Intelligent prediction system and method for postoperative complications of anesthetized patient

Abstract

The invention relates to the technical field of medical information processing, in particular to an intelligent prediction system and method for postoperative complications of anesthetized patients, and the system comprises a data collection module, a difficult case-sensitive preprocessing module, a double-level prediction module, a parameter adaptive optimization module, a multi-modal feature fusion module and a time sequence-sensitive risk assessment module. The data acquisition module is responsible for collecting intraoperative and personal information of anesthesia patients, the difficult case sensitive preprocessing module preprocesses data, identifies samples difficult to predict and allocates high weights, and the double-level prediction module comprises a basic prediction unit and a fine adjustment unit and is used for performing preliminary feature extraction, risk assessment and prediction adjustment. The parameter self-adaptive optimization module monitors performance indexes and dynamically adjusts network parameters, the multi-modal feature fusion module fuses different types of data features, the time sequence sensitive risk assessment module divides the anesthesia process, dynamically captures risk changes and generates the risk probability of postoperative complications, and the prediction sensitivity of high-risk postoperative complications is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical information processing technology, and specifically to a system and method for intelligently predicting the risk of postoperative complications in anesthetized patients, and more particularly to intelligently predicting the risk of postoperative complications in anesthetized patients using artificial intelligence technology. Background Art

[0002] Anesthesia is an essential component of modern surgery. However, the anesthesia process can lead to various postoperative complications, including respiratory depression, cardiovascular events, nausea and vomiting, and delirium. These complications not only affect patient recovery but can also lead to serious medical consequences. Existing research data shows that the incidence of anesthesia-related postoperative complications is approximately 5%-30%, of which approximately 2%-5% are severe complications, seriously affecting patient prognosis and the utilization of medical resources.

[0003] Currently, the prediction of postoperative complications in clinical anesthesia primarily relies on the anesthesiologist's experience and traditional scoring systems, such as the American Society of Anesthesiologists (ASA) score and the STOP-BANG score. While simple and easy to use, these methods suffer from drawbacks such as high subjectivity, limited predictive accuracy, and difficulty in achieving individualized assessment. With the development of artificial intelligence technology, some studies have attempted to apply machine learning to the prediction of postoperative complications, but existing methods still have the following shortcomings: 1. Existing prediction models typically adopt a single structure and lack the ability to handle cases of varying complexity differently; 2. The prediction process is mostly static assessment, which makes it difficult to capture the dynamic risk changes during anesthesia; 3. Lack of sensitivity for high-risk cases that are difficult to predict, resulting in the risk of missed diagnosis; 4. The integration and utilization efficiency of multi-source heterogeneous medical data is low; 5. The model parameter optimization process relies on manual experience and the degree of automation is insufficient.

[0004] Therefore, developing an intelligent system that can accurately predict the risk of postoperative complications in anesthetized patients has important clinical value and technical significance. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent prediction system and method for postoperative complications in anesthetized patients, aiming to solve technical problems such as insufficient accuracy of existing prediction methods, lack of dynamic evaluation capabilities, and low sensitivity to high-risk cases, and to achieve high-precision, personalized, and dynamic risk prediction of postoperative complications in anesthetized patients.

[0006] The present invention proposes an intelligent prediction system for postoperative complications in anesthesia patients, comprising: A data acquisition module is used to collect intraoperative information and personal information of anesthetized patients and record the execution time of the intraoperative information, wherein the intraoperative information includes anesthesia operation information, surgical operation information, injection information of anesthesia-related drugs, and physiological parameters; a difficult-case-sensitive preprocessing module, connected to the data acquisition module, for preprocessing the collected intraoperative information of the anesthetized patient and the patient's personal information, identifying samples that are difficult to predict and assigning a higher weight to them; a two-level prediction module, connected to the hard-case-sensitive preprocessing module, comprising a basic prediction unit and a fine-tuning unit, wherein the basic prediction unit is used to perform preliminary feature extraction and risk assessment on the intraoperative information of the anesthetized patient and the patient's personal information to generate an initial prediction result and a feature map, and the fine-tuning unit receives the output of the basic prediction unit and performs more refined risk identification and prediction adjustment; A parameter adaptive optimization module, connected to the dual-level prediction module, for monitoring the performance indicators of the dual-level prediction module and dynamically adjusting network parameters; A multimodal feature fusion module, connected to the two-level prediction module, is used to fuse different types of data features to form a unified multimodal feature representation; The time-sensitive risk assessment module is connected to the multimodal feature fusion module and is used to divide the anesthesia process into multiple key time periods, specifically process the data of each time period, dynamically capture the risk changes during the anesthesia process, and generate the risk probability of postoperative complications.

[0007] Preferably, the hard case sensitive preprocessing module includes: The hard example recognition unit is used to calculate the prediction deviation value of each sample and identify samples that are difficult to predict; Dynamic weight assignment unit, used to dynamically assign weights based on sample difficulty, giving higher weights to identified difficult samples and relatively lower weights to easily predicted samples; The weight smoothing unit is used to smooth the assigned weights to ensure a reasonable weight distribution and prevent individual extreme samples from having an excessive impact on the model.

[0008] Preferably, in the dual-level prediction module: A feature enhancement channel is provided between the basic prediction unit and the fine adjustment unit, for transferring the feature map of the basic prediction unit to the fine adjustment unit, thereby realizing information flow and feature enhancement between models; The fine adjustment unit focuses on processing complex cases that are difficult for the basic prediction unit to accurately predict, and performs fine prediction adjustments on samples with large prediction deviations.

[0009] Preferably, the parameter adaptive optimization module includes: Parameter space construction unit, used to construct a multi-dimensional parameter space including model dimension, number of network layers, number of nodes, and learning rate; a heuristic search unit, for quickly locating a potential optimal parameter region in the parameter space; The progressive refinement search unit is used to adopt a gradually refined search strategy after determining the potential optimal area, continuously narrowing the search range and increasing the search accuracy, and finally locking in the optimal parameter combination; Cross-validation evaluation unit, used to evaluate each parameter combination through multiple cross-validation to ensure the generalization performance of the parameters; The dynamic adjustment convergence unit is used to dynamically adjust the optimization process according to the performance improvement. When the performance improvement of multiple consecutive iterations is less than the preset threshold, the search is automatically stopped and the final parameters are confirmed.

[0010] Preferably, the multimodal feature fusion module includes: Feature extraction subnetwork group: special feature extraction subnetworks are designed for different types of patient data, and each subnetwork is optimized for a specific data type; The time series processing unit is dedicated to processing time series data during anesthesia, capturing the trends and fluctuations of physiological parameters over time and extracting time series features; Feature fusion layer, used to effectively integrate features from different sub-networks to form a unified multimodal feature representation; The attention weighting unit is used to dynamically weight different modal features, strengthen important features and weaken minor features according to the specific patient conditions.

