Disclosure of Invention
The invention discloses a deep learning-based anesthesia complication prediction model construction method which is used for improving the accuracy and efficiency of anesthesia complication prediction.
The invention provides a deep learning-based anesthesia complication prediction model construction method, which comprises the following steps:
acquiring a historical multi-dimensional dataset, preprocessing the historical multi-dimensional dataset and labeling the historical multi-dimensional dataset, wherein,
The historical multidimensional dataset comprises a preoperative dataset, an intraoperative dataset and a postoperative dataset;
Extracting key features capable of representing anesthesia complications from the preoperative data set, the postoperative data set and the postoperative data set respectively;
Constructing a time sequence prediction network, training the time sequence prediction network by using the extracted key features to obtain an anesthesia complication prediction model, and testing and verifying the anesthesia complication prediction model after training;
and (5) inputting preoperative data, intraoperative data or postoperative data into a trained anesthesia complication prediction model to predict anesthesia complications in different stages.
In one implementation manner, the step of acquiring the historical multidimensional dataset and preprocessing the historical multidimensional dataset specifically includes:
acquiring preoperative data, intraoperative data and postoperative data of a plurality of patients;
The method comprises the steps of associating and merging preoperative data, intraoperative data and postoperative data related to a patient by using a patient identifier to form multidimensional data containing all clinical data of the patient, wherein the data in the multidimensional data are arranged in time sequence;
Carrying out missing value processing, abnormal value detection and processing and formatting processing on the historical multidimensional data set;
dividing the processed historical multidimensional dataset according to a time sequence, and respectively defining time windows before, during and after operation;
the data quality of the historical multi-dimensional dataset is checked periodically using a data monitoring mechanism.
In one embodiment, the step of extracting key features capable of characterizing the anesthetic complications from the preoperative dataset, the intraoperative dataset, and the postoperative dataset, respectively, comprises:
setting an initial time window according to an operation flow and an operation time node, identifying key events influencing the risk of anesthesia complications for preoperation data, intraoperative data and postoperative data, dynamically adjusting the size and the position of the time window according to the identified key events, and extracting key features by utilizing the self-adaptive adjusted time window.
In one embodiment, in the step of extracting key features extracted by using the adaptively adjusted time window, features of each stage are extracted by using correlation analysis;
The extracted preoperative features comprise features reflecting baseline health and pre-anesthesia risk factors of a patient in a preoperative time window, the extracted intraoperative features comprise features covering surgical processes and anesthesia management in the intraoperative time window, and the extracted postoperative features comprise features for restoring dynamic and complication risks in the postoperative time window.
In one embodiment, the pre-operative features include physiological and biochemical indicators, cardiopulmonary function indicators, past medical history, drug history, physical features, lifestyle and psychological state features, the intra-operative features include narcotics, vasoactive drugs, surgical procedures, real-time vital signs, surgical complications, and the post-operative features include recovery, signs of complications, drug responses, trends in physiological indicators, readmission.
In one embodiment, the time-series predictive network structure includes three predictive network layers, a fusion layer, and an output layer, each predictive network layer including an LSTM layer, a convolution attention, and at least one hidden layer.
In one embodiment, the anesthetic complication prediction model adopts a layered training method, and the step of training the time sequence prediction network by using the key features to be extracted to obtain the anesthetic complication prediction model includes:
determining the sequence length of the preoperative, intraoperative and postoperative phases;
the key features of each stage are equally divided into a training set, a verification set and a test set by utilizing a rolling window;
Training the time sequence prediction network by using training set data, inputting time sequence data of a corresponding stage in each stage, and adjusting weight according to the output of the model;
And feeding back the dynamic adjustment sequence length according to the anesthesia complication prediction model.
In one embodiment, after the step of inputting pre-operative data, intra-operative data, or post-operative data into the trained prediction model of anesthetic complications, performing prediction of anesthetic complications at different stages, the method further comprises:
the method comprises the steps of obtaining the actual complication occurrence condition and treatment effect of a patient as feedback information, and utilizing the feedback information to adjust and optimize the prediction parameters of a model and the generation logic of personalized advice to form a continuously improved closed loop.
In one embodiment, after the step of inputting pre-operative data, intra-operative data, or post-operative data into the trained prediction model of anesthetic complications, performing prediction of anesthetic complications at different stages, the method further comprises:
and converting the output result of the anesthesia complication prediction model into an understandable risk score, and generating personalized advice information according to the risk score of each stage and preset advice generation rules.
