicians with standardized risk assessment logic; on the other hand, the model's ability to predict rare complications fills the gap in predictions for small sample sizes. Ultimately, through real-time risk warnings and personalized intervention recommendations, the safety of gynecological general anesthesia surgery has been significantly improved, postoperative recovery time has been shortened, and the approach has significant clinical translational value.

[0101] In Example 2, the multi-dimensional pathological and clinical data before gynecological general anesthesia in step 1 include basic clinical indicator data, pathological molecular marker detection data, and psychological state quantitative data;

[0102] Multi-dimensional pathological and clinical data during gynecological general anesthesia include real-time physiological index data, intraoperative pathological index data, and surgical operation parameters;

[0103] Multidimensional pathological and clinical data after gynecological general anesthesia surgery include complication records and repair and infection marker detection data.

[0104] By adopting the above technical solution, the collection time of basic clinical indicator data is completed 24-48 hours before surgery, through the combination of electronic medical record system and manual input;

[0105] Specific indicators and standards:

[0106] Demographic characteristics: Age, accurate to years; BMI: weight / height², retain one decimal place, fasting for 8 hours before measurement;

[0107] Medical history information: A structured questionnaire was used to record information, including the type of gynecological disease, such as uterine fibroids / ovarian cancer, surgical history (number of pelvic surgeries in the past 5 years), preoperative medication history (name and dosage of anticoagulants / hormonal drugs), and allergy history (allergies to anesthetic drugs should be marked with the specific drug and reaction);

[0108] Reduce prediction bias caused by missing medical history;

[0109] The data collection time for pathological molecular marker detection is 12-24 hours before surgery, 5 ml of venous blood (EDTA anticoagulant tube);

[0110] Testing indicators and methods:

[0111] Serum inflammatory factors: IL-6 (electrochemiluminescence method, detection limit 0.5 pg / ml), CRP (high-sensitivity latex-enhanced immunoturbidimetry, range 0.1-100 mg / L), detected by Roche Cobase601;

[0112] Stress gene polymorphisms: HTR2A and COMT genotyping was performed using the PCR fluorescent probe method (ABI7500 real-time fluorescence quantitative PCR instrument), and the primer sequences were verified by the NCBI database;

[0113] Genotyping results provide a reliable molecular basis for subsequent association analysis;

[0114] Quantitative data of psychological state:

[0115] Collection time: 24 hours before surgery, completed in a quiet clinic;

[0116] Specific methods:

[0117] SAI Anxiety Scale: Digitally collected through the WeChat mini-program, the total score (0-60 points) is automatically calculated. Auxiliary physiological signals: A wrist-worn heart rate variability (HRV) monitor is simultaneously worn to record 5 minutes of resting state data and extract the LF / HF ratio (sympathetic nerve activity indicator);

[0118] Data association: Binding the SAI score with the HRV index to form a combined psychological-physiological data set;

[0119] The quantitative dimension of psychological state has been expanded from a single scale to subjective scores plus objective physiological signals, and the correlation strength with intraoperative stress response has been enhanced.

[0120] Data collection during gynecological general anesthesia:

[0121] Real-time physiological indicator data:

[0122] Monitoring frequency: Blood pressure and heart rate are recorded every 5 minutes, and BIS and time of flight (TOF) values ​​are continuously sampled (1 Hz) and automatically uploaded to the database via an anesthesia monitor (e.g., Philips IntelliVue MX800).

[0123] Circulatory indicators: systolic blood pressure / diastolic blood pressure / mean arterial pressure, calculation of fluctuation amplitude;

[0124] Anesthesia status indicators: BIS value (target range 40-60, reflecting the depth of anesthesia), TOF ratio (recovery to 0.9 is considered as muscle relaxation);

[0125] Anesthetic medication: Real-time recording of propofol / sevoflurane dosage (mg / kg / h) and total amount of opioids (converted to morphine equivalent);

[0126] Physiological data time provides an accurate timing basis for real-time pathological-physiological correlation analysis.

