Anesthesia complication prediction and avoidance auxiliary decision-making system based on deep learning

Abstract

The invention discloses an anesthesia complication prediction and avoidance auxiliary decision-making system based on deep learning, and belongs to the technical field of medical information. According to the system, basic information, preoperative examination data, anesthesia induction data, real-time physiological indexes and intra-operative operation information of a patient are collected firstly, and pre-processed data are obtained through multi-dimensional pre-processing; extracting multi-source feature data through the constructed multi-modal sub-network, and performing weighted fusion to generate a global feature vector; and training the initial convolutional neural network model based on historical data to obtain an anesthesia complication prediction model, outputting a prediction result in combination with the global feature vector after optimization of a loss function and an optimizer, and performing risk grading and tracing key factors. And finally, generating a personalized avoidance strategy according to the risk level and the key factors, converting the personalized avoidance strategy into personalized suggestions and performing visual display, and accurately predicting complications and assisting clinical decision making.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically a deep learning-based auxiliary decision-making system for predicting and avoiding anesthesia complications, aiming to achieve intelligent support throughout the entire process from risk prediction to proactive intervention. Background Technology

[0002] In surgery, anesthesia is an essential and crucial step. However, complications such as hypotension and arrhythmia during anesthesia can not only interfere with the normal progress of the operation but may even pose a serious threat to the patient's life. Therefore, how to accurately predict and effectively prevent the occurrence of anesthetic complications has long been a core issue in medical research. In the past, the prediction of anesthetic complications mainly relied on the doctor's personal experience and the patient's past medical history. This method has obvious subjective bias and relatively low accuracy. However, in recent years, with the rapid development of deep learning technology, its application in the medical field has gradually emerged, achieving remarkable results. Deep learning can automatically extract features from massive amounts of medical data and then build predictive models, providing a new approach to accurately predicting anesthetic complications. However, even so, existing deep learning-based prediction methods still have problems such as complex model architecture, high computational cost, and less than ideal prediction accuracy.

[0003] Existing patent (CN116312958B) discloses an anesthesia risk early warning system, an emergency management system, and a method. The anesthesia risk early warning system includes: a monitoring terminal and a monitoring terminal. The monitoring terminal acquires patient monitoring data during anesthesia, stores the monitoring data, and transmits it encrypted to the monitoring terminal. The monitoring terminal receives and verifies the monitoring data, predicts the patient's monitoring data at a target time based on the current monitoring data, and provides an anesthesia risk warning based on the current and target time monitoring data. This invention securely transmits patient monitoring data through the monitoring terminal and the monitoring terminal, accurately predicts the patient's anesthesia risk, and monitors the on-duty status of the attending anesthesiologist, thus realizing a system design encompassing wireless monitoring data acquisition, anesthesia risk prediction, and emergency response. While the above invention can provide anesthesiologists with comprehensive, safe, and efficient anesthesia services, the early warning system only predicts future risks based on current monitoring data and cannot fully capture all risk factors, especially when facing the complex and ever-changing physiological states of patients. Therefore, the accuracy of the prediction may be limited.

[0004] Existing patent (CN118629649A) discloses a method for constructing a prediction model for anesthesia complications based on deep learning. The method includes the following steps: acquiring a historical multidimensional dataset, preprocessing and labeling the historical multidimensional dataset; the historical multidimensional dataset includes preoperative, intraoperative, and postoperative datasets; extracting key features characterizing anesthesia complications from the preoperative, intraoperative, and postoperative datasets respectively; constructing a time-series prediction network and training the time-series prediction network using the extracted key features to obtain an anesthesia complication prediction model; testing and validating the trained anesthesia complication prediction model; inputting preoperative, intraoperative, or postoperative data into the trained anesthesia complication prediction model to predict anesthesia complications at different stages. This invention can effectively predict anesthesia complications that may occur during surgery. The above invention focuses on the phased prediction of anesthesia complications, relying primarily on preoperative, intraoperative, and postoperative time-series data and an LSTM network to construct a prediction model, but it has the following limitations: (1) it only achieves risk prediction and does not involve proactive decision support for complication avoidance; (2) it cannot dynamically adjust intervention strategies based on real-time patient conditions.

[0005] Therefore, a deep learning-based decision support system for predicting and avoiding anesthesia complications is proposed. Summary of the Invention

[0006] To address the aforementioned technical problems, the present invention aims to provide a deep learning-based auxiliary decision-making system for predicting and avoiding anesthesia complications.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based anesthesia complication prediction and avoidance auxiliary decision-making system, the system comprising: a multi-source data acquisition module, a data preprocessing module, a data feature extraction module, a feature data fusion module, a prediction model construction module, a prediction model optimization module, a risk grading and key factor analysis module, and an avoidance auxiliary decision generation module;

[0008] The multi-source data acquisition module is used to collect patients' basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information;

[0009] The data preprocessing module is used to perform multidimensional preprocessing on the collected patient's basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators and intraoperative operation information (first perform basic preprocessing on the collected data, and then label the data after basic preprocessing) to obtain the corresponding preprocessed data.