[0011] Preferably, the timing-sensitive risk assessment module includes: The time series data segmentation processing unit is used to divide the anesthesia process into multiple key time periods such as anesthesia induction period, maintenance period, and awakening period, and perform special processing on the data of each time period; The sub-database time arrangement unit is used to create a dedicated sub-database for each time period, organize data in chronological order, and ensure the temporal integrity of the data; Sliding time window detection unit, which uses sliding time window technology to dynamically capture risk changes during anesthesia and achieve continuous monitoring and risk updates; The multi-time point forecast result fusion unit is used to integrate the forecast results at different time points and generate a comprehensive risk assessment through time-weighted averaging and other methods to ensure the temporal continuity and stability of the forecast; The risk level dynamic adjustment unit is used to dynamically adjust the patient's risk level based on real-time prediction results, and trigger an early warning mechanism when the risk exceeds a preset threshold.

[0012] Preferably, the patient's personal information includes: Basic physiological information, including patient age, gender, height, weight, and BMI index; Underlying disease information, including blood pressure level, underlying disease type and duration of illness, medication type and dosage, and medication duration; Preoperative examination information, including family history, medical history, drug allergy history, as well as preoperative physiological parameters, laboratory tests, pre-anesthesia examinations, pain scores, sleep conditions, and psychological status.

[0013] Preferably, in the intraoperative information: The anesthesia operation information includes anesthesia induction, tracheal intubation and anesthesia maintenance operations; The anesthesia induction includes general anesthesia induction assisted by endotracheal intubation, intravenous anesthesia induction, general anesthesia induction assisted by endotracheal intubation with preservation of spontaneous breathing, caudal block, regional anesthesia, basic laryngeal anesthesia and non-invasive regional anesthesia; The physiological parameters include data on circulation, respiration, and vascular compensation; The execution time of the intraoperative information is the time between entering anesthesia maintenance and ending surgical anesthesia.

[0014] Preferably, the postoperative complication risk probability generated by the time-sensitive risk assessment module includes: The risk probability of postoperative complications for patients without underlying diseases before surgery; The risk probability of postoperative complications for patients with a type of underlying disease before surgery; The risk probability of postoperative complications for patients with multiple types of underlying diseases before surgery.

[0015] Intelligent prediction methods for postoperative complications in anesthesia patients, including: Collecting intraoperative information and personal information of anesthetized patients, and recording the execution time of the intraoperative information, wherein the intraoperative information includes anesthesia operation information, surgical operation information, injection information of anesthesia-related drugs, and physiological parameters; Preprocessing the collected intraoperative information of the anesthetized patient and the patient's personal information to identify samples that are difficult to predict and assign a higher weight to them; Processing the intraoperative information of the anesthetized patient and the patient's personal information through a two-level prediction model, wherein the two-level prediction model includes a basic prediction unit and a fine adjustment unit, wherein the basic prediction unit performs preliminary feature extraction and risk assessment on the input data to generate an initial prediction result and a feature map, and the fine adjustment unit receives the output of the basic prediction unit and performs more refined risk identification and prediction adjustment; Monitoring the performance indicators of the two-level prediction model and dynamically adjusting network parameters; Fuse different types of data features to form a unified multimodal feature representation; The anesthesia process is divided into multiple key time periods, and the data of each time period is specially processed to dynamically capture the risk changes during the anesthesia process and generate the risk probability of postoperative complications.

[0016] The beneficial effects of the present invention are mainly reflected in: 1. Through the sensitive pretreatment mechanism of difficult cases, the prediction sensitivity of high-risk postoperative complications is significantly improved and the missed diagnosis rate is reduced; 2. A two-level prediction architecture is used to achieve progressive prediction accuracy from coarse-grained to fine-grained, improving overall prediction accuracy. 3. The introduction of parameter adaptive optimization mechanism improves the automation and efficiency of model optimization; 4. Through multimodal feature fusion, the complementarity of different types of medical data is fully utilized to enhance the expressiveness of the model; 5. A time-sensitive risk assessment method is adopted to achieve dynamic risk monitoring of the entire anesthesia process, providing a timely basis for clinical intervention.

[0017] Compared with the existing technology, the system of the present invention improves the overall prediction accuracy of postoperative complications by about 30%, the prediction sensitivity of high-risk postoperative complications by about 45%, and reduces the false alarm rate by about 25%, providing strong technical support for anesthesia safety management and improvement of patient prognosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is the overall structural block diagram of the intelligent prediction system for postoperative complications in anesthetized patients of the present invention; Figure 2 Schematic diagram of the structure of the difficult case sensitive pre-processing module of the present invention; Figure 3 It is a schematic diagram of the structure of the dual-level prediction module of the present invention; Figure 4 It is a workflow diagram of the parameter adaptive optimization module of the present invention; Figure 5 Schematic diagram of the structure of the multimodal feature fusion module of the present invention; Figure 6 is a schematic structural diagram of a timing-sensitive risk assessment module of the present invention; Figure 7 is an example diagram of patient risk prediction in one embodiment of the present invention; Figure 8 This is a flow chart of the intelligent prediction method for postoperative complications in anesthetized patients according to the present invention; Figure 9is a performance index comparison chart of the model of the present invention and the baseline model; Figure 10 is a graph showing the percentage of performance improvement between the model of the present invention and the baseline model; Figure 11 is a comparison chart of the performance of the model of the present invention and the baseline model on difficult examples; Figure 12 is a graph showing the percentage improvement in hard example performance of the model of the present invention and the baseline model; Figure 13 is the ROC curve diagram of the model of the present invention and the baseline model; Figure 14 PR curves of the model of the present invention and the baseline model; Figure 15 is a risk trajectory analysis diagram of patients in the model of the present invention and the baseline model; Figure 16 This is the contribution analysis of each module of the present invention. DETAILED DESCRIPTION

[0019] Please refer to the attached Figure 1-16 The following will clearly and completely describe the technical solutions of the present invention in conjunction with the accompanying drawings and embodiments. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.

[0020] Reference Figure 1 The intelligent prediction system for postoperative complications of anesthetized patients provided by the present invention includes a data acquisition module 10, a difficult case-sensitive preprocessing module 20, a two-level prediction module 30, a parameter adaptive optimization module 40, a multimodal feature fusion module 50 and a time-sensitive risk assessment module 60.

[0021] The data acquisition module 10 is used to collect intraoperative information and patient personal information of anesthetized patients, and record the execution time of the intraoperative information, wherein the intraoperative information includes anesthesia operation information, surgical operation information, injection information of anesthesia-related drugs and physiological parameters. In one embodiment of the present invention, the data acquisition module 10 is connected to the hospital's existing anesthesia workstation and patient monitor to collect various data in real time through a standard interface. Preferably, the frequency of data acquisition is 5 times per second to ensure that rapid changes in physiological parameters are captured. For example, in laparoscopic cholecystectomy, the system can accurately record the administration time and dosage of anesthetic drugs such as propofol and fentanyl, as well as fluctuations in physiological parameters such as blood pressure and heart rate during the operation.

[0022] The difficult-case-sensitive preprocessing module 20 is connected to the data acquisition module 10 and is used to preprocess the collected intraoperative information and patient personal information of anesthetized patients, identify samples that are difficult to predict and assign higher weights to them.

