An interpretable method for assessing the risk of hypoxia during painless endoscopy and a related device

Abstract

The present invention discloses an interpretable painless endoscopic hypoxia risk assessment method and related devices, belonging to the field of machine learning technology. The method comprises: obtaining clinical characteristics of a patient to be subjected to painless digestive endoscopy; inputting the clinical characteristics into a pre-trained hypoxia risk assessment model to obtain a hypoxia risk assessment result; wherein the method for obtaining the hypoxia risk assessment model comprises: obtaining an importance ranking result of each clinical characteristic in the historical sample data set; obtaining an interpretable model according to the importance ranking result; constructing a negative sample data set based on the historical sample data set, and training the improved interpretable model through the historical sample data set and the negative sample data set to obtain a hypoxia risk assessment model, thereby improving the problem of difficulty in learning corresponding rules during model training and overfitting due to small sample data, and realizing accurate and efficient interpretable prediction of hypoxia event risk.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to an explainable painless endoscopic hypoxia risk assessment method and related devices. Background Art

[0002] With the development of medical technology, painless gastroenteroscopy is widely used in the diagnosis of gastrointestinal diseases. However, due to the use of sedatives during painless digestive endoscopy, some patients may experience hypoxia events. If this hypoxia event is not discovered and treated in time, it may lead to serious consequences, including respiratory depression, cardiovascular complications, and even life-threatening. Traditional monitoring methods rely on continuous observation of the patient's vital signs during the examination, but this may have the risk of delayed response, especially for certain high-risk patients, where early warning means are limited.

[0003] At present, the clinical risk assessment of hypoxic events during painless digestive endoscopy mainly relies on the clinical experience of anesthesiologists. However, this risk assessment method is somewhat subjective and it is difficult to accurately predict the occurrence of hypoxic events, especially in large-scale outpatient settings. Traditional manual assessment methods are difficult to efficiently and accurately identify high-risk patients. Therefore, there is an urgent need for a systematic artificial intelligence model dedicated to the prediction of hypoxic risks during painless digestive endoscopy, to help doctors effectively predict and prevent the occurrence of hypoxic events before painless gastrointestinal endoscopy, and to ensure patient safety. Summary of the invention

[0004] The present invention provides an interpretable painless endoscopic hypoxia risk assessment method and related devices. By performing feature selection and importance sorting in advance, an interpretable model that integrates hypoxia risk and feature correlation rules is constructed based on the sorting results, and negative sample data is constructed based on the original sample data, thereby achieving efficient and accurate interpretable prediction of the patient's hypoxia event risk based on small sample data.

[0005] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0006] In a first aspect, the present invention provides an interpretable method for assessing the risk of hypoxia during painless endoscopic examination, comprising:

[0007] Acquiring clinical characteristics of patients who are to undergo painless digestive endoscopy, wherein the clinical characteristics include individual characteristics, medical history information, and relevant pre-endoscopy assessment parameters;

[0008] Inputting the clinical characteristics into a pre-trained hypoxia risk assessment model to obtain a hypoxia risk assessment result;

[0009] Wherein, the method for obtaining the hypoxia risk assessment model includes:

[0010] Obtain a historical sample data set of patients who underwent painless digestive endoscopy, and construct a negative sample data set based on the historical sample data set, wherein each sample data set includes clinical characteristics and risk category labels;

[0011] The importance of each clinical feature in the historical sample data set is ranked;

[0012] An improved interpretable model is obtained based on the importance ranking result, wherein the improved interpretable model includes an encoder and a decoder connected to each other, the encoder algorithm incorporates the learning of correlation rules between clinical features and hypoxia risk, and is used to encode input data to obtain a feature vector containing the correlation rule, and the decoder is used to perform risk category label mapping according to the feature vector;

[0013] The improved interpretable model is trained by using the historical sample data set and the negative sample data set to obtain a hypoxia risk assessment model.

[0014] Optionally, the individual characteristics include demographic characteristics and basic physiological indicators;

[0015] The demographic characteristics include: age, sex, height, weight, BMI;

[0016] The basic physiological indicators include: basic blood pressure, basic heart rate, basic finger pulse oxygen saturation;

[0017] The medical history information includes: history of basic hypertension, history of diabetes, history of smoking, history of alcoholism, and snoring;

[0018] The relevant evaluation parameters before the mirror include: neck circumference, Mahalanobis grade, and ASA grade.