[0010] The data feature extraction module constructs a multimodal subnetwork, inputs preprocessed data into the multimodal subnetwork for feature extraction, and obtains corresponding multi-source feature data;

[0011] The feature data fusion module is used to perform weighted fusion of the feature data to generate a global feature vector;

[0012] The prediction model building module is used to train the initial convolutional neural network model based on the basic information of the historical collection period, preoperative examination data, anesthesia induction data, real-time physiological indicators and intraoperative operation information, so as to obtain the anesthesia complication prediction model.

[0013] The prediction model optimization module optimizes the parameters and structure of the anesthesia complication prediction model based on the loss function and Adam optimizer, and obtains the optimized anesthesia complication prediction model.

[0014] The risk grading and key factor analysis module obtains the prediction results of anesthesia complications for the corresponding patients based on the anesthesia complication prediction model and global feature vector, performs risk grading on the prediction results, obtains the corresponding risk level, and traces the key factors that lead to anesthesia complications.

[0015] The risk avoidance auxiliary decision generation module generates personalized avoidance strategies based on the risk level and key factors, and then obtains personalized avoidance suggestions and visualizations based on the personalized avoidance strategies.

[0016] Furthermore, the process of collecting basic patient information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative procedure information includes:

[0017] Through the interfaces of the hospital information system and electronic medical record system, the system automatically captures the patient's basic information and preoperative examination data; records anesthesia induction data manually and through drug infusion pump equipment; connects to multi-parameter monitoring equipment to collect physiological indicators in real time during anesthesia induction and maintenance; collects intraoperative operation information through the surgical anesthesia information system; and unifies the collection cycle, which includes several collection moments.

[0018] Furthermore, the process of performing multidimensional preprocessing on the collected basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information to obtain the corresponding preprocessed data includes:

[0019] The 3σ rule was used to identify outliers in physiological indicators, and secondary verification was performed using clinical common sense to remove invalid values. For isolated points that deviated significantly from the trend, the mean of adjacent time points was used for replacement. For short-term missing data, linear interpolation was used for imputation. For long-term missing data, K-nearest neighbor algorithm was used for imputation based on historical data of patients of the same age group and disease. Z-score standardization was used for continuous data to eliminate the influence of dimensions. The preprocessed data was associated with the patient's unique identifier and combined with the unprocessed data to generate preprocessed data.

[0020] Furthermore, the process of constructing a multimodal subnetwork includes:

[0021] The multimodal subnetwork includes a temporal data subnetwork, a static data subnetwork, and an image data subnetwork;

[0022] Based on the preprocessed data from several historical collection periods, corresponding training sample sets and test sample sets are formed; the training sample sets are divided into time-series data training sample sets, structured data training sample sets, and unstructured data training sample sets; the test sample sets are divided into time-series data test sample sets, structured data test sample sets, and unstructured data test sample sets.

[0023] The training sample set of time-series data, the training sample set of structured data, the training sample set of unstructured data, the test sample set of time-series data, the test sample set of structured data, and the test sample set of unstructured data are input into the deep learning model for training. The trained deep learning model is denoted as a multimodal sub-network. The multimodal sub-network includes a time-series data branch network, a structured data branch network, and an unstructured data branch network.

[0024] Furthermore, the process of inputting the preprocessed data into the multimodal subnetwork for feature extraction to obtain the corresponding feature data includes:

[0025] The preprocessed data is input into a multimodal subnetwork. The patient's preprocessed data is then subjected to feature extraction through the temporal data branch network, structured data branch network, and unstructured data branch network in the multimodal subnetwork. This process yields the corresponding patient's temporal feature data, structured feature data, and unstructured feature data, which are then referred to as multi-source feature data.

[0026] Furthermore, the process of weighted fusion of the multi-source feature data to generate a global feature vector includes:

[0027] The time-series feature data, structured feature data, and unstructured feature data are aligned. The time-series feature data, structured feature data, and unstructured feature data after feature alignment are combined to form a temporary feature matrix.

[0028] Based on a single-layer fully connected network, a nonlinear transformation is performed on the temporary feature matrix to output the weight coefficients of each feature data.

[0029] A global feature vector is generated based on the time-series feature data, structured feature data, unstructured feature data, and the weight coefficients of the corresponding feature data.

[0030] Furthermore, the process of training the initial convolutional neural network model based on historical data collection period information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information to obtain an anesthesia complication prediction model includes:

[0031] The basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators and intraoperative operation information of several historical collection periods are grouped and labeled as f = 1, 2, 3, ..., e; e is a natural number.

[0032] The basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information of the historical collection period of group eh are used as sample data, where h is a natural number less than e. The mean of the sample data is obtained using the sample data and is denoted as the sample set. The basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information of the historical collection period of the other groups are used as the test set. A training sample set is formed based on the sample set and the test set.

[0033] A standard prediction model is constructed based on convolutional neural networks.

[0034] The training sample set is then input into the standard prediction model to train it, and the trained standard prediction model is recorded as the anesthesia complication prediction model.

[0035] Furthermore, based on the loss function and the Adam optimizer, the process of optimizing the parameters and structure of the anesthesia complication prediction model to obtain the optimized anesthesia complication prediction model includes:

[0036] Based on the anesthetic complication prediction model, the failure risk coefficient of the anesthetic complication prediction model is obtained, and the anesthetic complication prediction model is optimized using the failure risk coefficient.