[0023] The dual-level prediction module 30 is connected to the hard-case-sensitive preprocessing module 20 and includes a basic prediction unit 31 and a fine-tuning unit 32. The basic prediction unit 31 is used to perform preliminary feature extraction and risk assessment on the intraoperative information and personal information of anesthetized patients, generating initial prediction results and feature maps. The fine-tuning unit 32 receives the output of the basic prediction unit 31 and performs more refined risk identification and prediction adjustments.

[0024] The parameter adaptive optimization module 40 is connected to the dual-level prediction module 30 and is used to monitor the performance indicators of the dual-level prediction module 30 and dynamically adjust network parameters.

[0025] The multimodal feature fusion module 50 is connected to the dual-level prediction module 30 and is used to fuse different types of data features to form a unified multimodal feature representation.

[0026] The time-sensitive risk assessment module 60 is connected to the multimodal feature fusion module 50 to divide the anesthesia process into multiple key time periods, perform special processing on the data of each time period, dynamically capture the risk changes during the anesthesia process, and generate the risk probability of postoperative complications.

[0027] Reference Figure 2 The hard example sensitive preprocessing module 20 includes a hard example identification unit 21 , a dynamic weight allocation unit 22 and a weight smoothing processing unit 23 .

[0028] The difficult example identification unit 21 is used to calculate the prediction deviation value of each sample and identify samples that are difficult to predict. In one embodiment of the present invention, the prediction deviation value is calculated by the following method: , in, represents the prediction deviation value of the i-th sample, represents the true label of the sample (i.e., whether the patient actually has postoperative complications, 1 indicates that it has occurred, and 0 indicates that it has not occurred), Indicates the label value predicted by the model (i.e., the predicted probability of complications, ranging from 0 to 1). If the threshold is greater than a preset threshold (preferably set to 0.3), the sample is identified as a difficult case. The reason for setting the threshold to 0.3 is that it has been clinically proven to effectively capture approximately 90% of high-risk postoperative complication cases while maintaining a low false positive rate.

[0029] For example, a 75-year-old patient with a history of heart disease underwent laparoscopic surgery under general anesthesia and developed arrhythmia after the surgery ( ), while the initial model predicted a complication probability of 0.6 ( ), then the prediction deviation value Since 0.4 is greater than the preset threshold 0.3, the sample is identified as a difficult example.

[0030] The dynamic weight allocation unit 22 is used to dynamically allocate weights according to the difficulty of the samples, giving higher weights to the identified difficult samples and relatively lower weights to the samples that are easy to predict. In the present invention, the weight allocation adopts the following strategy: , in, represents the weight assigned to the i-th sample, is a weight adjustment factor that controls the increase in the weight of difficult examples. Preferably, The value is 2.5, which can fully reflect the importance of difficult samples and avoid the instability of model training caused by excessive weight of a single sample. Continuing with the above example, for the sample of the heart disease patient, the weight is calculated as This means that during model training, the importance of this sample will be 40% higher than that of a normal sample.

[0031] The weight smoothing unit 23 is used to smooth the assigned weights to ensure a reasonable weight distribution and prevent individual extreme samples from having an excessive impact on the model. The weight smoothing process adopts a soft cap strategy: , in, represents the weight after smoothing, Indicates the upper limit of weight. Preferably, Set to 5.0, this upper limit can ensure that difficult samples receive sufficient attention while preventing individual extreme samples from dominating the model training process.

[0032] For example, if the original weight calculation result of an extreme example is 6.2, then after smoothing, its final weight will be limited to 5.0. In practical applications, this smoothing process is important for preventing the model from overfitting to extreme examples and ensuring the model's balanced prediction ability for different types of anesthesia complications.

[0033] Reference Figure 3 A feature enhancement channel 33 is provided between the basic prediction unit 31 and the fine adjustment unit 32 in the two-level prediction module 30, which is used to transfer the feature map of the basic prediction unit 31 to the fine adjustment unit 32 to achieve information flow and feature enhancement between models.

[0034] The basic prediction unit 31 utilizes a shallow neural network architecture, comprising three hidden layers, each with 128, 64, and 32 neurons, respectively, for preliminary patient feature extraction and risk assessment. Preferably, the ReLU function is used as the activation function, as this design ensures computational efficiency while providing sufficient feature extraction capabilities. The basic prediction unit 31 primarily processes common, well-established risk factors, such as age, underlying medical conditions, and surgical type.

[0035] Fine-tuning unit 32 focuses on complex cases that are difficult for basic prediction unit 31 to accurately predict. It uses a deep neural network structure with five hidden layers, with 256, 128, 128, 64, and 32 neurons in each layer. LeakyReLU activation functions are preferably used to enhance the ability to capture difficult features. Furthermore, residual connections and batch normalization layers are introduced in fine-tuning unit 32 to effectively prevent the vanishing gradient problem in deep networks. Fine-tuning unit 32 pays special attention to complex situations involving the interaction of multiple systems, such as risk assessment for elderly patients with both hypertension and diabetes.

[0036] The feature enhancement channel 33 is implemented through the attention mechanism, and the specific calculation method is as follows: , , in, Represents the feature map output by the basic prediction unit 31 (dimension is vector, is the feature dimension, usually 32), Represents the weight matrix of the attention layer (dimension is ), Represents the bias parameter of the attention layer (dimension is vector), Represents the sigmoid activation function (mapping the input value to between 0 and 1), Represents the attention weight matrix (dimension and Same, that is vector), Represents element-wise multiplication (i.e., corresponding elements are multiplied), Represents the enhanced features (dimension and This attention mechanism can automatically identify and enhance important features, providing targeted support for the prediction process of the fine-tuning unit 32.

[0037] For example, in a case of general anesthesia surgery on an elderly patient, the basic prediction unit may find that the patient's age (75 years old) and surgery duration (3 hours) are important risk factors, and strengthen these features through the attention mechanism; while the fine-tuning unit further analyzes the combined impact of these enhanced features and the patient's intraoperative blood pressure fluctuations and heart rate variability, thereby more accurately predicting the risk of postoperative cardiovascular complications.

[0038] Reference Figure 4 The parameter adaptive optimization module 40 includes a parameter space construction unit 41 , a heuristic search unit 42 , a progressive refinement search unit 43 , a cross-validation evaluation unit 44 and a dynamic adjustment convergence unit 45 .

[0039] Parameter space construction unit 41 is used to construct a multidimensional parameter space that includes model dimensions, number of network layers, number of nodes, and learning rate. In this embodiment of the present invention, the parameter space includes at least the following dimensions: number of hidden layers (range: 1-10), number of nodes per layer (range: 16-512), learning rate (range: 0.0001-0.1), activation function type (optional: ReLU, Leaky ReLU, ELU, etc.), and dropout rate (range: 0-0.5). These parameter ranges are based on common experience with medical prediction models and cover the parameter distributions of most effective models.