[0127] Intraoperative pathological index data:

[0128] Frozen section inflammation grade:

[0129] Sampling time: within 30 minutes after adnexectomy, three tissues (cortex / medullary / junctional area) were collected for sectioning;

[0130] Grading process: Two pathologists blindly review the samples, and a digital pathology system (such as 3DHISTECH) is used to quantitatively analyze the density of inflammatory cells;

[0131] HMGB1 detection:

[0132] Time points: 30 minutes after anesthesia induction, after key surgical steps (such as tumor resection), and before abdominal closure, for a total of 3 times;

[0133] Portable microfluidic chip detection (detection time 15 minutes, lower limit 0.1ng / ml), with results automatically synchronized to the intraoperative database;

[0134] Dynamic detection of HMGB1 captured the intraoperative peak time point (occurring on average 28 minutes after tumor resection), providing key time node data for the correlation matrix;

[0135] Surgical operation parameters:

[0136] Recording criteria: type of surgery (laparoscopic / open) was recorded using the procedure code, and the duration of the surgery was recorded to the nearest minute (from skin incision to abdominal closure).

[0137] Bleeding volume measurement: double check using the suction fluid volume - flushing fluid volume + gauze weighing method;

[0138] The correlation analysis between surgical parameters and pathological indicators provided a basis for introducing surgical procedure weights into the model;

[0139] Data collection after gynecological general anesthesia:

[0140] Complications recorded:

[0141] Recording criteria: Clavien-Dindo classification (grades I-V) was used, and the patient was evaluated jointly by nurses and doctors 24 / 48 / 72 hours after surgery;

[0142] Specific indicators: agitation: Riker scale is scored once every hour, and a score greater than or equal to 5 is defined as agitation;

[0143] Nausea and vomiting: PONV grading (0-3), record the time of first occurrence and duration;

[0144] Pain: VAS score (0-10 points) was recorded every 6 hours, and intervention was required if the score was greater than or equal to 4 points;

[0145] Standardization of complication grading makes data comparable across centers, and the sensitivity of subsequent models in identifying severe complications of grade III or above has been greatly improved.

[0146] Repair and infection marker detection data:

[0147] Testing time: venous blood was collected 24h, 48h, and 72h after surgery to test VEGF (vascular repair), TGF-β (tissue repair), and PCT (infection warning);

[0148] Dynamic analysis: Calculate the 48h / 24h ratio. A ratio greater than 1.2 indicates delayed repair or infection risk.

[0149] The dynamic changes of markers are significantly associated with complications, providing the model with enhanced indicators of postoperative risk;

[0150] Multi-dimensional data covers the entire cycle of pre-operative, intra-operative and post-operative, which expands the feature dimension of the prediction model in step 3 and increases the AUC value to.

[0151] Real-time data (such as intraoperative HMGB1) enables early warning of complications during surgery (compared to traditional postoperative evaluation), reducing the incidence of severe postoperative complications.

[0152] In Example 3, in step 2, the time is divided into several key time nodes;

[0153] The spatial dimension is defined as several types of associated subjects, and several groups of core associated pairs are formed based on the combination of representative indicators of several types of associated subjects.

[0154] In Example 4, in step 2, the spatiotemporal correlation matrix uses several key time nodes as row labels and several groups of core correlation pairs as column labels. The initial values ​​of the matrix are all 0. The correlation coefficients obtained by statistical calculation are subsequently used to fill the matrix cells to form a complete spatiotemporal correlation matrix.

[0155] In Example 5, several key time points are: 24 hours before surgery, after anesthesia induction, 1 hour after surgery, at the end of surgery, and 24 hours after surgery;

[0156] Several categories of related subjects are: three categories of related subjects, namely psychological state, pathological indicators and physiological reactions;

[0157] Representative indicators of psychological status are anxiety assessment, heart rate variability, and galvanic skin response;

[0158] Representative pathological indicators include preoperative inflammatory factors, inflammation grade, and high-mobility group protein B1 detection value;

[0159] Representative indicators of physiological responses are blood pressure fluctuation amplitude, heart rate variability and BIS value.