[0037] The fault risk coefficient is classified into multiple risk thresholds to obtain the category prediction value; the category prediction value and the true label are used to calculate the error through a loss function to obtain the corresponding loss value;

[0038] The model parameters of the anesthesia complication prediction model are updated based on the Adam optimizer and the loss value, thereby obtaining the optimized anesthesia complication prediction model.

[0039] Furthermore, based on the anesthetic complication prediction model and global feature vector, the prediction results of anesthetic complications for the corresponding patients are obtained. The prediction results are then risk-classified to obtain the corresponding risk level. The process of tracing the key factors leading to anesthetic complications includes:

[0040] The global feature vector of the patient to be predicted is input into the anesthesia complication prediction model, and the prediction results of the corresponding anesthesia complications are output.

[0041] The prediction results include the probability of occurrence of various complications. Based on the probability of occurrence of each complication, a risk level is divided, and the probability of occurrence is marked as φ. If the probability of occurrence φ < 10%, it is classified as a low-risk level; if the probability of occurrence φ ∈ [10%, 30%], it is classified as a medium-risk level; if the probability of occurrence φ > 30%, it is classified as a high-risk level.

[0042] The SHAP analysis model is used to calculate the SHAP value of each feature data in the global feature vector for the prediction result; the corresponding key factors are extracted by sorting them according to the absolute value of the SHAP value, and a weighted list of key factors is generated.

[0043] Furthermore, based on the aforementioned risk level and key factors, a personalized avoidance strategy is generated. Then, based on this personalized avoidance strategy, personalized avoidance suggestions and visual representations are obtained. This process includes:

[0044] Construct a knowledge base for complication avoidance strategies; based on risk levels and key factors, and according to the knowledge base for complication avoidance strategies, generate basic strategies; the basic strategies include reinforcement intervention strategies, preventive intervention strategies, and monitoring strategies;

[0045] The basic strategy is dynamically optimized by reinforcement learning model to generate personalized avoidance strategy, including: constructing state space, defining action space, designing reward function and iterative optimization; the effect of intervention measures is pre-simulated in a simulation environment by deep deterministic policy gradient algorithm, the policy parameters are adjusted according to reward function, and finally personalized avoidance strategy adapted to individual patient characteristics is output.

[0046] Transform personalized avoidance strategies into visualized personalized avoidance suggestions and visualizations, including: visualizations and personalized avoidance suggestions.

[0047] Compared with existing technologies, the beneficial effects of this invention are as follows: A multi-source data acquisition module integrates various types of data, including patient basic information, preoperative examination data, and real-time physiological indicators, unifying the acquisition cycle and classifying and storing them to provide comprehensive information support for prediction; a data preprocessing module performs outlier identification, missing value imputation, and standardization, combined with the patient's unique identifier to ensure data quality and traceability, providing high-quality input for model training; the constructed multimodal subnetwork includes time-series, static, and image data subnetworks, selectively extracting features from different types of data to fully explore the potential value of the data; and a feature data fusion module uses an attention mechanism to weightedly fuse multi-source features, highlighting key features. The system utilizes key feature contributions to generate global feature vectors that accurately reflect risk associations. The prediction model construction and optimization module combines convolutional neural networks, loss functions, and optimizers to dynamically adjust parameter structures, introducing hierarchical optimization of fault risk coefficients to improve model prediction accuracy and generalization ability. The risk grading and key factor analysis module clarifies risk levels through probability thresholds and traces key factors using SHAP values, enhancing decision interpretability. The avoidance-assisted decision generation module generates basic strategies based on a knowledge base integrating guidelines, expert experience, and cases. These strategies are then dynamically optimized through reinforcement learning to output personalized avoidance strategies, which are then transformed into a visual interface, achieving a closed-loop support system of "prediction-decision-feedback." The overall system provides full-process assistance for anesthesia complications, improving prediction accuracy, intervention precision, and clinical applicability, effectively reducing the incidence of complications, and improving anesthesia safety and decision-making efficiency. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0049] Figure 1 This is a schematic diagram illustrating the steps of a deep learning-based auxiliary decision-making system for predicting and avoiding anesthesia complications.

[0050] Figure 2 This is a schematic diagram of a module of a deep learning-based auxiliary decision-making system for predicting and avoiding anesthesia complications. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0052] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0053] like Figure 1 As shown, a deep learning-based anesthesia complication prediction and avoidance auxiliary decision-making system includes: a multi-source data acquisition module, a data preprocessing module, a data feature extraction module, a feature data fusion module, a prediction model construction module, a prediction model optimization module, a risk classification and key factor analysis module, and an avoidance auxiliary decision generation module.

[0054] The multi-source data acquisition module is used to collect patients' basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information;

[0055] The data preprocessing module is used to perform multidimensional preprocessing on the collected patient's basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information to obtain the corresponding preprocessed data.

[0056] The data feature extraction module constructs a multimodal subnetwork, inputs preprocessed data into the multimodal subnetwork for feature extraction, and obtains corresponding multi-source feature data;

[0057] The feature data fusion module is used to perform weighted fusion of the feature data to generate a global feature vector;

[0058] The prediction model building module is used to train the initial convolutional neural network model based on the basic information of the historical collection period, preoperative examination data, anesthesia induction data, real-time physiological indicators and intraoperative operation information, so as to obtain the anesthesia complication prediction model.