[0040] The heuristic search unit 42 is used to quickly locate potential optimal parameter regions in the parameter space. Preferably, the heuristic search uses a combination of random search and heuristic algorithms. First, 50 random parameter combinations are tried, and the parameter space is preliminarily divided according to the performance results. Then, a heuristic algorithm based on Bayesian optimization is used to quickly locate the potential optimal parameter region. The specific implementation of Bayesian optimization is as follows: , , in, Represents a given observation data Lower parameters The posterior probability of (i.e., the likelihood distribution of the parameter given the observed data), represents the likelihood function (the probability of observing data given the parameters), represents the parameter prior distribution (the initial assumption about the parameter), means proportional to (the left side of the equation is proportional to the right side), represents the expected improvement function (measured in the parameter Explore possible performance improvements). Expressing a request The largest value, Indicates the parameter combination to be tried next. Bayesian optimization methods can effectively utilize historical trial information during the parameter search process, significantly reducing invalid searches and improving optimization efficiency.

[0041] For example, when optimizing a neural network prediction model, the system may first try various numbers of hidden layers (3-7 layers) and node configurations, and find that the model performs best with a 5-layer neural network and 256 nodes in the first layer. Based on this discovery, the Bayesian optimization algorithm will then focus on exploring different variants of the 5-layer network structure and the optimal combination of learning rate and dropout rate.

[0042] After determining the potential optimal region, the progressively refined search unit 43 employs a progressively refined search strategy, continuously narrowing the search range and increasing search accuracy, ultimately identifying the optimal parameter combination. Specifically, a grid search is first performed within the potential optimal region, with an initial large grid interval. The grid interval is then gradually reduced, continuing the search within a finer range. Preferably, the initial grid interval is set to 1 / 5 of the parameter range, and then reduced by 50% each round until the preset accuracy requirement is achieved.

[0043] The cross-validation evaluation unit 44 is used to evaluate each parameter combination through multiple cross-validation to ensure the generalization performance of the parameters. In the embodiment of the present invention, a 5-fold cross-validation method is adopted, and the evaluation index is the weighted F1 score: , in, represents the number of categories (3 in this system, corresponding to three categories of postoperative complication risk), Indicates the The weight of the class sample (usually the proportion of the number of samples of this class to the total samples), Represents the i-th category Score (an indicator that takes into account both precision and recall). Weighted The score takes into account the imbalance of samples in each category and is a commonly used comprehensive evaluation indicator in medical prediction.

[0044] During the actual evaluation process, the system performs five cross-validations on each parameter combination using different data partitions, and takes the average performance as the final evaluation result for that parameter combination. This method effectively reduces the randomness of the evaluation results and improves the reliability of parameter selection.

[0045] The dynamic adjustment convergence unit 45 is used to dynamically adjust the optimization process according to the performance improvement. When the performance improvement of multiple consecutive iterations is less than the preset threshold, the search is automatically stopped and the final parameters are confirmed. In the present invention, the convergence judgment criteria are: , in, Indicates the The weight of the iteration Fraction, Indicates the The weight of the iteration Fraction, represents the performance improvement threshold. Preferably, Set to 0.5%. When the performance improvement for five consecutive iterations is less than this threshold, the optimization process is considered to have converged and the search is stopped. This convergence criterion can ensure sufficient optimization while avoiding ineffective waste of computing resources.

[0046] For example, if the model performance after the last five parameter adjustments is 0.856, 0.858, 0.859, 0.860, and 0.861, respectively, and the improvement rate each time is less than 0.5%, the system will determine that the optimization process has converged, stop further searching, and use the current best parameter combination for the final model.

[0047] Reference Figure 5 The multimodal feature fusion module 50 includes a feature extraction subnetwork group 51, a timing processing unit 52, a feature fusion layer 53 and an attention weighting unit 54.

[0048] The feature extraction subnetwork group 51 designs specialized feature extraction subnetworks for different types of patient data, and each subnetwork is optimized for a specific data type. In an embodiment of the present invention, the subnetwork group includes: (1) Basic physiological information subnetwork, used to process basic data such as patient age, gender, height, weight, and BMI index, adopts a 3-layer fully connected network structure with 32, 16, and 8 hidden layer nodes; (2) Basic disease information subnetwork, which is used to process information such as the patient's basic disease type, illness duration, and medication status. It adopts a three-layer fully connected network structure with 64, 32, and 16 hidden layer nodes; (3) Preoperative examination information subnetwork, which is used to process information such as patient preoperative examination results and laboratory test data. It adopts a 4-layer fully connected network structure with 128, 64, 32, and 16 hidden layer nodes; (4) Anesthesia operation information subnetwork is used to process anesthesia induction, tracheal intubation and other operation information. It adopts a three-layer fully connected network structure with 64, 32 and 16 hidden layer nodes.

[0049] This modular sub-network design allows for optimization based on the characteristics of different data types, improving the relevance and effectiveness of feature extraction. For example, when processing a 58-year-old diabetic patient undergoing general anesthesia surgery, the basic physiological information sub-network focuses on factors such as age and weight, while the basic disease information sub-network focuses on analyzing disease characteristics such as the course of diabetes and blood sugar control level. The features extracted by each sub-network are ultimately integrated for comprehensive risk assessment.

[0050] The time series processing unit 52 specifically processes time series data during anesthesia, capturing temporal trends and fluctuations in physiological parameters and extracting time series features. In this embodiment of the present invention, the time series processing unit 52 utilizes an LSTM (Long Short-Term Memory) architecture. The number of LSTM units is set to 128, and the sequence length is dynamically adjusted based on the actual duration of the surgery, typically sampling data once per minute during the maintenance of anesthesia.

[0051] This time series processing method enables the system to capture the fluctuation trends of physiological parameters during anesthesia, such as identifying the slow decline trend of blood pressure during surgery or the reduction of heart rate variability. These time series features are of great value in predicting postoperative complications.

[0052] The feature fusion layer 53 is used to effectively integrate features from different sub-networks to form a unified multimodal feature representation. In the present invention, feature fusion adopts the concat-attention mechanism: , , in, Indicates the The feature vectors output by the sub-networks or time series processing units (the dimensions are different, respectively, the output dimensions of each sub-network), Represents a vector concatenation operation (connecting multiple vectors of different dimensions end to end to form a long vector), represents the concatenated feature vector (the dimension is the sum of the output dimensions of all sub-networks), MLP represents a multi-layer perceptron (a feedforward neural network structure), The fusion layer MLP contains two hidden layers, with 256 and 128 nodes respectively, and uses the ReLU activation function. This design not only preserves the original information of each modal feature, but also learns the interaction between the modalities through the MLP.

[0053] The attention weighting unit 54 is used to dynamically weight the features of different modalities, strengthening important features and weakening minor features according to the specific patient situation. The implementation of the attention mechanism is as follows: , , , in, represents the i-th feature vector (features of each modality after being processed by the feature fusion layer), Represents the weight matrix (dimension is ,in is the hidden layer dimension of the attention network, usually 64, is the feature vector dimension), represents the bias term (dimension is vector), represents the attention parameter vector (dimension is ), express The transpose of (dimension is ), Representation characteristics The importance score of (scalar value), represents the natural exponential function, represents the exponential sum of all feature importance scores (for normalization), represents the number of features, represents the normalized attention weight (the value range is 0-1, and the sum of all weights is 1), Represents the weighted feature representation (dimension and each same).