[0160] By adopting the above technical scheme, we selected 24 hours before surgery, after anesthesia induction, 1 hour after surgery, at the end of surgery, and 24 hours after surgery as key nodes, and designed the following based on the pathophysiological laws of gynecological general anesthesia surgery:

[0161] 24 hours before surgery: Capture the patient's baseline status (such as chronic inflammation level and anxiety baseline) to provide a reference for intraoperative changes;

[0162] After anesthesia induction: reflects the initial effects of anesthetic drugs on the physiological state, when psycho-patho-physiological interactions begin to emerge;

[0163] One hour after surgery: Pathological reactions caused by surgical trauma (such as the release of inflammatory factors) reach their peak, which is a key window for early warning of complication risks;

[0164] At the end of surgery: the cumulative trauma effect and the residual anesthetic drug are superimposed, which is directly related to early postoperative complications;

[0165] 24 hours after surgery: a turning point in tissue repair and infection risk, during which the effectiveness of intraoperative interventions can be assessed;

[0166] Three types of related subjects are divided: Based on the psychological-pathological-physiological triple interaction mechanism, psychological state (anxiety) affects pathological indicators (such as inflammatory factors) through the neuro-endocrine-immune network, and pathological indicators and physiological reactions (such as blood pressure fluctuations) are directly related to the occurrence of complications, forming a complete chain of psychological stress, pathological changes, physiological disorders, and complications.

[0167] Representative indicator selection:

[0168] Psychological state: The anxiety assessment value (SAI scale) reflects the subjective anxiety level, and heart rate variability and skin galvanic response serve as objective physiological evidence, achieving dual quantification of subjective and objective factors;

[0169] Pathological indicators: Preoperative inflammatory factors (IL-6 / CRP) reflect the basic inflammatory state, intraoperative inflammation grade reflects local tissue damage, and HMGB1 detection value assesses ischemia-reperfusion injury, covering the three-level pathological dimensions of systemic, local, and cellular.

[0170] Physiological response: Blood pressure fluctuation amplitude and heart rate reflect circulatory stability, and BIS value monitors the depth of anesthesia. The three together constitute the core indicators of anesthesia safety;

[0171] The logic for forming core association pairs: Adopting the principles of intra-subject verification and inter-subject cross-pollination, similar subject indicators (such as heart rate variability, which is both psychological and physiological) are used to verify consistency, and cross-subject indicator combinations (such as anxiety assessment value-inflammatory factor-BIS value) are used to capture interactive relationships. Ultimately, 9 groups of core association pairs are formed to achieve comprehensive coverage of multi-dimensional influences.

[0172] The matrix uses time nodes as the vertical axis and core association pairs as the horizontal axis, and quantifies the dynamic association strength through correlation coefficients. Its essence is to transform the abstract interaction of psychology, pathology and physiology into a visual mathematical model.

[0173] Time dimension: capturing the temporal nature of the association through the changes in coefficients at different nodes (e.g., the effect of anxiety on inflammation is strongest 24 hours before surgery and weakens 24 hours after surgery);

[0174] Spatial dimension: Key driving factors are identified by the differences in coefficients of association pairs across subjects (e.g., the association strength between intraoperative inflammation grade and blood pressure fluctuation is usually higher than that between anxiety and BIS value).

[0175] Example 6, specific statistical calculation process of correlation coefficient:

[0176] Step a: extracting the associated pair data of the corresponding time period according to the time node from the standardized time series database;

[0177] Step b: If the correlation pair is a continuous variable and a continuous variable, a normality test is performed first. If it conforms to the normal distribution, the Pearson correlation coefficient is used; if it is a continuous variable and a rank variable, the Spearman rank correlation coefficient is used, and the rank correlation is calculated by sorting the data;

[0178] Step c: perform a t-test on the calculated correlation coefficient and calculate the P value. Calculate the above steps for each of the five key time nodes.

[0179] Step d: Perform clinical logic verification on the statistical results. If any results conflict with medical common sense, they will be revised based on the opinions of clinical experts.

[0180] By employing the above technical solution, we extract association pair data from a standardized time series database by time node, essentially achieving data alignment through time anchoring. Due to the varying duration and pace of gynecological surgery among different patients, it's difficult to unify analysis dimensions solely based on absolute time (e.g., 3 hours post-surgery). However, using key time nodes, such as the 24 hours pre-operatively and the 1 hour post-surgery period, as benchmarks ensures clinical comparability of association pair data at the same time node. For example, all patients' SAI scores and blood pressure data after anesthesia induction are at the same stage of diagnosis and treatment.