[0059] The prediction model optimization module optimizes the parameters and structure of the anesthesia complication prediction model based on the loss function and Adam optimizer, and obtains the optimized anesthesia complication prediction model.

[0060] The risk grading and key factor analysis module obtains the prediction results of anesthesia complications for the corresponding patients based on the anesthesia complication prediction model and global feature vector, performs risk grading on the prediction results, obtains the corresponding risk level, and traces the key factors that lead to anesthesia complications.

[0061] The risk avoidance auxiliary decision generation module generates personalized avoidance strategies based on the risk level and key factors, and then obtains personalized avoidance suggestions and visualizations based on the personalized avoidance strategies.

[0062] It should be further explained that, in the specific implementation process, the collection of the patient's basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information includes:

[0063] Optionally, in this embodiment, the patient's basic information and preoperative examination data are automatically captured through the interface of the hospital information system and electronic medical record system; anesthesia induction data are recorded manually and through drug infusion pump equipment; multi-parameter monitoring equipment is connected to collect physiological indicators during anesthesia induction and maintenance in real time; and intraoperative operation information is collected through the surgical anesthesia information system.

[0064] It should be noted that the above basic information includes, but is not limited to, age, gender, height, weight, past medical history, allergy history, and surgical history. The above preoperative examination data includes, but is not limited to, complete blood count reports, liver and kidney function reports, coagulation function reports, electrocardiogram reports, and chest X-ray reports. The above multi-parameter monitoring equipment includes, but is not limited to, electrocardiogram monitors, invasive blood pressure monitors, pulse oximeters, and bispectral index (BIS) monitors. The above physiological indicators include, but are not limited to, heart rate (HR), systolic blood pressure (SBP), diastolic blood pressure (DBP), mean arterial pressure (MAP), blood oxygen saturation (SpO2), and BIS values. The above anesthesia induction data includes, but is not limited to, the type, dosage, infusion rate, and administration time of anesthetic induction drugs (such as propofol and fentanyl) and maintenance drugs (such as sevoflurane and remifentanil). The above intraoperative operation information includes, but is not limited to, the type of surgery (such as laparoscopic surgery and open thoracotomy), operation duration, incision location, blood loss, fluid infusion volume, blood transfusion status, and special intraoperative procedures (such as endotracheal intubation depth and mechanical ventilation parameters).

[0065] To further explain, a unified collection cycle is established for the above data collection process. This collection cycle includes several collection moments, the specific number of which is set according to the actual situation. Data is classified according to its type, including but not limited to structured data, time-series data, and unstructured data. The collected structured data (such as age and blood pressure values) is stored in CSV format, time-series data (such as heart rate changes over time) is stored in a time-series database (such as InfluxDB), and unstructured data (such as intraoperative images and doctors' handwritten records) is converted into structured text using OCR technology before storage, thereby improving the authenticity and reliability of the data.

[0066] It should be further explained that, in the specific implementation process, the collection of basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information undergoes multidimensional preprocessing to obtain the corresponding preprocessed data. The specific process includes:

[0067] Optionally, in the embodiments of this application, the 3σ rule is used to identify abnormal values ​​in physiological indicators (such as a sudden drop in blood pressure to 0 due to equipment detachment), and combined with clinical common sense (such as the heart rate cannot exceed 300 beats / minute) for secondary verification to eliminate invalid values; for isolated points that deviate significantly from the trend, the average of adjacent time points is used for replacement, and the specific process will not be described in this application.

[0068] Optionally, in the embodiments of this application, for short-term missing data (such as missing physiological indicators within 10 seconds), linear interpolation is used to fill in the missing data; for long-term missing data (such as a certain indicator not detected in the preoperative examination), the missing data is filled in using the K-nearest neighbor (KNN) algorithm based on historical data of patients of the same age and the same disease. The specific process will not be described in detail in this application.

[0069] Optionally, in the embodiments of this application, continuous data (such as blood pressure, blood loss, infusion volume, etc.) are standardized using Z-score to eliminate the influence of dimensions. The specific process will not be described in detail in this application.

[0070] Optionally, in this embodiment of the application, the preprocessed data is associated with the patient's unique identifier (such as hospital number) and combined with the unprocessed data to generate preprocessed data, ensuring data traceability.

[0071] It should be noted that the preprocessed data includes preprocessed data, unprocessed data, and a unique identifier, which are not limited in this application.

[0072] It should be further explained that, in the specific implementation process, the construction of the multimodal subnetwork includes:

[0073] It should be noted that the multimodal subnetwork includes a time-series data subnetwork, a static data subnetwork, and an image data subnetwork.

[0074] Optionally, in this embodiment of the application, a corresponding training sample set and a test sample set are formed based on the preprocessed data of several groups of historical collection periods;

[0075] It should be noted that the above training sample sets are divided into time-series data training sample sets, structured data training sample sets, and unstructured data training sample sets. The above test sample sets are divided into time-series data test sample sets, structured data test sample sets, and unstructured data test sample sets. The specific classification process of the training and test sample sets will not be elaborated here.