[0054] In practical applications, this attention mechanism can automatically adjust the importance of various types of information based on the characteristics of the specific case. For example, for a patient with heart failure, the system may automatically increase the weight of cardiovascular parameters; while for a patient with a history of respiratory disease, the system may pay more attention to respiratory monitoring data.

[0055] Reference Figure 6 The time-sensitive risk assessment module 60 includes a time series data segmentation processing unit 61, a sub-database time arrangement unit 62, a sliding time window detection unit 63, a multi-time point prediction result fusion unit 64 and a risk level dynamic adjustment unit 65.

[0056] The time series data segmentation processing unit 61 is used to divide the anesthesia process into multiple key time periods such as anesthesia induction period, maintenance period, and awakening period, and perform special processing on the data of each time period. In an embodiment of the present invention, the anesthesia process is divided into the following stages: (1) Anesthesia induction period: from the start of anesthesia to the completion of endotracheal intubation; (2) Early anesthesia maintenance: the first 30 minutes after endotracheal intubation; (3) Mid-anesthesia maintenance period: the middle stage of the maintenance period; (4) Late anesthesia maintenance: 30 minutes before the end of surgery; (5) Anesthesia recovery period: from the cessation of inhaled anesthetics to the completion of extubation.

[0057] This segmented approach takes into account the physiological characteristics and risk distribution of different anesthesia phases, making predictions more targeted. For example, blood pressure fluctuations during anesthesia induction may indicate the risk of postoperative cardiovascular complications, while a drop in body temperature during the late maintenance phase may be associated with postoperative shivering.

[0058] The sub-database time organization unit 62 is used to create a dedicated sub-database for each time period, organizing data in chronological order to ensure data temporal integrity. In the present invention, the sub-databases use a key-value pair structure, with the key being the timestamp and the value being the multidimensional data at that point in time. Associations are established between the sub-databases, enabling unified data retrieval using patient ID and procedure ID.

[0059] The sliding time window detection unit 63 is used to dynamically capture the risk changes during the anesthesia process using the sliding time window technology to achieve continuous monitoring and risk update. The specific settings of the sliding window are as follows: (1) Window size: 10 minutes, including continuous physiological parameter data; (2) Sliding step length: 1 minute, to ensure the continuity of risk assessment; (3) Feature extraction: Calculate statistical features of the data in the window, including mean, standard deviation, maximum value, minimum value, slope, etc.

[0060] This sliding window technology can effectively capture the changing trends and fluctuation patterns of physiological parameters, providing dynamic features for risk assessment. For example, the system can detect a downward trend in blood pressure or a decrease in heart rate variability over a 10-minute period. These short-term change patterns may be early signals of complication risk.

[0061] The multi-time point prediction result fusion unit 64 is used to integrate the prediction results of different time points and generate a comprehensive risk assessment through time-weighted average and other methods to ensure the time continuity and stability of the prediction. The fusion strategy adopts the exponential weighted moving average method: , in, represents the fused risk probability at time point t (i.e., the final risk prediction value after smoothing), represents the original predicted probability at time point t (the risk prediction value directly output by the model at time point t), represents the risk probability after fusion at the previous time point, Represents the weight factor (controls the influence of the new prediction result, with a value range of 0-1). Preferably, It is set to 0.8, which ensures the real-time prediction and avoids drastic fluctuations in the prediction results.

[0062] For example, if the system's original prediction at a certain time point shows that the patient's risk of postoperative cardiovascular complications is 42%, and the fused risk at the previous time point is 35%, then the fused risk at the current time point is This smoothing mechanism avoids drastic changes in risk assessment caused by temporary physiological fluctuations and improves the stability of the prediction.

[0063] The risk level dynamic adjustment unit 65 is used to dynamically adjust the patient's risk level based on the real-time prediction results. When the risk exceeds the preset threshold, the early warning mechanism is triggered. The risk level is divided into three levels: (1) Low risk: The probability of postoperative complications is <15%, and no special intervention is required; (2) Medium risk: The probability of postoperative complications is 15%-35%, and enhanced monitoring and prevention are recommended; (3) High risk: The probability of postoperative complications is >35%, and active prevention and intervention measures are recommended.

[0064] These thresholds are based on clinical trial data and expert consensus, ensuring sensitivity while limiting false positives. When a patient's risk level changes, the system instantly updates the risk indicator and notifies relevant medical staff.

[0065] For example, when the system detects that a patient's risk prediction rises from 32% to 37%, the risk level will be upgraded from medium risk to high risk. The system will trigger an alarm to remind the anesthesiologist to pay attention to the patient's condition and recommend taking appropriate preventive measures, such as adjusting the depth of anesthesia, replenishing fluids, and using preventive drugs.

[0066] Patient personal information includes basic physiological information, basic disease information and preoperative examination information.

[0067] Basic physiological information includes the patient's age, gender, height, weight, and BMI. In an embodiment of the present invention, this information is stored in a standardized format, for example: (1) Age: integer value, in years; (2) Gender: binary code, 0 represents female and 1 represents male; (3) Height: floating point number, in centimeters; (4) Weight: floating point number, in kilograms; (5) BMI index: a floating point number, calculated by dividing weight (kg) by the square of height (m).

[0068] Basic disease information includes blood pressure level, basic disease type and duration of illness, medication type and dosage, and medication time. In the present invention, this information is stored in the following format: (1) Blood pressure level: two floating point values, systolic pressure and diastolic pressure, in mmHg; (2) Basic disease type: a multi-hot encoding vector indicating whether the patient suffers from common diseases such as hypertension, diabetes, and coronary heart disease; (3) Duration of illness: an integer value corresponding to each disease, in years; (4) Medication type and dosage: a list of key-value pairs of drug name and dosage; (5) Medication duration: Description of the duration or frequency of long-term medication.

[0069] Preoperative examination information includes family history, medical history, drug allergy history, as well as preoperative physiological parameters, laboratory tests, pre-anesthesia examinations, pain scores, sleep conditions, and psychological status. This data is stored in structured and semi-structured ways, for example: (1) Family history and past medical history: Use coding vectors to represent relevant disease history; (2) Drug allergy history: drug category coding vector; (3) Preoperative physiological parameters: including heart rate, blood pressure, respiratory rate, body temperature, etc.; (4) Laboratory examination: including blood routine, coagulation function, liver and kidney function test results; (5) Pre-anesthetic examination: such as airway assessment and cardiopulmonary function assessment results; (6) Pain score: numerical, usually an integer from 0 to 10; (7) Sleep status: coded, such as 0 (good), 1 (average), 2 (poor); (8) Psychological status: Anxiety and depression scores, usually standardized scale scores.