[0019] Optionally, the step of obtaining a historical sample data set of patients who underwent painless digestive endoscopy includes:

[0020] The historical clinical characteristics of patients admitted to the hospital for painless gastrointestinal endoscopy were collected;

[0021] Data cleaning and standardization of each patient's historical clinical characteristics;

[0022] The processed historical clinical characteristics of each patient are combined with the corresponding hypoxia risk labels to obtain a historical sample data set.

[0023] Optionally, the clinical features in the historical sample dataset are ranked by importance using a pre-trained XGBoost model;

[0024] Wherein, the training method of the XGBoost model includes:

[0025] Segmenting the historical sample data set to obtain corresponding training sets and test sets;

[0026] Perform hyperparameter adjustment and optimization on the XGBoost model on the training set and test set to obtain the optimal parameter combination;

[0027] The XGBoost model is set according to the optimal parameter combination to obtain a trained XGBoost model.

[0028] Optionally, the hyperparameter adjustment and optimization of the XGBoost model on the training set and the test set are performed to obtain the optimal parameter combination, including:

[0029] Input the training set into the XGBoost model, and preliminarily set the XGBoost model according to conventional parameters and the specified number of classification categories;

[0030] Using a grid search algorithm to perform hyperparameter adjustment on the initially set XGBoost model on the training set to obtain different parameter combinations of the XGBoost model;

[0031] The XGBoost models with different parameter combinations are input into the test set for verification to obtain the optimal parameter combination of the XGBoost model.

[0032] Optionally, constructing a negative sample dataset based on the historical sample dataset includes:

[0033] Clinical features that were positively correlated with the risk of hypoxic events were selected as target features;

[0034] Searching for sample data containing the target feature in the historical sample data set;

[0035] The target feature value is increased in the found sample data to obtain new sample data.

[0036] Optionally, the encoder includes a plurality of feature extraction modules connected in sequence and a rule layer connected to the last feature extraction module, wherein each feature extraction module includes a feature concatenation layer, the input of the feature concatenation layer in the nth feature extraction module includes a feature vector output by the n-1th feature extraction module and a feature vector of the clinical feature at the nth position according to the result of sorting in reverse order of importance, and the rule layer stores the encoding value of the target feature, which is used to perform a dot product between the feature vector output by the last feature extraction module and the encoding value of the target feature;

[0037] The decoder includes a softmax layer.

[0038] Optionally, the step of training the improved interpretable model by using the historical sample data set and the negative sample data set to obtain a hypoxia risk assessment model includes:

[0039] Input each sample data in the historical sample data set and the negative sample data set into the improved interpretable model, extract features from the feature vectors in each sample data through an encoder, and map risk category labels according to the feature vectors output by the encoder through a decoder;

[0040] The loss function value of the improved interpretable model is calculated based on the risk category label output by the decoder and the true label corresponding to each sample data;

[0041] The model parameters were adjusted to minimize the loss function value to obtain the hypoxia risk assessment model.

[0042] In a second aspect, the present invention provides an interpretable painless endoscopic hypoxia risk assessment device, comprising:

[0043] Patient information acquisition module: used to obtain clinical characteristics of patients who are to undergo painless digestive endoscopy, wherein the clinical characteristics include individual characteristics, medical history information, and relevant evaluation parameters before endoscopy;

[0044] A hypoxia risk assessment result acquisition module is used to input the clinical characteristics into a pre-built hypoxia risk assessment model to obtain a hypoxia risk assessment result;

[0045] Wherein, the method for obtaining the hypoxia risk assessment model includes:

[0046] Obtain a historical sample data set of patients who underwent painless digestive endoscopy, and construct a negative sample data set based on the historical sample data set, wherein each sample data set includes clinical characteristics and risk category labels;

[0047] The importance of each clinical feature in the historical sample data set is ranked;

[0048] An improved interpretable model is obtained based on the importance ranking result, wherein the improved interpretable model includes an encoder and a decoder connected to each other, the encoder algorithm incorporates the learning of correlation rules between clinical features and hypoxia risk, and is used to encode input data to obtain a feature vector containing the correlation rule, and the decoder is used to perform risk category label mapping according to the feature vector;

[0049] The improved interpretable model is trained by using the historical sample data set and the negative sample data set to obtain a hypoxia risk assessment model.