[0076] Optionally, in this embodiment, the time-series data training sample set, the structured data training sample set, the unstructured data training sample set, the time-series data test sample set, the structured data test sample set, and the unstructured data test sample set are input into the deep learning model for training, and the trained deep learning model is denoted as a multimodal sub-network; wherein, the multimodal sub-network includes a time-series data branch network, a structured data branch network, and an unstructured data branch network.

[0077] It should be noted that the aforementioned time-series data branch network is used to extract feature data corresponding to the time-series data in the preprocessed data. The aforementioned structured data branch network is used to extract feature data corresponding to the structured data in the preprocessed data. The aforementioned unstructured data branch network is used to extract feature data corresponding to the unstructured data in the preprocessed data.

[0078] It should be further explained that, in the specific implementation process, the preprocessed data is input into the multimodal subnetwork for feature extraction to obtain the corresponding feature data. The specific process includes:

[0079] Optionally, in this embodiment of the application, the patient's preprocessed data is input into a multimodal subnetwork, and the patient's preprocessed data is feature extracted through the temporal data branch network, structured data branch network and unstructured data branch network in the multimodal subnetwork, thereby obtaining the corresponding patient's temporal feature data, structured feature data and unstructured feature data, and the temporal feature data, structured feature data and unstructured feature data are recorded as multi-source feature data.

[0080] It should be noted that the time-series feature data, structured feature data, and unstructured feature data are stored to facilitate model training and optimization.

[0081] It should be further explained that, in the specific implementation process, the weighted fusion of the multi-source feature data to generate a global feature vector includes:

[0082] Optionally, in this embodiment, the temporal feature data, structured feature data, and unstructured feature data in the multi-source feature data are denoted as R, T, and K, respectively; the temporal feature data R, structured feature data T, and unstructured feature data K are subjected to feature alignment processing, and the feature-aligned temporal feature data R, structured feature data T, and unstructured feature data K are used to form a temporary feature matrix. in, The temporal feature data after feature alignment processing. The structured feature data after feature alignment processing. This refers to unstructured feature data after feature alignment processing.

[0083] It should be noted that, since the feature vectors output by different branches of the network have different dimensions, the above feature alignment process needs to be mapped to the same dimension d (e.g., uniformly 128 dimensions) through a projection layer to ensure that they can be spliced ​​and calculated subsequently.

[0084] It should be noted that, based on a single-layer fully connected network, a nonlinear transformation is performed on the temporary feature matrix H to output the weight coefficients of each feature data, specifically:

[0085] Fully connected layer transformation, let the weight matrix of the fully connected layer be W. a ∈R k×d The bias is b a ∈R k , a = 1, 2, 3; using the activation function tanh, obtain the intermediate features U = tanh(W) corresponding to one modality in each column. a ·H+b a );

[0086] The k-dimensional intermediate features U are compressed into 1-dimensional weight coefficients through a linear layer. Let the weight of this linear layer be Q. a ∈R 1×k Then w a =Q a ·U; among them, w a This represents the weight coefficient of the corresponding feature data;

[0087] Based on the time-series feature data R, structured feature data T, unstructured feature data K, and the corresponding weight coefficients w of the feature data. a Generate global feature vectors;

[0088] To further explain, the global feature vector is: G = w1 × R + w2 × T + w3 × K; where G represents the global feature vector; w1 represents the weight coefficient of the time-series feature data R; w2 represents the weight coefficient of the structured feature data T; and w3 represents the weight coefficient of the unstructured feature data K.

[0089] It should be further explained that, in the specific implementation process, the initial convolutional neural network model is trained based on basic information from historical data collection periods, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information to obtain an anesthesia complication prediction model. The specific process includes:

[0090] Optionally, in the embodiments of this application, the basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators and intraoperative operation information of several historical collection periods are grouped and labeled as f = 1, 2, 3, ..., e; e is a natural number;

[0091] The basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information of the historical collection period of group eh are used as sample data, where h is a natural number less than e. The mean of the sample data is obtained using the sample data and is denoted as the sample set. The basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information of the historical collection period of the other groups are used as the test set. A training sample set is formed based on the sample set and the test set.

[0092] A standard prediction model is constructed based on convolutional neural networks.

[0093] The training sample set is then input into the standard prediction model to train it, and the trained standard prediction model is recorded as the anesthesia complication prediction model.

[0094] It should be further explained that, in the specific implementation process, the process of optimizing the parameters and structure of the anesthesia complication prediction model based on the loss function and the Adam optimizer to obtain the optimized anesthesia complication prediction model includes:

[0095] Optionally, in this embodiment of the application, based on the anesthetic complication prediction model, a failure risk coefficient γ of the anesthetic complication prediction model is obtained, and the anesthetic complication prediction model is optimized using the failure risk coefficient γ, specifically as follows:

[0096] The failure risk coefficient γ is classified into multiple risk thresholds to obtain the category prediction value D; the category prediction value D is:

[0097]

[0098] It should be noted that the category prediction values ​​include: unoptimized category prediction values, slightly optimized category prediction values, and deeply optimized category prediction values.