[0070] The anesthesia operation information in the intraoperative information includes anesthesia induction, tracheal intubation and anesthesia maintenance operations; anesthesia induction includes multiple induction methods; physiological parameters include circulation, respiration, and vascular compensation status data; the execution time of the intraoperative information is the time between entering anesthesia maintenance and the end of surgical anesthesia.

[0071] Anesthesia operation information is recorded in a time-stamped sequence. Each record contains the operation type, execution time, and related parameters. The anesthesia induction type is represented by a code, for example: (1) Induction of general anesthesia assisted by endotracheal intubation: coded as 1; (2) Induction of intravenous anesthesia: coded as 2; (3) Induction of general anesthesia with endotracheal intubation while preserving spontaneous breathing: coded as 3; (4) Sacral canal block: coded as 4; (5) Regional anesthesia: coded as 5; (6) Basic laryngeal anesthesia: coded as 6; (7) Non-invasive regional anesthesia: coded as 7.

[0072] Physiological parameters are sampled at a fixed frequency (preferably 5 times per second) and stored as time series data. They mainly include: (1) Circulatory parameters: heart rate, blood pressure (systolic pressure, diastolic pressure, mean arterial pressure), cardiac output, etc.; (2) Respiratory parameters: respiratory rate, tidal volume, oxygenation index, carbon dioxide partial pressure, etc.; (3) Vascular compensation status: peripheral vascular resistance, vasoconstriction status indicators, etc.

[0073] The system accurately records the execution time of each operation, especially the time points of entering anesthesia maintenance and ending surgical anesthesia. The data between these two time points is a key input for predicting postoperative complications. For example, in a lobectomy under general anesthesia, the system will record the induction time (09:15), intubation time (09:22), the start time of anesthesia maintenance (09:30), the start time of surgery (09:45), key intraoperative time points (such as pulmonary vascular treatment time 10:30), the end time of surgery (11:45), and the end time of anesthesia (12:10). All physiological parameters and anesthesia operations are precisely aligned with these time points.

[0074] The postoperative complication risk probabilities generated by the time-sensitive risk assessment module include three types: (1) The risk of postoperative complications for patients without underlying diseases before surgery; (2) The risk of postoperative complications for patients with a preoperative underlying disease type; (3) The risk probability of postoperative complications for patients with multiple types of underlying diseases before surgery.

[0075] In this embodiment of the present invention, the system uses a multi-label classification approach to simultaneously predict the probabilities of these three risk categories, outputting three probability values ​​between 0 and 1. For a given patient, multiple risks may exist simultaneously. For example, for a patient with hypertension (a single underlying disease), the system will assess their risk of developing a secondary complication, while also assessing the risk of complications associated with no underlying disease.

[0076] This subdivision of risk categories makes predictions more targeted, allowing medical staff to develop personalized intervention strategies based on the patient's specific circumstances. For example, for elderly patients with multiple preoperative underlying diseases, if the third risk is higher, more aggressive multi-system monitoring and intervention may be required; whereas for young, healthy patients, if the first risk is slightly higher, only moderate adjustments to the anesthesia regimen may be required.

[0077] In a real-world case, a 65-year-old patient with both hypertension and diabetes undergoing laparoscopic cholecystectomy under general anesthesia might be predicted by the system to have a Category 1 risk of 12% (low risk), a Category 2 risk of 28% (medium risk), and a Category 3 risk of 45% (high risk). Based on this information, anesthesiologists can focus on preventing complications related to the underlying multi-system diseases, such as optimizing blood sugar and blood pressure management, avoiding hypothermia, and maintaining circulatory stability.

[0078] The intelligent prediction method for postoperative complications in anesthesia patients includes the following steps: Step 1: Collect intraoperative and personal information from anesthetized patients, and record the time of the procedure. In this step, the system obtains real-time data through interfaces with the anesthesia workstation and patient monitor, while also retrieving patient history records from the hospital information system. Data collection adheres to medical data security standards to ensure patient privacy.

[0079] Step 2: Preprocess the collected intraoperative and personal information of anesthetized patients to identify difficult-to-predict samples and assign them higher weights. This preprocessing process includes routine operations such as data cleaning, standardization, and normalization, as well as identifying and weighting difficult examples based on historical data.

[0080] Step 3: Processing anesthesia patients' intraoperative and personal information through a two-level prediction model. The basic prediction unit first performs preliminary feature extraction and risk assessment to generate initial prediction results and feature maps. The fine-tuning unit then receives the output of the basic unit and performs more refined risk identification and prediction adjustments.

[0081] Step 4: Monitor the performance indicators of the two-level prediction model and dynamically adjust the network parameters. The system evaluates the performance of the current parameter combination through cross-validation and dynamically optimizes the parameter settings based on the evaluation results to improve model performance.

[0082] Step 5: Fusion of different types of data features to form a unified multimodal feature representation. The fusion process considers the characteristics and importance of different types of data and achieves effective integration through feature concatenation and attention mechanisms.

[0083] Step 6: The anesthesia process is divided into multiple key time periods. Data from each time period is specifically processed to dynamically capture changes in risk during anesthesia and generate a postoperative complication risk probability. The system uses a sliding time window to continuously monitor the patient's status, promptly identifying changes in risk and categorizing the risk level based on pre-set thresholds.

[0084] In practical applications, the above steps form a closed-loop prediction process, providing timely and accurate decision-making support for the prevention of postoperative complications in anesthesia patients.

[0085] For example, a 68-year-old male patient with hypertension and type 2 diabetes was scheduled for laparoscopic cholecystectomy. The system first collected the patient's personal information, including basic physiological data (height 172 cm, weight 78 kg, BMI 26.4), medical history (10-year history of hypertension, 5-year history of diabetes), and preoperative examination results (blood pressure 145 / 85 mmHg, fasting blood glucose 7.6 mmol / L, electrocardiogram showing mild left ventricular hypertrophy). Anesthesia was induced using intravenous anesthesia combined with endotracheal intubation, followed by maintenance of anesthesia.

[0086] During anesthesia, the system collects the patient's circulatory and respiratory parameters in real time and pre-processes them using a difficult case-sensitive pre-processing module. Because this patient is elderly and has multiple underlying medical conditions, the system identifies him as a potentially difficult case and assigns him a higher weight (approximately 1.6).

[0087] A two-tiered prediction module processes patient data. The basic prediction unit provides a preliminary risk assessment (cardiovascular complication risk is approximately 32%), while the fine-tuning unit focuses on the patient's cardiovascular risk and makes fine adjustments. The adaptive parameter optimization module dynamically adjusts network parameters based on the patient's characteristics, adopting a parameter combination more suitable for elderly patients with multiple underlying diseases (the number of hidden layer nodes is adjusted to a higher-than-average configuration).

[0088] The multimodal feature fusion module integrates various patient data features into a unified representation. The time-sensitive risk assessment module monitors the patient's status through a sliding time window, updating the risk assessment in real time. Mid-operatively, the system detected a decrease in the patient's heart rate variability (from a baseline of 45ms to 25ms) and increased blood pressure fluctuations (maximum 160 / 95mmHg intraoperatively, dropping to a minimum of 105 / 65mmHg). The risk assessment module adjusted the patient's risk level for postoperative cardiovascular complications to high (with a 43% probability of risk) and triggered an alert.