[0050] In a third aspect, the present invention provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the interpretable painless endoscopic hypoxia risk assessment method as described in any step of the first aspect.

[0051] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: the present invention applies an interpretable deep learning network to relevant scenarios in the medical field, and combines it with the known relevant rules in the medical scenario to construct an interpretable model with better learning effect, thereby solving the problem that it is difficult to learn the corresponding rules during model training, and constructing negative sample data based on the original small-range sample data, so as to train the constructed interpretable model to obtain a hypoxic event risk assessment model, thereby improving the model overfitting problem caused by the small number of sample data in the medical field, collecting multiple basic characteristics of the patient before the examination, and using the risk assessment model to obtain an interpretable risk prediction result based on these basic characteristics, thereby achieving accurate and efficient prediction of the risk of hypoxic events, and predicting the risk in advance through the model can provide an early warning before the patient's hypoxic event occurs, and take timely intervention measures, which significantly improves the safety of the painless digestive endoscopy process, reduces anesthesia-related complications, and helps to optimize the use of medical resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Shown is a flow chart of a method for assessing hypoxia risk during painless endoscopy that can be explained in one embodiment of the present invention;

[0053] Figure 2 Shown is a schematic diagram of the structure of a feature extraction module in an interpretable deep learning network in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0055] Example 1

[0056] like Figure 1 As shown, this embodiment provides an interpretable method for assessing hypoxia risk during painless endoscopy, comprising:

[0057] Acquiring clinical characteristics of patients who are to undergo painless digestive endoscopy, wherein the clinical characteristics include individual characteristics, medical history information, and relevant pre-endoscopy assessment parameters;

[0058] Inputting the clinical characteristics into a pre-trained hypoxia risk assessment model to obtain a hypoxia risk assessment result;

[0059] Wherein, the method for obtaining the hypoxia risk assessment model includes:

[0060] Obtain a historical sample data set of patients who underwent painless digestive endoscopy, and construct a negative sample data set based on the historical sample data set, wherein each sample data set includes clinical characteristics and risk category labels;

[0061] The importance of each clinical feature in the historical sample data set is ranked;

[0062] An improved interpretable model is obtained based on the importance ranking result, wherein the improved interpretable model includes an encoder and a decoder connected to each other, the encoder algorithm incorporates the learning of correlation rules between clinical features and hypoxia risk, and is used to encode input data to obtain a feature vector containing the correlation rule, and the decoder is used to perform risk category label mapping according to the feature vector;

[0063] The improved interpretable model is trained by using the historical sample data set and the negative sample data set to obtain a hypoxia risk assessment model.

[0064] By combining the feature correlation rules in the hypoxic event risk assessment scenario with the interpretable model, an improved interpretable model is obtained, which enables the interpretable model to better learn the correct rules. By constructing negative samples and training the improved interpretable model together with real samples, the obtained hypoxic risk assessment model is more suitable for medical scenarios with small samples, and the generalization ability of the model is improved, thereby achieving accurate and reliable prediction of hypoxic event risks through an interpretable artificial intelligence model.

[0065] Example 2

[0066] Based on Example 1, this example also makes the following design.

[0067] Combining artificial intelligence with the medical field presents the following problems in the collection of sample data.

[0068] Small sample problem: Sample data collection is difficult and costly, resulting in very limited data available for training.

[0069] Inaccurate data collection: Due to the complexity of medical data and the limitations of collection methods, the collected data is often not completely accurate or noisy.

[0070] Fewer and fixed features: Compared to the millions or even tens of millions of features in large deep learning models, our models have fewer features and these features are usually fixed.

[0071] These problems may cause the following situations during model training:

[0072] Model uninterpretability: Model output is difficult for medical professionals to understand and lacks interpretability in a medical context.

[0073] Overfitting: The model overfits to small sample data and has difficulty generalizing to new data.