In one embodiment, the recommendation generation rules include threshold determinations, risk ranking, and recommendation content matching.
The method for constructing the anesthesia complication prediction model based on the deep learning technology has the advantages that the method is high in accuracy and practicality, can effectively predict possible anesthesia complications of a patient in an operation process, provides important auxiliary tools for a clinical anesthesiologist, reduces medical risks, and improves treatment effects.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims thereof as well as the appended drawings.
The technical scheme of the invention is further described in detail through the drawings and the embodiments.
Detailed Description
The preferred embodiments of the present invention will be described below with reference to the accompanying drawings, it being understood that the preferred embodiments described herein are for illustration and explanation of the present invention only, and are not intended to limit the present invention.
During medical procedures, many procedures require anesthesia procedures to be performed on the patient, which, while capable of helping the physician perform the procedure, help the patient lose consciousness, present some risk. General anesthesia is divided into three stages of anesthesia induction, anesthesia maintenance and anesthesia recovery, and generally, the blood pressure, heart rate and bis value decrease immediately after anesthesia induction is mainly caused by the inhibition of circulation by anesthetic, blood pressure, pulse and heart rate are maintained in proper ranges during anesthesia, namely, hemodynamic stability is maintained, and if severe hemodynamic fluctuation occurs or the blood pressure and heart rate are difficult to maintain, vasoactive drugs are needed to maintain blood pressure. The Bis value reflects the degree of sedation and previous studies have suggested that low mean arterial pressure, low Bis, low Mac (minimum alveolar gas effective concentration) are closely related to increased post-operative mortality. As shown in fig. 4. In the general anesthesia process, the physical indexes of a patient need to be detected, wherein the general detection indexes in the general anesthesia process include blood pressure, pulse, heart rate, oxygen saturation, end-tidal carbon dioxide and brain electrical double frequency index (bis), and critical patients periodically check blood gas analysis, wherein the blood gas analysis comprises arterial blood oxygen partial pressure, carbon dioxide partial pressure, blood ph value, electrolyte levels of lactic acid, potassium sodium calcium and the like. And judging whether anesthesia complications exist or not according to the indexes. The anesthesia related complications mainly comprise cardiovascular and cerebrovascular accidents, malignant hyperthermia, agitation in a waking period, waking delay, nausea and vomiting after operation, pulmonary complications and the like, are generally not much related to the resected part of a patient, and can mainly treat the postoperative pulmonary complications (pleural effusion, pulmonary atelectasis, pneumothorax and the like) through imaging reaction. The schematic diagram of the lung after anesthesia is shown in fig. 5, the visible light transmittance of the two lungs is reduced in 3 hours after operation, the flaky ground glass density shadow is visible, the heart shadow is not large, the chest film shows the diffuse patch-shaped high density shadow of the two lungs in 19 hours after operation, the two lung doors are not large, and the chest film has no obvious abnormality of the two lungs, the heart and the diaphragm in 64 hours after operation. In the anesthesia process, if there is no abnormality, as shown in fig. 6, the CT images of the right diaphragmatic top 1cm layer are respectively before anesthesia, after anesthesia induction intubation and after operation, and the position indicated by the arrow is a lung non-stretch area, which indicates that the lung of the patient is normal.
The embodiment 1 provides a deep learning-based anesthesia complication prediction model construction method, which comprises the following steps:
acquiring a historical multi-dimensional dataset, preprocessing the historical multi-dimensional dataset and labeling the historical multi-dimensional dataset, wherein,
The historical multidimensional dataset comprises a preoperative dataset, an intraoperative dataset and a postoperative dataset;
Extracting key features capable of representing anesthesia complications from the preoperative data set, the postoperative data set and the postoperative data set respectively;
Constructing a time sequence prediction network, training the time sequence prediction network by using the extracted key features to obtain an anesthesia complication prediction model, and testing and verifying the anesthesia complication prediction model after training;
and (5) inputting preoperative data, intraoperative data or postoperative data into a trained anesthesia complication prediction model to predict anesthesia complications in different stages.