[0181] Continuous variable × continuous variable (such as preoperative inflammatory factor IL-6 and blood pressure fluctuation amplitude): The Pearson correlation coefficient is applicable to data with a linear relationship and a normal distribution. It can accurately quantify the strength of the linear association between variables (range: -1 to 1). Its mathematical principle is based on the ratio of covariance to standard deviation, reflecting the degree of coordinated change of variables.

[0182] Continuous variable × ordinal variable (such as SAI score and inflammation grade): The Spearman rank correlation coefficient avoids the non-continuous and non-normal characteristics of ordinal variables by converting the data into ranks. It is more suitable for describing the monotonic association between ordered categorical data and continuous data, such as the consistency of the trend between increased inflammation grade and increased HMGB1 detection value.

[0183] The t-test determines whether the strength of an association is due to random error by calculating the t-statistic and P-value corresponding to the correlation coefficient. A P-value less than 0.05 indicates statistical significance, ruling out the possibility of random error. Calculating each of the five key nodes can capture dynamic changes in the strength of the association. For example, inflammation grade and heart rate showed a significant correlation one hour after surgery, but no correlation 24 hours before surgery.

[0184] Statistical results may be affected by sampling bias, leading to conclusions that conflict with medical common sense, such as a strong correlation between increased inflammation levels and decreased blood pressure. Clinical logic verification incorporates expert experience to distinguish true weak associations from false associations, ensuring that statistical results align with pathophysiological principles. For example, inflammatory responses are often accompanied by increased blood pressure, thus achieving a balance between statistical significance and clinical plausibility.

[0185] In Example 7, in step 3, the model architecture of the complication prediction model includes an input layer, a fusion layer, and an output layer;

[0186] The input layer includes static data channel, time series data channel and image data channel;

[0187] The fusion layer introduces an attention mechanism based on the spatiotemporal correlation matrix: dynamic attention weights are assigned to the indicators involved in the core correlation pairs;

[0188] The output layer is designed based on multi-task learning to simultaneously predict multiple types of complications and risk probabilities;

[0189] The Sigmoid function was used to predict the probability of a single complication, and the Softmax function was used to rank multiple complications.

[0190] By adopting the above technical solutions, static data reflects the patient's basic condition, time series data captures dynamic changes during surgery, and image data provides pathological morphological information. The three types of data complement each other to cover the complete feature dimensions required for prediction.

[0191] The attention mechanism embeds clinical priors and guides weight allocation through the association matrix, allowing the model to automatically focus on strongly correlated features with clear clinical significance, such as the high correlation pair between inflammation grade and HMGB1, avoiding the black box defects of data-driven models.

[0192] The Sigmoid function independently predicts the probability of various complications, and the Softmax function implements risk ranking, taking into account the accuracy of single indicators and the priority requirements of clinical decision-making.

[0193] Multimodal fusion improves the AUC value compared to a single data model, especially the early warning sensitivity of complications reaches over 85%.

[0194] Attention weight visualization can directly display key information such as the contribution of inflammation grading characteristics, improving doctors' acceptance.

[0195] The lightweight design of LSTM and ResNet50 makes single-sample prediction take less than 0.5 seconds, meeting the real-time warning requirements during surgery.

[0196] By constraining the model's learning direction through an association matrix, the model's performance decay rate during cross-center validation was less than 5%. This architecture achieves a dual-wheel drive of data-driven and clinical knowledge-guided development, leveraging the advantages of deep learning for feature extraction while ensuring that prediction logic aligns with medical cognition through association rules, providing reliable support for clinical decision-making.

[0197] Example 8 further includes step 5 of generating clinical intervention recommendations based on the complication prediction results.

[0198] In Example 9, the generation of clinical intervention suggestions in step 5 includes the following steps:

[0199] The complication prediction model outputs the complication risk probability and the contribution of each feature, sets the contribution threshold, screens out the core risk factors, and sorts them in descending order of contribution to form a risk factor priority list;

[0200] Establish a mapping library of risk factors and intervention measures, including: pathological indicator intervention, psychological state intervention, and physiological indicator intervention;

[0201] The corresponding intervention measures are called from the mapping library according to the risk factor type.

[0202] Example 10: Introducing gene polymorphism correction factors to adjust intervention measures in combination with the patient's basic characteristics.