[0074] Does not conform to real physical or medical rules: For example, although in medicine a larger feature value means a higher probability of the patient getting the disease, it may be difficult for the model to learn this relationship.

[0075] Due to the above technical background, this embodiment constructs negative samples and an improved interpretable deep learning network to ensure that the model can follow the known rules in the medical field in the case of small samples and limited features, thereby improving the generalization ability and interpretability of the model. The interpretable deep learning network trained with positive and negative samples is used as a hypoxia event risk assessment model to assess the risk of hypoxia that may occur in patients during painless gastrointestinal endoscopy based on patient information, wherein the acquisition of the hypoxia event risk assessment model is divided into the following steps.

[0076] Step 1: Data Collection

[0077] In the outpatient clinic, basic data were collected from 2,237 patients who underwent painless gastrointestinal endoscopy. The collected patient information included but was not limited to the following parameters:

[0078] Demographic characteristics: age, sex, height, weight, BMI;

[0079] Basic physiological indicators: basic blood pressure, basic heart rate, basic pulse oxygen saturation;

[0080] Medical history information: history of basic hypertension, diabetes, smoking, alcoholism, and snoring;

[0081] Other information: neck circumference, Marsh grade, ASA grade, etc.

[0082] Step 2: Data processing

[0083] The collected patient information is cleaned and standardized to ensure the availability and consistency of various indicators. Missing data is processed by interpolation or deletion to obtain a standardized historical sample data set.

[0084] Step 3: Get the XGBoost model

[0085] XGBoost is based on the GradientBoosting method and uses multiple weak learners (usually decision trees) to build a model. Decision trees are non-parametric models that do not rely on assumptions about data distribution, so they can still adapt flexibly in small sample situations. In addition, XGBoost introduces regularization terms to prevent overfitting problems with small sample data.

[0086] The objective function of XGBoost consists of a loss function and a regularization term:

[0087] ,

[0088] in, Indicates The objective function at the iteration is is a differentiable loss function, It is in The predicted value at the iteration, It is in The number of tree pairs added in the iteration The predicted value of It is in The true value at the iteration, is the regularization term, which means t The new addition in the iteration t Tree The complexity of is added to the objective function as a penalty term to avoid overfitting of the model.

[0089] Regularization term The form is as follows:

[0090] ,

[0091] in, It is t Tree The number of leaf nodes, It is j The weight of a leaf node, j For the t Tree The index of the upper leaf node, and It is a regularization parameter used to control the complexity of the model. In the case of small samples, the existence of the regularization term can effectively prevent the model from overfitting and help improve the generalization ability of the model.

[0092] When the feature dimension is small, the advantage of XGBoost is that it can implicitly perform feature selection. The decision tree splitting criterion is based on the optimization of the gain function, and only the most informative features are selected for splitting.

[0093] XGBoost Gain Function Used to evaluate whether to split, the formula is as follows:

[0094] ,

[0095] in, , are the gradient of the left child node and the first-order derivative of the loss function of the right child node, respectively. , are the second-order derivatives of the left and right child nodes, respectively. is the regularization parameter, controlling the sensitivity of the split, is the penalty term for leaf nodes.

[0096] Through the gain function, XGBoost only selects those features with large gains for splitting. This is especially effective when the feature dimensions are small because it can automatically select the most informative features.

[0097] XGBoost has a built-in mechanism to handle missing features. During each tree building process, XGBoost assigns a default direction (left or right) to the missing values ​​at each node split and optimizes the choice of this direction by minimizing the loss function. This ensures that missing features do not have a significant negative impact on the model.

[0098] Specifically, when dividing nodes, XGBoost will assign missing values ​​to sub-nodes that can maximize gains based on the results of the gain function, which enables the model to maintain good predictive capabilities when dealing with missing features.

[0099] XGBoost uses the gradient boosting framework to efficiently optimize the objective function by approximating the second-order Taylor expansion, that is, using the first-order and second-order derivatives of the loss function to accelerate iterations. Even if the data is sparse or the sample size is small, this optimization method can still converge efficiently.

[0100] The second-order Taylor expansion approximation formula of the objective function is as follows:

[0101] ,

[0102] in, is the first-order gradient of the loss function with respect to the predicted value, is the second-order gradient of the loss function with respect to the predicted value.