Referring to fig. 1, the application firstly acquires a historical multidimensional dataset containing information of multiple aspects such as medical history, physiological indexes, medicine use condition and the like of a patient. The historical multidimensional dataset may be derived from a hospital database, clinical trial record, or the like. And preprocessing and labeling the historical multidimensional dataset. Preprocessing can improve the quality of subsequent analysis and modeling. And during labeling, a manual labeling or machine learning automatic labeling method can be adopted to label the type of anesthesia complications for each sample in the data set. Then, key features are extracted from the historical multidimensional data set, and key features related to anesthesia complications can be extracted from the marked historical multidimensional data set through technologies such as cluster analysis, association rule mining and the like. And then, constructing an anesthesia complication prediction model by utilizing the extracted key features and combining a time sequence prediction network. And finally, inputting the key features with the labels into the constructed time sequence prediction model, training the model, and testing and verifying the training-completed anesthesia complication prediction model. And (3) inputting preoperative data, intraoperative data or postoperative data into a trained anesthesia complication prediction model, so that the anesthesia complications in different stages can be predicted.
The method for constructing the anesthesia complication prediction model has the beneficial effects that the anesthesia complication prediction model can effectively predict the risk probability of the anesthesia complication of a patient in the operation process, provides a powerful supporting tool for a clinical anesthesiologist, is beneficial to reducing medical risks and improving the treatment effect.
Based on embodiment 1, the step of obtaining the historical multidimensional dataset and preprocessing the historical multidimensional dataset specifically includes:
acquiring preoperative data, intraoperative data and postoperative data of a plurality of patients;
The method comprises the steps of associating and merging preoperative data, intraoperative data and postoperative data related to a patient by using a patient identifier to form multidimensional data containing all clinical data of the patient, wherein the data in the multidimensional data are arranged in time sequence;
Carrying out missing value processing, abnormal value detection and processing and formatting processing on the historical multidimensional data set;
Dividing the processed historical multidimensional data set according to time sequence, and respectively defining time windows before, during and after operation to obtain data sets applicable to different analysis stages;
the data quality of the historical multi-dimensional dataset is checked periodically using a data monitoring mechanism.
The technical scheme has the working principle and beneficial effects that preoperative data, intraoperative data and postoperative data of a plurality of patients are firstly obtained, then the preoperative data, intraoperative data and postoperative data related to the patients are associated and combined by using the patient identifiers, and the association between the data is established by the patient identifiers. Next, missing value processing, outlier detection and processing, and formatting processing are performed on the historical multidimensional dataset. Through the processing steps, the quality and consistency of the data are ensured. For the missing value, strategies such as filling, deleting or replacing can be selected, and for the abnormal value, detection and processing are needed, so that the influence on the model is avoided. Then, the processed historical multidimensional data sets are divided according to time sequence, and time windows before, during and after operation are defined respectively to obtain training sets, verification sets and test sets which are applicable to different analysis stages, so that a definite time flow exists for training, verifying and testing of the model, and the effectiveness and reliability of the result are guaranteed. Finally, the data quality of the historical multidimensional dataset is checked regularly by using a data monitoring mechanism. The data monitoring mechanism can detect the new data immediately when the new data is added, so that the problems can be found and solved in time.
On the basis of embodiment 1, the step of extracting key features capable of characterizing the anesthetic complications from the preoperative dataset, the intraoperative dataset and the postoperative dataset, respectively, comprises:
setting an initial time window according to the operation flow and the operation time node;
Identifying key events affecting the risk of anesthesia complications for preoperative data, intraoperative data and postoperative data, and dynamically adjusting the size and the position of a time window according to the identified key events;
and extracting key features by using the self-adaptive adjusted time window.
The working principle of the technical scheme is that referring to fig. 2, the application firstly sets an initial time window according to the operation flow and operation time nodes. When setting the initial window, starting from the starting time of the operation and ending at the expected ending time of the operation, thereby obtaining an initial time window. Critical events affecting the risk of anesthetic complications, such as certain physiological changes, use of drugs, manner of surgery, etc., are then identified for the pre-operative data, the intra-operative data, and the post-operative data. And dynamically adjusting the size and the position of the time window according to the identified key event. The time window size and position are adjusted based on the recognition result of the key event to better capture possible variations. And finally, extracting key features capable of best representing the risk of the anesthesia complications by using the self-adaptive adjusted time window so as to improve the effect of a final prediction model.