[0203] By adopting the above technical solution, we can see that the occurrence of complications is the result of the combined effects of multiple factors, and the contribution of different risk factors varies significantly. For example, inflammation grade may have a greater impact on postoperative infection than heart rate fluctuations. By setting a contribution threshold, such as 10% or greater, to screen core risk factors and sort them by contribution, we can focus on the factors that most significantly impact complications and avoid the dispersion of intervention resources. Essentially, this approach uses interpretable algorithms such as SHAP values ​​to convert the importance of black-box features output by the model into clinically understandable risk weights, ensuring that intervention measures directly target the core causes.

[0204] The core of establishing a risk factor-intervention mapping library is causal association matching based on evidence-based medicine. For example:

[0205] Pathological indicators, such as the mapping of inflammation grade ≥2 to anti-inflammatory measures (glucocorticoids), are derived from the pathological mechanisms of excessive release of inflammatory factors, tissue damage, and complications;

[0206] Psychological states, such as the mapping of SAI scores greater than 40 to sedative medications, are based on the physiological logic of anxiety, sympathetic nervous system excitation, elevated stress hormones, and increased risk of complications;

[0207] The mapping of physiological indicators to vasoactive drugs is based on the clinical consensus of hemodynamic instability, organ hypoperfusion, and delayed recovery.

[0208] The establishment of a mapping library ensures the direct correlation between intervention measures and risk causes, avoiding blind medication or operations.

[0209] Genetic polymorphisms (such as COMTVal158Met and HTR2Ars6311) can significantly affect drug metabolism and neuroendocrine responses, leading to significant differences in the effectiveness of the same intervention across different patients (e.g., patients with the Met / Met genotype are more sensitive to opioids). By adjusting for baseline patient characteristics (age, BMI, and underlying medical conditions), group intervention plans can be further refined into personalized plans. This principle aligns with the core principle of precision medicine, which optimizes treatment based on individual biological characteristics.

[0210] By screening core risk factors, 30%-40% of non-critical factor interventions can be reduced, avoiding the waste of wide-coverage medical resources.

[0211] Interventions based on the evidence-based mapping library are directly matched to risk factors. For example, the use of dexamethasone in patients with grade 3 inflammation can reduce the incidence of postoperative agitation by more than 40%. After adjusting for genetic polymorphisms, the incidence of adverse drug reactions (such as nausea and vomiting) is further reduced.

[0212] The structured intervention recommendation generation process significantly shortens doctors' decision-making time, especially in emergency scenarios during surgery (such as sudden blood pressure fluctuations). It can quickly identify core risks and push corresponding measures to buy time for rescue.

[0213] Combining genetic correction factors with adjustments to baseline characteristics allows intervention plans to be tailored to the physiological characteristics of different patients. For example, reducing the opioid dose by 20% for elderly patients with COMTMet / Met syndrome can maintain analgesia while reducing the risk of respiratory depression, improving clinical suitability by over 50%.

[0214] The following uses specific embodiments to illustrate the implementation principle of the present invention:

[0215] The patient was a 45-year-old female who was scheduled to undergo laparoscopic ovarian cystectomy for ovarian cyst. She was ASA grade II and had no history of hypertension or diabetes. She had not used hormonal drugs before the operation.

[0216] Step 1: Multi-dimensional data collection:

[0217] Preoperative data:

[0218] Basic clinical indicators: age 45 years, BMI 23.5 kg / m², gynecological disease type is benign cyst, no history of pelvic surgery, no history of allergy to anesthetics.

[0219] Pathological molecular markers: IL-6 = 12.3 pg / ml (electrochemiluminescence method), CRP = 8.5 mg / L (high-sensitivity latex method) detected 18 hours before surgery; HTR2A genotype was C / C, COMT genotype was Met / Met.

[0220] Psychological status: Preoperative 24h SAI score was 42 points (collected by WeChat applet), and synchronous HRV detection LF / HF=2.1 (sympathetic nerve activity was slightly high).

[0221] Intraoperative data:

[0222] Real-time physiological indicators: BIS value after anesthesia induction was 45, blood pressure fluctuation range was 8 mmHg; heart rate was 78 beats / min 1 hour after surgery, and propofol dose was 4 mg / kg / h.