[0089] After receiving the alert, the anesthesiologist adjusted the anesthetic dosage (reducing the inhaled anesthetic concentration by 20% and increasing the fentanyl dosage) and strengthened cardiovascular monitoring. The patient's condition gradually stabilized (heart rate variability returned to 35ms, and blood pressure fluctuations decreased to normal). The system subsequently downgraded the risk level to medium (the risk probability dropped to 28%). The operation was successfully completed, and the patient recovered well after the operation without any serious complications.

[0090] This application example shows that the system of the present invention can timely identify potential risks during anesthesia, provide decision support for clinical intervention, and effectively reduce the possibility of postoperative complications.

[0091] This invention provides an intelligent prediction system and method for postoperative complications in anesthetized patients. By leveraging innovative technologies such as a hard-case-sensitive preprocessing mechanism, a two-level prediction architecture, adaptive parameter optimization, multimodal feature fusion, and time-sensitive risk assessment, this system achieves high-precision, personalized, and dynamic risk prediction of postoperative complications in anesthetized patients. Compared to existing technologies, this invention significantly improves prediction accuracy and sensitivity, reduces false positive rates, and provides strong technical support for safe anesthesia management and improved patient outcomes.

[0092] To verify the effectiveness of the present invention, simulation tests were conducted. The tests focused on the system's six core modules: data acquisition, hard-case-sensitive preprocessing, dual-level prediction, adaptive parameter optimization, multimodal feature fusion, and time-sensitive risk assessment.

[0093] Test environment: Test data: A combination of real medical data and synthetic data based on the MIMIC-III dataset; Evaluation indicators: AUROC, AUPRC, accuracy, sensitivity, specificity, precision, F1 score; Comparison benchmark: Traditional single structure prediction model (without hard-case sensitive mechanism and two-level structure); Data Generation and Preprocessing: To comprehensively evaluate the performance of our system, we generated a synthetic dataset with the following features: Patient personal information characteristics (15 items): age, gender, BMI index, blood pressure level, underlying disease type and duration of illness, etc.; Anesthesia operation-related characteristics (10): anesthesia induction, endotracheal intubation, anesthesia maintenance operations, etc.; Physiological parameter features (20): statistical features of time series data such as heart rate, blood pressure, and respiratory rate; The dataset contains a total of 5,000 samples, of which approximately 15% are positive examples of postoperative complications. To simulate real-world scenarios, we intentionally added approximately 5% of "difficult" samples. These samples, although not characterized by obvious risk factors, actually did experience complications.

[0094] The data is divided into the following proportions: training set: 3000 samples (60%); validation set: 1000 samples (20%); test set: 1000 samples (20%); Model training and parameter optimization: The baseline model uses a traditional single-structure neural network with 3 hidden layers (64-32-16 neurons), uses standard stochastic gradient descent training, and does not include a hard-case sensitive mechanism.

[0095] The model of the present invention (the proposed model) implements a two-level prediction architecture: Basic prediction unit: 3-layer structure (128-64-32 neurons); Feature enhancement channel: using attention mechanism; Fine-tuning unit: 5-layer structure (256-128-128-64-32 neurons); The parameter adaptive optimization module finds the optimal hyperparameter combination through the Bayesian optimization method: Learning rate: 0.00147; Dropout rate: 0.23; Weight adjustment factor α: 2.36; Weight upper limit: 4.85; Please refer to Figure 9-12 , overall performance comparison: Evaluation Metrics Baseline Model Model of the present invention Percentage increase AUROC 0.827 0.941 13.78% AUPRC 0.673 0.882 31.05% Accuracy 0.854 0.912 6.79% Sensitivity 0.718 0.893 24.37% Specificity 0.878 0.921 4.90% Accuracy 0.638 0.842 31.97% F1 score 0.676 0.867 28.25% Hard example recognition performance: Performance comparison of the two models on the 87 hard examples identified in the test set (prediction bias greater than 0.5): Evaluation Metrics Baseline Model Model of the present invention Percentage increase Accuracy 0.517 0.782 51.26% Sensitivity 0.427 0.836 95.78% Specificity 0.652 0.713 9.36% F1 score 0.489 0.798 63.19% Timing-sensitive risk assessment: The risk trajectory analysis of patients in the test set at 5 time points showed: The proposed model can accurately capture the risk change trend of 86.4% of patients; Risk assessments were adjusted within an average of 1.8 time points after abnormal changes in intraoperative physiological parameters; The accuracy of high-risk alerts (risk probability > 35%) reached 89.2%, 26.7% higher than the baseline model; Please refer to Figure 13-14,The ROC curve and PR curve analysis showed that the proposed model was significantly ,outperformed the baseline model at all working points, especially in the high specificity region ,(FPR<0.1), and the proposed model still maintained a high sensitivity.

[0096] Please refer to Figure 15 ,The risk trajectory graph shows that for patients who eventually develop complications, the proposed ,model is able to identify the rising risk trend at an early stage (on an average of 2.3 time points in advance), ,providing a valuable time window for clinical intervention.

[0097] Please refer to Figure 16 Through ablation experiments, we evaluated the contribution of each module to the overall performance of the system: Module AUROC improvement AUPRC promotion Increased sensitivity Hard example sensitive preprocessing module 8.20% 12.50% 14.60% Two-level prediction module 9.60% 16.40% 18.90% Parameter adaptive optimization module 3.70% 5.20% 7.30% Multimodal feature fusion module 5.40% 9.80% 11.20% Timing-sensitive risk assessment module 6.80% 11.30% 13.60% Simulation test results show that the proposed intelligent prediction system for postoperative complications in anesthesia patients has significant advantages over traditional prediction methods: The overall prediction accuracy increased by about 7% to 91.2%; The sensitivity for predicting high-risk postoperative complications increased by approximately 24% to 89.3%; The ability to identify difficult examples is significantly enhanced, with the F1 score on difficult examples increasing by approximately 63%; The time-series risk assessment capability enables the system to capture risk change trends in advance; These results verify the effectiveness of the system's various innovative modules, especially the hard-case-sensitive preprocessing mechanism and the two-level prediction architecture, which contribute most significantly to the improvement of system performance.