[0099] The error between the predicted category value and the true label is calculated using a loss function, as shown in the formula:

[0100] Where F is the loss value; The true label is represented by a value of 0 or 1; O1 represents the probability that the corresponding category prediction value D is 0; O2 represents the probability that the corresponding category prediction value D is 1.

[0101] The model parameters of the anesthesia complication prediction model are updated using the Adam optimizer to obtain the optimized anesthesia complication prediction model.

[0102] To further explain, the update is based on minimizing the loss function, and the update formula is:

[0103] in, τ represents the learning rate; τ represents the model parameters of the anesthesia complication prediction model.

[0104] It should be further explained that, in the specific implementation process, based on the anesthesia complication prediction model and global feature vector, the prediction results of the corresponding patient's anesthesia complications are obtained. The prediction results are then risk-classified to obtain the corresponding risk level. The specific process of tracing the key factors leading to anesthesia complications includes:

[0105] Optionally, in this embodiment of the application, the global feature vector of the patient to be predicted is input into the anesthesia complication prediction model, and the prediction result of the corresponding patient's anesthesia complications is output;

[0106] It should be noted that the above prediction results include the probability of occurrence of various complications. The risk level is divided according to the probability of occurrence of each type of complication, and the probability of occurrence is marked as φ. If the probability of occurrence φ < 10%, it is classified as low risk level; if the probability of occurrence φ ∈ [10%, 30%], it is classified as medium risk level; if the probability of occurrence φ > 30%, it is classified as high risk level.

[0107] Optionally, in the embodiments of this application, the SHAP analysis model is used to calculate the SHAP value of each feature data in the global feature vector for the prediction result (a positive SHAP value indicates that the feature increases the risk of complications, and a negative value indicates that the risk is reduced).

[0108] Sort by absolute SHAP value, extract corresponding key factors (such as "propofol infusion rate > 5 mg / kg / h" and "preoperative MAP < 80 mmHg"), and generate a weighted list of key factors.

[0109] It should be noted that the weights mentioned above are the percentage of the absolute value of the SHAP value.

[0110] It should be further explained that, in the specific implementation process, the process of generating personalized avoidance strategies based on the aforementioned risk levels and key factors, and then obtaining personalized avoidance suggestions and visualizations based on these personalized avoidance strategies, includes the following:

[0111] Optionally, in this embodiment of the application, a complication avoidance strategy knowledge base is constructed as the basic support for generating personalized avoidance strategies.

[0112] It should be noted that the aforementioned knowledge base for complication avoidance strategies includes authoritative guideline data, expert experience data, and historical case data.

[0113] To further clarify, the aforementioned authoritative guideline data is derived from international guidelines such as the *American Society of Anesthesiologists (ASA) Clinical Practice Guidelines* and the *European Society of Anesthesiology (ESA) Guidelines for the Prevention and Treatment of Complications*, as well as the domestic *Expert Consensus on the Management of Anesthesia Complications*. These guidelines extract standardized intervention frameworks for different types of complications, clearly defining core elements such as intervention timing, drug selection, and dosage range. The aforementioned expert experience data includes the practical experience of chief physicians in the anesthesiology departments of tertiary hospitals. The aforementioned historical case data includes cases of successful complication avoidance screened from the hospital's anesthesia case database over the past five years. Patient characteristics, intervention measures, and effectiveness evaluation indicators from these cases were extracted to form a case database and effectiveness mapping table. This application does not impose further limitations.

[0114] Optionally, in this embodiment of the application, a basic strategy is generated based on the risk level and key factors, and according to a knowledge base of complication avoidance strategies, specifically as follows:

[0115] It should be noted that the basic strategies include reinforcement intervention strategies, preventive intervention strategies, and monitoring strategies.

[0116] For high-risk levels, high-priority measures directly related to key factors are matched from the complication avoidance strategy knowledge base, thereby generating enhanced intervention strategies.

[0117] In an exemplary embodiment, if the key factor is "propofol infusion rate > 5 mg / kg / h" and the risk level is high-risk hypotension, then the command "immediately reduce propofol rate to below 3 mg / kg / h + intravenous infusion of norepinephrine 0.05-0.1 μg / kg / min + monitor blood pressure every 5 minutes" is invoked.

[0118] For medium-risk levels, preventive measures are selected based on the potential impact of key factors, thereby generating preventive intervention strategies.

[0119] In an exemplary embodiment, if the key factor is "female + history of postoperative nausea and vomiting" and the risk level is medium risk, then "intravenous injection of ondansetron 4mg 30 minutes before surgery + intraoperative restriction of opioid dosage" is invoked.

[0120] For low-risk levels, the focus is on increasing monitoring frequency, thereby generating monitoring strategies.

[0121] It should be noted that the aforementioned enhanced intervention strategies, preventive intervention strategies, and monitoring strategies can be adjusted by doctors according to the actual situation, and this application does not impose further restrictions.

[0122] Optionally, in this embodiment of the application, the basic policy is dynamically optimized using a reinforcement learning model to generate a personalized avoidance policy, specifically as follows:

[0123] State Space Construction: Real-time physiological indicators, key factors, and risk level of the patient are used as state variables. Action Space Definition: Includes adjustment options for intervention measures (such as the magnitude of drug dosage increase / decrease, route of administration switch, and monitoring frequency adjustment). Reward Function Design: The core reward is the reduction in the probability of complication, with the incidence of side effects of the intervention as the penalty. Iterative Optimization: The effect of the intervention is pre-simulated in a simulated environment using the Deep Deterministic Policy Gradient (DDPG) algorithm. Policy parameters are adjusted based on the reward function, ultimately outputting a personalized avoidance strategy adapted to the individual characteristics of the patient.