[0103] This second-order approximation method speeds up the optimization process of gradient boosting, so that even in the case of small samples or sparse data, XGBoost can still converge quickly and show good performance.

[0104] The sample data processed in step 2 is input into the constructed XGBoost to optimize the model parameters.

[0105] In terms of hyperparameter optimization, XGBoost has relatively few hyperparameters, mainly the maximum depth of the tree, the minimum weight of the leaf node, etc., so we use the grid search method. Grid search is a hyperparameter tuning method that exhaustively enumerates the specified parameter combinations, calculates the performance of each set of parameters on the test set, and finally selects the best performing parameter combination.

[0106] In the process of model optimization, the performance of the model is usually evaluated not only by accuracy, but also by sensitivity (i.e. True Positive Rate) and specificity (i.e. True Negative Rate) for comprehensive judgment. At this time, ROC curve and AUC value are very commonly used evaluation indicators.

[0107] The ROC curve is a curve drawn with the false positive rate FPR as the horizontal axis and the true positive rate TPR as the vertical axis, which is used to evaluate the performance of the model under different thresholds. The closer the ROC curve is to the upper left corner, the better the classification performance of the model.

[0108] The calculation formula of true positive rate TPR is:

[0109] ,

[0110] The calculation formula of false positive rate FPR is:

[0111] ,

[0112] Among them, TP is the number of true positive examples, FN is the number of false negative examples, FP is the number of false positive examples, TN is the number of true negative examples, and the AUC value is the area under the ROC curve. The closer the AUC value is to 1, the better the classification performance of the model. An AUC value of 0.5 means that the model has no distinguishing ability (equivalent to random classification).

[0113] By calculating the ROC curve and AUC value, the overall classification effect of the model can be intuitively evaluated. Especially in the case of unbalanced sample categories, the AUC value can better reflect the true performance of the model than the accuracy rate.

[0114] Step 4: Extract model features through the trained XGBoost model, and then sort them in reverse order according to the model feature weight scores: Sort(Feature[1−N]).

[0115] Step 5: Obtain an improved interpretable deep learning network

[0116] The interpretable deep learning network is divided into an encoder and a decoder, where the encoder includes several feature extraction modules connected in sequence, and the decoder includes a softmax layer.

[0117] like Figure 2 As shown, each of the feature extraction modules includes a feature concatenation layer, a dense layer Dense, a normalized batch norm and a regularized dropout layer.

[0118] First, the two feature vectors feature 1 and feature 2 with the lowest importance weight are concatenated, and then normalized batch norm and regularized dropout are performed to obtain the first layer vector L1 Embedding (128 dimensions).

[0119] Then feature 3 and L1 Embedding are concatenated, and then passed through the dense layer Dense, normalized batch norm and regularized dropout to obtain the second layer vector L2 Embedding (128 dimensions).

[0120] Then feature 4 is concatenated with L1 Embedding and L2 Embedding, and then passed through the dense layer Dense and normalized batch norm and regularized dropout to obtain the third layer vector L3 Embedding (128 dimensions).

[0121] By analogy, we finally get the feature vector L(N-1) Embedding (128 dimensions) output by the last feature extraction module.

[0122] In the last layer of the encoder, add a meaningful rule layer, such as the feature value A. As A increases, the probability of illness increases. Encode A. Assuming that the value of A is in the intervals {1.1, 1.2, 1.3, 1.4, 1.5, 1.6}, then when A falls between 1.1 and 1.2, we encode it as 0000000001, and when it falls between 1.2 and 1.3, we encode it as 0000000011, and so on. Similar to Label Encoding, the larger the original value of A, the larger the binary value of the encoded value.

[0123] Perform a dot product between the encoded value of A and the feature vector L(N-1) Embedding output by the last feature extraction module to obtain the final feature vector value extracted by the encoder.

[0124] Next, the feature vector output by the encoder is input into the decoder, that is, the risk category label is mapped according to the input feature vector value through the softmax activation function.