[0223] Intraoperative pathological indicators: Frozen sections were completed 25 minutes after appendectomy, and the inflammation grade was grade 2 (5-10 inflammatory cells / HPF) by blind review by two pathologists; the HMGB1 detection value was 6.2 ng / ml 1 hour after surgery (microfluidic chip method).

[0224] Surgical operation parameters: laparoscopic surgery, duration 95 minutes, blood loss 50ml.

[0225] Postoperative data:

[0226] Complications recorded: 6 h after surgery, Riker scale score 4 points (no agitation), PONV grade 1 (mild nausea), VAS score 3 points (no intervention required).

[0227] Repair markers: VEGF = 350 pg / ml 24 hours after surgery, 48 hours / 24 hours ratio = 1.1 (no repair delay). Step 2: Construction of spatiotemporal correlation matrix:

[0228] Key time nodes and core association pairs: 24 hours before surgery, after anesthesia induction, 1 hour after surgery, at the end of surgery, and 24 hours after surgery. The core association pairs include 9 groups such as SAI score-IL-6 inflammation grade-blood pressure fluctuation.

[0229] Correlation coefficient calculation:

[0230] The Spearman coefficient was calculated as 0.62 (P less than 0.01) for the ratio of inflammation grade (grade 2) minus blood pressure fluctuation amplitude, which was a continuous variable × a rank variable, at 1 hour after surgery.

[0231] The preoperative 24h SAI score (42)-IL-6 (12.3 pg / ml) was a continuous variable × continuous variable, and the normality test was met using the Pearson coefficient, which was 0.58 (P less than 0.05).

[0232] Clinical verification: All correlation coefficients are consistent with medical logic such as increased inflammation → increased blood pressure fluctuations, and there are no conflicting results.

[0233] Step 3: Complication prediction model output:

[0234] Multimodal fusion: Preoperative genetic data (COMTMet / Met type), intraoperative pathological images (inflammation level 2), and temporal physiological signals (blood pressure fluctuations) are input. The fusion layer attention mechanism assigns a weight of 0.32 to the inflammation level-HMGB1 association pair.

[0235] Prediction results: The risk probability of postoperative agitation is 18%, and the risk probability of nausea and vomiting is 42% (Sigmoid function); the ranking of multiple complications is nausea and vomiting > pain > agitation (Softmax function).

[0236] Step 4-5: Generation of clinical intervention recommendations:

[0237] Core risk factors: Factors with a contribution of ≥10% were screened based on the SHAP value, and the ranking was HMGB1 (6.2 ng / ml, 35%) > SAI score (42 points, 28%).

[0238] Interventions:

[0239] Call HMGB1>6ng / ml from the mapping library → monitor PCT 6h after surgery (pathological indicator intervention).

[0240] Combined with the COMTMet / Met genotype (opioid sensitivity), the postoperative analgesia regimen was adjusted: the morphine dose was reduced by 20%.

[0241] The results of the comparison of the technical effects of this technical solution and the traditional complication prediction solution are shown in Table 1:

[0242] Table 1

[0243] index This technical solution Traditional method (Logistic regression) Traditional method (Apfel score) AUC (prediction accuracy) 0.92 0.75 0.68 Early warning sensitivity 85% 60% 55% Incidence of severe postoperative complications 4.2% 11.5% 13.8% Clinical intervention decision time 2.1 minutes 8.5 minutes 10.3 minutes Adverse drug reaction rate 3.5% (Gene Adjustment) 9.8% 10.2%

[0244] This technical solution, through multi-dimensional data fusion, dynamic correlation analysis and individualized intervention, is significantly superior to traditional methods in terms of predictive accuracy, clinical efficiency and safety.

[0245] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A modeling method for predicting complications after gynecological general anesthesia based on pathological data analysis, characterized in that: The following steps are involved: Step 1: Collect multidimensional pathological and clinical data before, during, and after gynecological general anesthesia; Step 2: Establish a spatiotemporal correlation matrix among anxiety, intraoperative pathology, and complications; Step 3: Based on the spatiotemporal correlation matrix, a gynecological general anesthesia postoperative complication prediction model based on a multimodal deep learning prediction architecture was constructed; In step 4, the complication prediction model inputs the multi-dimensional pathological and clinical data of the current patient before, during, and after surgery, and outputs the type of complication and the probability of its occurrence.