[0124] Optionally, in this embodiment of the application, the personalized avoidance strategy is transformed into a visualized personalized avoidance suggestion and a visual display, specifically as follows:

[0125] Visualization: A three-color dashboard of red (high risk), yellow (medium risk), and green (low risk) is used to display the real-time risk level of various complications, mark the probability of occurrence and key factors, and display the risk trend (probability change over the past 30 minutes) through a dynamic curve.

[0126] Personalized avoidance recommendations: sorted by implementation priority (marked with ★, ★★★ being the highest). Each recommendation includes the following steps (e.g., "Step 1: Adjust the propofol infusion pump to 3 mg / kg / h; Step 2: Connect the norepinephrine injector and set the rate to 0.08 μg / kg / min"), expected effect (e.g., "MAP rises to above 70 mmHg within 10 minutes"), and precautions (e.g., "Monitor heart rate; if it is >120 beats / minute, the norepinephrine dose needs to be reduced").

[0127] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0128] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0129] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more electronic devices to execute all or part of the steps of the methods described in the various embodiments of this application.

[0130] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0131] In the several embodiments provided in this application, it should be understood that the disclosed application can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0132] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0133] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0134] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A deep learning-based decision support system for predicting and avoiding anesthesia complications, characterized in that, The system includes: a multi-source data acquisition module, a data preprocessing module, a data feature extraction module, a feature data fusion module, a prediction model construction module, a prediction model optimization module, a risk classification and key factor analysis module, and an avoidance auxiliary decision generation module; The multi-source data acquisition module is used to collect patients' basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information; The data preprocessing module is used to perform multidimensional preprocessing on the collected patient's basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information to obtain the corresponding preprocessed data. The data feature extraction module constructs a multimodal subnetwork, inputs preprocessed data into the multimodal subnetwork for feature extraction, and obtains corresponding multi-source feature data; The feature data fusion module is used to perform weighted fusion of the feature data to generate a global feature vector; The prediction model building module is used to train the initial convolutional neural network model based on the basic information of the historical collection period, preoperative examination data, anesthesia induction data, real-time physiological indicators and intraoperative operation information, so as to obtain the anesthesia complication prediction model. The prediction model optimization module optimizes the parameters and structure of the anesthesia complication prediction model based on the loss function and Adam optimizer, and obtains the optimized anesthesia complication prediction model. The risk grading and key factor analysis module obtains the prediction results of anesthesia complications for the corresponding patients based on the anesthesia complication prediction model and global feature vector, performs risk grading on the prediction results, obtains the corresponding risk level, and traces the key factors that lead to anesthesia complications. The risk avoidance auxiliary decision generation module generates personalized avoidance strategies based on the risk level and key factors, and then obtains personalized avoidance suggestions and visualizations based on the personalized avoidance strategies.

2. The deep learning-based anesthetic complication prediction and avoidance auxiliary decision-making system according to claim 1, characterized in that, The process of collecting basic patient information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative procedure information includes: Through the interfaces of the hospital information system and electronic medical record system, the system automatically captures the patient's basic information and preoperative examination data; records anesthesia induction data manually and through drug infusion pump equipment; connects to multi-parameter monitoring equipment to collect physiological indicators in real time during anesthesia induction and maintenance; collects intraoperative operation information through the surgical anesthesia information system; and unifies the collection cycle, which includes several collection moments.

3. The deep learning-based anesthetic complication prediction and avoidance auxiliary decision-making system according to claim 2, characterized in that, The process of performing multidimensional preprocessing on the collected basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information to obtain the corresponding preprocessed data includes: The 3σ rule was used to identify outliers in physiological indicators, and secondary verification was performed using clinical common sense to remove invalid values. For isolated points that deviated significantly from the trend, the mean of adjacent time points was used for replacement. For short-term missing data, linear interpolation was used for imputation. For long-term missing data, K-nearest neighbor algorithm was used for imputation based on historical data of patients of the same age group and disease. Z-score standardization was used for continuous data to eliminate the influence of dimensions. The preprocessed data was associated with the patient's unique identifier and combined with the unprocessed data to generate preprocessed data.

4. The deep learning-based anesthetic complication prediction and avoidance auxiliary decision-making system according to claim 3, characterized in that, The process of constructing a multimodal subnetwork includes: The multimodal subnetwork includes a temporal data subnetwork, a static data subnetwork, and an image data subnetwork; Based on the preprocessed data from several historical collection periods, corresponding training sample sets and test sample sets are formed; the training sample sets are divided into time-series data training sample sets, structured data training sample sets, and unstructured data training sample sets; the test sample sets are divided into time-series data test sample sets, structured data test sample sets, and unstructured data test sample sets. The training sample set of time-series data, the training sample set of structured data, the training sample set of unstructured data, the test sample set of time-series data, the test sample set of structured data, and the test sample set of unstructured data are input into the deep learning model for training. The trained deep learning model is denoted as a multimodal sub-network. The multimodal sub-network includes a time-series data branch network, a structured data branch network, and an unstructured data branch network.