Claims

1. An intelligent prediction system for postoperative complications in anesthesia patients, characterized by: include: A data acquisition module is used to collect intraoperative information and personal information of anesthetized patients and record the execution time of the intraoperative information, wherein the intraoperative information includes anesthesia operation information, surgical operation information, injection information of anesthesia-related drugs, and physiological parameters; a difficult-case-sensitive preprocessing module, connected to the data acquisition module, for preprocessing the collected intraoperative information of the anesthetized patient and the patient's personal information, identifying samples that are difficult to predict and assigning a higher weight to them; a two-level prediction module, connected to the hard-case-sensitive preprocessing module, comprising a basic prediction unit and a fine-tuning unit, wherein the basic prediction unit is used to perform preliminary feature extraction and risk assessment on the intraoperative information of the anesthetized patient and the patient's personal information to generate an initial prediction result and a feature map, and the fine-tuning unit receives the output of the basic prediction unit and performs more refined risk identification and prediction adjustment; A parameter adaptive optimization module, connected to the dual-level prediction module, for monitoring the performance indicators of the dual-level prediction module and dynamically adjusting network parameters; A multimodal feature fusion module, connected to the two-level prediction module, is used to fuse different types of data features to form a unified multimodal feature representation; The time-sensitive risk assessment module is connected to the multimodal feature fusion module and is used to divide the anesthesia process into multiple key time periods, specifically process the data of each time period, dynamically capture the risk changes during the anesthesia process, and generate the risk probability of postoperative complications.

2. The intelligent prediction system for postoperative complications of anesthesia patients according to claim 1, characterized in that: The hard case sensitive preprocessing module includes: The hard example recognition unit is used to calculate the prediction deviation value of each sample and identify samples that are difficult to predict; Dynamic weight assignment unit, used to dynamically assign weights based on sample difficulty, giving higher weights to identified difficult samples and relatively lower weights to easily predicted samples; The weight smoothing unit is used to smooth the assigned weights to ensure a reasonable weight distribution and prevent individual extreme samples from having an excessive impact on the model.

3. The intelligent prediction system for postoperative complications of anesthesia patients according to claim 1, characterized in that: In the two-level prediction module: A feature enhancement channel is provided between the basic prediction unit and the fine adjustment unit, for transferring the feature map of the basic prediction unit to the fine adjustment unit, thereby realizing information flow and feature enhancement between models; The fine adjustment unit focuses on processing complex cases that are difficult for the basic prediction unit to accurately predict, and performs fine prediction adjustments on samples with large prediction deviations.

4. The intelligent prediction system for postoperative complications of anesthesia patients according to claim 1, characterized in that: The parameter adaptive optimization module includes: Parameter space construction unit, used to construct a multi-dimensional parameter space including model dimension, number of network layers, number of nodes, and learning rate; a heuristic search unit, for quickly locating a potential optimal parameter region in the parameter space; The progressive refinement search unit is used to adopt a gradually refined search strategy after determining the potential optimal area, continuously narrowing the search range and increasing the search accuracy, and finally locking in the optimal parameter combination; Cross-validation evaluation unit, used to evaluate each parameter combination through multiple cross-validation to ensure the generalization performance of the parameters; The dynamic adjustment convergence unit is used to dynamically adjust the optimization process according to the performance improvement. When the performance improvement of multiple consecutive iterations is less than the preset threshold, the search is automatically stopped and the final parameters are confirmed.

5. The intelligent prediction system for postoperative complications of anesthesia patients according to claim 1, characterized in that: The multimodal feature fusion module includes: Feature extraction subnetwork group: special feature extraction subnetworks are designed for different types of patient data, and each subnetwork is optimized for a specific data type; The time series processing unit is dedicated to processing time series data during anesthesia, capturing the trends and fluctuations of physiological parameters over time and extracting time series features; Feature fusion layer, used to effectively integrate features from different sub-networks to form a unified multimodal feature representation; The attention weighting unit is used to dynamically weight different modal features, strengthen important features and weaken minor features according to the specific patient situation.

6. The intelligent prediction system for postoperative complications of anesthesia patients according to claim 1, characterized in that: The timing-sensitive risk assessment module includes: The time series data segmentation processing unit is used to divide the anesthesia process into multiple key time periods such as anesthesia induction period, maintenance period, and awakening period, and perform special processing on the data of each time period; The sub-database time arrangement unit is used to create a dedicated sub-database for each time period, organize data in chronological order, and ensure the temporal integrity of the data; Sliding time window detection unit, which uses sliding time window technology to dynamically capture risk changes during anesthesia and achieve continuous monitoring and risk updates; The multi-time point forecast result fusion unit is used to integrate the forecast results at different time points and generate a comprehensive risk assessment through time-weighted averaging and other methods to ensure the temporal continuity and stability of the forecast; The risk level dynamic adjustment unit is used to dynamically adjust the patient's risk level based on real-time prediction results, and trigger an early warning mechanism when the risk exceeds a preset threshold.

7. The intelligent prediction system for postoperative complications of anesthesia patients according to claim 1, characterized in that: The patient's personal information includes: Basic physiological information, including patient age, gender, height, weight, and BMI index; Underlying disease information, including blood pressure level, underlying disease type and duration of illness, medication type and dosage, and medication duration; Preoperative examination information, including family history, medical history, drug allergy history, as well as preoperative physiological parameters, laboratory tests, pre-anesthesia examinations, pain scores, sleep conditions, and psychological status.

8. The intelligent prediction system for postoperative complications of anesthesia patients according to claim 1, characterized in that: The intraoperative information includes: The anesthesia operation information includes anesthesia induction, endotracheal intubation and anesthesia maintenance operations; The anesthesia induction includes general anesthesia induction assisted by endotracheal intubation, intravenous anesthesia induction, general anesthesia induction assisted by endotracheal intubation with preservation of spontaneous breathing, caudal block, regional anesthesia, basic laryngeal anesthesia and non-invasive regional anesthesia; The physiological parameters include data on circulation, respiration, and vascular compensation; The execution time of the intraoperative information is the time between entering anesthesia maintenance and ending surgical anesthesia.

9. The intelligent prediction system for postoperative complications of anesthesia patients according to claim 1, characterized in that: The postoperative complication risk probabilities generated by the time-sensitive risk assessment module include: The risk probability of postoperative complications for patients without underlying diseases before surgery; The risk probability of postoperative complications for patients with a type of underlying disease before surgery; The risk probability of postoperative complications for patients with multiple types of underlying diseases before surgery.

10. An intelligent prediction method for postoperative complications in anesthesia patients, characterized by: include: Collecting intraoperative information and personal information of anesthetized patients, and recording the execution time of the intraoperative information, wherein the intraoperative information includes anesthesia operation information, surgical operation information, injection information of anesthesia-related drugs, and physiological parameters; Preprocessing the collected intraoperative information of the anesthetized patient and the patient's personal information to identify samples that are difficult to predict and assign a higher weight to them; Processing the intraoperative information of the anesthetized patient and the patient's personal information through a two-level prediction model, wherein the two-level prediction model includes a basic prediction unit and a fine adjustment unit, wherein the basic prediction unit performs preliminary feature extraction and risk assessment on the input data to generate an initial prediction result and a feature map, and the fine adjustment unit receives the output of the basic prediction unit and performs more refined risk identification and prediction adjustment; Monitoring the performance indicators of the two-level prediction model and dynamically adjusting network parameters; Fuse different types of data features to form a unified multimodal feature representation; The anesthesia process is divided into multiple key time periods, and the data of each time period is specially processed to dynamically capture the risk changes during the anesthesia process and generate the risk probability of postoperative complications.

Images