[0125] Step 6: Obtain the Hypoxia Risk Assessment Model

[0126] Expand the modeling space and construct negative samples based on historical samples. For example, a sample data is as follows: Feature 1, Feature 2, Feature 3, Feature A, Feature 5, Label 1 is sick. According to known experience, the larger the feature A, the greater the probability of illness. Then you can construct data like the following to constrain the model so that the model can learn more data. The negative sample data example is as follows:

[0127] [Feature 1, Feature 2, Feature 3, Feature A+1, Feature 5, Label 1

[0128] Feature 1, Feature 2, Feature 3, Feature A+2, Feature 5, Label 1

[0129] Feature 1, Feature 2, Feature 3, Feature A+3, Feature 5, Label 1

[0130] feature 1, feature 2, feature 3, feature A+4, feature 5, Label1].

[0131] The constructed sample data are combined into a negative sample data set, and the constructed interpretable deep learning network is trained using the historical sample data set and the negative sample data set to finally obtain a hypoxia risk assessment model.

[0132] Example 3

[0133] This embodiment provides an interpretable painless endoscopic hypoxia risk assessment device, comprising:

[0134] Patient information acquisition module: used to obtain clinical characteristics of patients who are to undergo painless digestive endoscopy, wherein the clinical characteristics include individual characteristics, medical history information, and relevant evaluation parameters before endoscopy;

[0135] A hypoxia risk assessment result acquisition module is used to input the clinical characteristics into a pre-built hypoxia risk assessment model to obtain a hypoxia risk assessment result;

[0136] Wherein, the method for obtaining the hypoxia risk assessment model includes:

[0137] Obtain a historical sample data set of patients who underwent painless digestive endoscopy, and construct a negative sample data set based on the historical sample data set, wherein each sample data set includes clinical characteristics and risk category labels;

[0138] The importance of each clinical feature in the historical sample data set is ranked;

[0139] An improved interpretable model is obtained based on the importance ranking result, wherein the improved interpretable model includes an encoder and a decoder connected to each other, the encoder algorithm incorporates the learning of correlation rules between clinical features and hypoxia risk, and is used to encode input data to obtain a feature vector containing the correlation rule, and the decoder is used to perform risk category label mapping according to the feature vector;

[0140] The improved interpretable model is trained by using the historical sample data set and the negative sample data set to obtain a hypoxia risk assessment model.

[0141] Example 4

[0142] This embodiment provides a computer storage medium having a computer program stored thereon. When the computer program is executed by a processor, the explainable painless endoscopic hypoxia risk assessment method as described in any step of Embodiment 2 is implemented.

[0143] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0144] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0145] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0146] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0147] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. An interpretable method for assessing the risk of hypoxia during painless digestive endoscopy, characterized in that: include: Obtaining clinical characteristics of patients who are to undergo painless digestive endoscopy, wherein the clinical characteristics include demographic characteristics, basic physiological indicators, medical history information, and relevant pre-endoscopy assessment parameters; Inputting the clinical characteristics into a pre-trained hypoxia risk assessment model to obtain a hypoxia risk assessment result; The demographic characteristics include: age, gender, height, weight, and BMI; The basic physiological indicators include: basic blood pressure, basic heart rate, basic finger pulse oxygen saturation; The medical history information includes: history of basic hypertension, history of diabetes, history of smoking, history of alcoholism, and snoring; The relevant evaluation parameters before the mirror include: neck circumference, Mahalanobis grade, ASA grade; Wherein, the method for obtaining the hypoxia risk assessment model includes: Obtain a historical sample data set of patients who underwent painless digestive endoscopy, and construct a negative sample data set based on the historical sample data set, wherein each sample data set includes clinical characteristics and risk category labels; Using a pre-trained XGBoost model to rank the importance of each clinical feature in the historical sample data set, and obtaining the importance ranking result of each clinical feature in the historical sample data set; An improved interpretable model is obtained based on the importance ranking result, wherein the improved interpretable model includes an encoder and a decoder connected to each other, the encoder algorithm incorporates the learning of correlation rules between clinical features and hypoxia risk, and is used to encode input data to obtain a feature vector containing the correlation rule, and the decoder is used to perform risk category label mapping according to the feature vector; The improved interpretable model is trained by using the historical sample data set and the negative sample data set to obtain a hypoxia risk assessment model; Wherein, the training method of the XGBoost model includes: Segmenting the historical sample data set to obtain corresponding training sets and test sets; Perform hyperparameter adjustment and optimization on the XGBoost model on the training set and test set to obtain the optimal parameter combination; The XGBoost model is set according to the optimal parameter combination to obtain a trained XGBoost model; Wherein, constructing a negative sample data set based on the historical sample data set includes: Clinical features that were positively correlated with the risk of hypoxic events were selected as target features; Searching for sample data containing the target feature in the historical sample data set; The target feature value is increased in the found sample data to obtain new sample data.