5. The deep learning-based anesthetic complication prediction and avoidance auxiliary decision-making system according to claim 4, characterized in that, The process of inputting preprocessed data into a multimodal subnetwork for feature extraction to obtain the corresponding feature data includes: The preprocessed data is input into a multimodal subnetwork. The patient's preprocessed data is then subjected to feature extraction through the temporal data branch network, structured data branch network, and unstructured data branch network in the multimodal subnetwork. This process yields the corresponding patient's temporal feature data, structured feature data, and unstructured feature data, which are then referred to as multi-source feature data.

6. The deep learning-based anesthetic complication prediction and avoidance auxiliary decision-making system according to claim 5, characterized in that, The process of weighted fusion of the multi-source feature data to generate a global feature vector includes: The time-series feature data, structured feature data, and unstructured feature data are aligned. The time-series feature data, structured feature data, and unstructured feature data after feature alignment are combined to form a temporary feature matrix. Based on a single-layer fully connected network, a nonlinear transformation is performed on the temporary feature matrix to output the weight coefficients of each feature data. A global feature vector is generated based on the time-series feature data, structured feature data, unstructured feature data, and the weight coefficients of the corresponding feature data.

7. The deep learning-based anesthetic complication prediction and avoidance auxiliary decision-making system according to claim 6, characterized in that, The process of training an initial convolutional neural network model based on historical data collection period information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information to obtain an anesthesia complication prediction model includes: The basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators and intraoperative operation information of several historical collection periods are grouped and labeled as f = 1, 2, 3, ..., e; e is a natural number. The basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information of the historical collection period of group eh are used as sample data, where h is a natural number less than e. The mean of the sample data is obtained using the sample data and is denoted as the sample set. The basic information, preoperative examination data, anesthesia induction data, real-time physiological indicators, and intraoperative operation information of the historical collection period of the other groups are used as the test set. A training sample set is formed based on the sample set and the test set. A standard prediction model is constructed based on convolutional neural networks. The training sample set is then input into the standard prediction model to train it, and the trained standard prediction model is recorded as the anesthesia complication prediction model.

8. The deep learning-based anesthetic complication prediction and avoidance auxiliary decision-making system according to claim 7, characterized in that, The process of optimizing the parameters and structure of the anesthesia complication prediction model based on the loss function and Adam optimizer to obtain the optimized anesthesia complication prediction model includes: Based on the anesthetic complication prediction model, the failure risk coefficient of the anesthetic complication prediction model is obtained, and the anesthetic complication prediction model is optimized using the failure risk coefficient. The fault risk coefficient is classified into multiple risk thresholds to obtain the category prediction value; the category prediction value and the true label are used to calculate the error through a loss function to obtain the corresponding loss value; The model parameters of the anesthesia complication prediction model are updated based on the Adam optimizer and the loss value, thereby obtaining the optimized anesthesia complication prediction model.

9. The deep learning-based anesthetic complication prediction and avoidance auxiliary decision-making system according to claim 8, characterized in that, Based on the anesthesia complication prediction model and global feature vector, the prediction results of anesthesia complications for corresponding patients are obtained. The prediction results are then risk-classified to obtain the corresponding risk level. The process of tracing the key factors leading to anesthesia complications includes: The global feature vector of the patient to be predicted is input into the anesthesia complication prediction model, and the prediction results of the corresponding anesthesia complications are output. The prediction results include the probability of occurrence of various complications. Based on the probability of occurrence of each complication, a risk level is divided, and the probability of occurrence is marked as φ. If the probability of occurrence φ < 10%, it is classified as a low-risk level; if the probability of occurrence φ ∈ [10%, 30%], it is classified as a medium-risk level; if the probability of occurrence φ > 30%, it is classified as a high-risk level. The SHAP analysis model is used to calculate the SHAP value of each feature data in the global feature vector for the prediction result; the corresponding key factors are extracted by sorting them according to the absolute value of the SHAP value, and a weighted list of key factors is generated.

10. The deep learning-based anesthetic complication prediction and avoidance auxiliary decision-making system according to claim 9, characterized in that, Based on the risk level and key factors, a personalized avoidance strategy is generated. Then, based on this personalized avoidance strategy, personalized avoidance suggestions and visualizations are obtained. The process includes: Construct a knowledge base for complication avoidance strategies; based on risk levels and key factors, and according to the knowledge base for complication avoidance strategies, generate basic strategies; the basic strategies include reinforcement intervention strategies, preventive intervention strategies, and monitoring strategies; The basic strategy is dynamically optimized by reinforcement learning model to generate personalized avoidance strategy, including: constructing state space, defining action space, designing reward function and iterative optimization; the effect of intervention measures is pre-simulated in a simulation environment by deep deterministic policy gradient algorithm, the policy parameters are adjusted according to reward function, and finally personalized avoidance strategy adapted to individual patient characteristics is output. Transform personalized avoidance strategies into visualized personalized avoidance suggestions and visualizations, including: visualizations and personalized avoidance suggestions.

Images