On the basis of the embodiment 1, in the step of extracting key features by using an adaptive adjusted time window, the features of each stage are extracted by using correlation analysis, the extracted preoperative features comprise features reflecting baseline health and risk factors before anesthesia of a patient in a preoperative time window, the extracted intraoperative features comprise features covering the surgical process and anesthesia management in the intraoperative time window, and the extracted postoperative features comprise features of response recovery dynamics and complications risks in the postoperative time window.
The postoperative features comprise physiological and biochemical indexes, heart and lung function indexes, past medical history, medicine history, physical features, living habit and psychological state features, wherein the intraoperative features comprise anesthetic medicines, vasoactive medicines, operation operations, real-time vital signs and operation complications, and the postoperative features comprise recovery conditions, complication signs, medicine reactions, physiological index trend and readmission conditions.
The technical scheme has the working principle and beneficial effects that in the process of extracting the characteristics by utilizing the self-adaptive adjusted time window, corresponding characteristics are extracted in a targeted manner according to the different characteristics of three stages of preoperative, intraoperative and postoperation. The pre-operative features include features reflecting baseline health and pre-anesthesia risk factors of the patient within a pre-operative time window, the intra-operative features include features covering surgical procedures and anesthesia management within the intra-operative time window, and the extracted post-operative features include features recovering dynamic and complication risk of the reaction within the post-operative time window.
In particular, preoperative features include physiological and biochemical indicators such as heart rate, blood pressure, respiratory rate, etc., which directly affect the patient's anesthetic process. The past medical history is that the health condition and the past disease condition of the patient are known so as to be convenient for predicting possible complications. Drug history-the use of drugs may directly affect the patient's anesthetic process, and is therefore an important feature. Lifestyle, such as whether smoking, drinking, or having a history of allergies, all of which affect the anesthetic process. Psychological state characteristics the psychological state of the patient such as anxiety, depression and the like can influence the anesthesia process. Physical characteristics and cardiopulmonary function index. Intraoperative characteristics include anesthetic drug, the type of anesthetic drug used, the dosage, etc., which are factors that directly affect the anesthetic process. Vasoactive drugs, surgical procedures, the specific procedure, mode of surgery, etc., which are all important factors in the surgical procedure. Real-time vital signs such as oxygen saturation, electrocardiogram, blood oxygen, etc., which are vital signs that must be closely monitored during surgery. Surgical complications-complications during surgery conditions, which are important for predicting postoperative complications. The postoperative features comprise recovery conditions such as body temperature, heart rate and blood pressure of the patient, and the recovery conditions are all indexes which need to be closely observed after the operation. Signs of complications such as nausea, vomiting, pain, etc., which may be precursors to complications. Drug response-the response of the patient to the anesthetic drug after surgery, which is important for adjusting the medication regimen. The trend of the physiological index is the trend of the change of the physiological index such as blood oxygen, heart rate, blood pressure and the like. Readmission, the condition of patient readmission after surgery, which can reflect the duration and effectiveness of treatment.
On the basis of embodiment 1, the time-series prediction network structure comprises three prediction network layers, a fusion layer and an output layer, wherein each prediction network layer comprises an LSTM layer, a convolution attention and at least one hidden layer.
The technical scheme adopts the working principle that three independent LSTMs (long-short-term memory network) are adopted, and data of each stage is firstly processed through the independent LSTMs so as to capture time sequence characteristics. Three attention mechanism layers, followed by a separate LSTM, weight key points in time for each phase. The hidden layer may be located after the attention mechanism layer, and may be one or more. These layers receive weighted outputs from the attention mechanism layer and perform further feature extraction and integration. The hidden layer may be a fully connected layer, a convolutional layer, or other type of neural network layer, depending on the model design and task requirements. Fusion layer for the intraoperative and postoperative phases, the fusion layer combines the LSTM layer output (weighted by attention) of the previous phase with the output of the current phase to realize information fusion. The fusion layer may also be considered part of the hidden layer, especially when designing a multi-level fusion structure. Finally, the output of the fusion layer (or the last layer of the hidden layer) enters the output layer, which generates the final prediction result.
The technical scheme has the beneficial effects that the time sequence prediction network structure not only considers the characteristics of time sequence data, but also fully fuses information of different modes, has stronger expression capability, and can be used for predicting various time sequence data. Meanwhile, the model has good expandability, and can be applied to prediction of various anesthesia complications.