2. The method for predicting complications after gynecological general anesthesia based on pathological data analysis according to claim 1, characterized in that: The multi-dimensional pathological and clinical data before gynecological general anesthesia in step 1 include basic clinical index data, pathological molecular marker detection data, and psychological status quantitative data; Multi-dimensional pathological and clinical data during gynecological general anesthesia include real-time physiological index data, intraoperative pathological index data, and surgical operation parameters; Multidimensional pathological and clinical data after gynecological general anesthesia surgery include complication records and repair and infection marker detection data.

3. The method for predicting complications after gynecological general anesthesia based on pathological data analysis according to claim 2, characterized in that: In step 2, the time is divided into several key time nodes; The spatial dimension is defined as several types of associated subjects, and several groups of core associated pairs are formed based on the combination of representative indicators of several types of associated subjects.

4. The method for predicting complications after gynecological general anesthesia based on pathological data analysis according to claim 3, characterized in that: In step 2, the spatiotemporal correlation matrix uses several key time nodes as row labels and several groups of core correlation pairs as column labels. The initial values ​​of the matrix are all 0. The correlation coefficients obtained through statistical calculations are then used to fill the matrix cells to form a complete spatiotemporal correlation matrix.

5. The method for predicting complications after gynecological general anesthesia based on pathological data analysis according to claim 4, characterized in that: Several key time points are: 24 hours before surgery, after anesthesia induction, 1 hour after surgery, at the end of surgery, and 24 hours after surgery; Several categories of related subjects are: three categories of related subjects, namely psychological state, pathological indicators and physiological reactions; Representative indicators of psychological status are anxiety assessment, heart rate variability, and galvanic skin response; Representative pathological indicators include preoperative inflammatory factors, inflammation grade, and high-mobility group protein B1 detection value; Representative indicators of physiological responses are blood pressure fluctuation amplitude, heart rate variability and BIS value.

6. The method for predicting complications after gynecological general anesthesia based on pathological data analysis according to claim 5, characterized in that: Specific statistical calculation process of the correlation coefficient: Step a: extracting the associated pair data of the corresponding time period according to the time node from the standardized time series database; Step b: If the correlation pair is a continuous variable and a continuous variable, a normality test is performed first. If it conforms to the normal distribution, the Pearson correlation coefficient is used; if it is a continuous variable and a rank variable, the Spearman rank correlation coefficient is used, and the rank correlation is calculated by sorting the data; Step c: perform a t-test on the calculated correlation coefficient and calculate the P value. Calculate the P value for each of the five key time nodes according to steps a to c. Step d: Perform clinical logic verification on the statistical results. If any results conflict with medical common sense, they will be revised based on the opinions of clinical experts.

7. The method for predicting complications after gynecological general anesthesia based on pathological data analysis according to claim 6, characterized in that: In step 3, the model architecture of the complication prediction model includes input layer, fusion layer and output layer; The input layer includes static data channel, time series data channel and image data channel; The fusion layer introduces an attention mechanism based on the spatiotemporal correlation matrix: dynamic attention weights are assigned to the indicators involved in the core correlation pairs; The output layer is designed based on multi-task learning to simultaneously predict multiple types of complications and risk probabilities; The Sigmoid function was used to predict the probability of a single complication, and the Softmax function was used to rank multiple complications.

8. The method for predicting complications after gynecological general anesthesia based on pathological data analysis according to claim 7, characterized in that: The method also includes step 5, generating clinical intervention recommendations based on the complication prediction results.

9. The method for predicting complications after gynecological general anesthesia based on pathological data analysis according to claim 8, characterized in that: The generation of clinical intervention recommendations in step 5 includes the following steps: The complication prediction model outputs the complication risk probability and the contribution of each feature, sets the contribution threshold, screens out the core risk factors, and sorts them in descending order of contribution to form a risk factor priority list; Establish a mapping library of risk factors and intervention measures, including: pathological indicator intervention, psychological state intervention, and physiological indicator intervention; The corresponding intervention measures are called from the mapping library according to the risk factor type.

10. The method for predicting complications after gynecological general anesthesia based on pathological data analysis according to claim 9, characterized in that: Gene polymorphism correction factors were introduced to adjust intervention measures in combination with the patient's basic characteristics.

Images