Based on embodiment 1, the anesthesia complication prediction model adopts a layered training method, and the step of training the time sequence prediction network by using the key features to be extracted to obtain the anesthesia complication prediction model comprises the following steps:
determining the sequence length of the preoperative, intraoperative and postoperative phases;
the key features of each stage are equally divided into a training set, a verification set and a test set by utilizing a rolling window;
Training the time sequence prediction network by using training set data, inputting time sequence data of a corresponding stage in each stage, and adjusting weight according to the output of the model;
And feeding back the dynamic adjustment sequence length according to the anesthesia complication prediction model.
The working principle of the above technical solution is that referring to fig. 3, firstly, the sequence lengths of the preoperative, intraoperative and postoperative phases are determined first to ensure that the prediction model can adapt to the characteristics of different phases well. In determining sequence strength, it may be desirable to increase the sequence length of the post-operative stage, as the data of the post-operative stage tends to be more abundant than the pre-operative and intra-operative stages, more conducive to training of the model. Then, the key feature data of each stage is equally divided into a training set, a verification set and a test set by using a rolling window. The training set is used for training the model, the verification set is used for checking the performance of the model, and the test set is used for evaluating the generalization capability of the model on unknown data. The training set data is then used to train the time series prediction network. This process is typically performed in steps, with time series data of the corresponding phase being input first, and the prediction result being output. And then, according to the output result of the model, adjusting the weight of the model to improve the prediction performance of the model. And finally, dynamically adjusting the sequence length according to feedback of the anesthesia complication prediction model.
The technical scheme has the beneficial effects that the training process of the anesthesia complication prediction model is beneficial to improving the prediction accuracy, generalization capability and clinical practicality of the model, and powerful technical support is provided for anesthesia management.
On the basis of embodiment 1, after the step of inputting preoperative data, intraoperative data or postoperative data into the trained anesthesia complication prediction model to perform prediction of the anesthesia complications of different phases, the method further includes:
the method comprises the steps of obtaining the actual complication occurrence condition and treatment effect of a patient as feedback information, and utilizing the feedback information to adjust and optimize the prediction parameters of a model and the generation logic of personalized advice to form a continuously improved closed loop.
The technical scheme has the working principle and beneficial effects that firstly, the actual complication occurrence condition and treatment effect of a patient are obtained and used as feedback information. Which may be obtained from an electronic medical record system of a medical institution, medical records, or clinical surveys. The feedback information obtained may be used to determine the accuracy of the predictive model, as well as the validity of the personalized advice. Second, feedback information is utilized to adjust and optimize the prediction parameters of the model and the generation logic of the personalized advice. In particular, the predictive performance of the model may be improved by changing the architecture of the model, adjusting hyper-parameters, adding new features, and the like. At the same time, the generation logic of the provided personalized advice can be adjusted according to the feedback information, for example, more specific drug advice or surgical plan recommendation can be provided according to the actual illness state and medication condition of the patient. Finally, this process needs to be iterated continually, allowing the model to learn and optimize continually. Each feedback can help find the model's shortcomings, thereby improving the model to make it more accurate and reliable.
On the basis of embodiment 1, after the step of inputting preoperative data, intraoperative data or postoperative data into the trained anesthesia complication prediction model to perform prediction of the anesthesia complications of different phases, the method further includes:
and converting the output result of the anesthesia complication prediction model into an understandable risk score, and generating personalized advice information according to the risk score of each stage and preset advice generation rules.
The suggestion generation rule comprises threshold judgment, risk classification and suggestion content matching.
The working principle of the technical scheme is that the prediction result of the model is converted into a risk score, and the risk score reflects the possibility of the anesthesia complication of the patient at each stage. The higher this score, the greater the risk of the patient developing anesthesia complications at this stage. Conversely, a lower score indicates a lower risk of patient developing anesthesia complications. Personalized advice information may then be generated based on the risk scores for each stage and preset advice generation rules. For example, if a patient is at some point post-operatively, the predictive outcome of the model indicates that the patient is at a high risk of developing anesthesia complications within a few days of the future, then additional advice may be provided to the physician or nurse, such as avoiding certain actions or practices that may lead to complications, or taking special care to reduce the incidence of complications.
The technical scheme has the beneficial effects that doctors and nurses can make optimal decisions according to the prediction result and the suggestion information of the model at different stages of operation so as to reduce medical risks.
It will be apparent to those skilled in the art that various modifications and variations can be made to the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.