2. The method for risk assessment of hypoxic events during painless digestive endoscopy according to claim 1, characterized in that: The historical sample data set of patients undergoing painless digestive endoscopy is obtained, including: The historical clinical characteristics of outpatients and inpatients who underwent painless gastroenteroscopy were collected; Data cleaning and standardization of each patient's historical clinical characteristics; The processed historical clinical characteristics of each patient are combined with the corresponding hypoxia risk labels to obtain a historical sample data set.

3. The method for risk assessment of hypoxic events during painless digestive endoscopy according to claim 2, characterized in that: The hyperparameters of the XGBoost model are adjusted and optimized on the training set and the test set to obtain the optimal parameter combination, including: Input the training set into the XGBoost model, and preliminarily set the XGBoost model according to conventional parameters and the specified number of classification categories; Using a grid search algorithm to perform hyperparameter adjustment on the initially set XGBoost model on the training set to obtain different parameter combinations of the XGBoost model; The XGBoost models with different parameter combinations are input into the test set for verification to obtain the optimal parameter combination of the XGBoost model.

4. The method for risk assessment of hypoxic events during painless digestive endoscopy according to claim 2, characterized in that: The encoder includes a plurality of feature extraction modules connected in sequence and a rule layer connected to the last feature extraction module, wherein each feature extraction module includes a feature splicing layer, the input of the feature splicing layer in the nth feature extraction module includes the feature vector output by the n-1th feature extraction module and the feature vector of the clinical feature at the nth position according to the result of sorting in reverse order of importance, and the rule layer stores the encoding value of the target feature, which is used to perform a dot product between the feature vector output by the last feature extraction module and the encoding value of the target feature; The decoder includes a softmax layer.

5. The method for risk assessment of hypoxic events during painless digestive endoscopy according to claim 4, characterized in that: The improved interpretable model is trained by using the historical sample data set and the negative sample data set to obtain a hypoxia risk assessment model, including: Input each sample data in the historical sample data set and the negative sample data set into the improved interpretable model, extract features from the feature vectors in each sample data through an encoder, and map risk category labels according to the feature vectors output by the encoder through a decoder; The loss function value of the improved interpretable model is calculated based on the risk category label output by the decoder and the true label corresponding to each sample data; The model parameters were adjusted to minimize the loss function value to obtain the hypoxia risk assessment model.

6. An interpretable painless digestive endoscopy hypoxic event risk assessment device, characterized in that: include: Patient information acquisition module: used to obtain clinical characteristics of patients who are to undergo painless digestive endoscopy, wherein the clinical characteristics include demographic characteristics, basic physiological indicators, medical history information, and relevant pre-endoscopy assessment parameters; A hypoxia risk assessment result acquisition module is used to input the clinical characteristics into a pre-built hypoxia risk assessment model to obtain a hypoxia risk assessment result; The demographic characteristics include: age, gender, height, weight, and BMI; The basic physiological indicators include: basic blood pressure, basic heart rate, basic finger pulse oxygen saturation; The medical history information includes: history of basic hypertension, history of diabetes, history of smoking, history of alcoholism, and snoring; The relevant evaluation parameters before the mirror include: neck circumference, Mahalanobis grade, ASA grade; Wherein, the method for obtaining the hypoxia risk assessment model includes: Obtain a historical sample data set of patients who underwent painless digestive endoscopy, and construct a negative sample data set based on the historical sample data set, wherein each sample data set includes clinical characteristics and risk category labels; Using a pre-trained XGBoost model to rank the importance of each clinical feature in the historical sample data set, and obtaining the importance ranking result of each clinical feature in the historical sample data set; An improved interpretable model is obtained based on the importance ranking result, wherein the improved interpretable model includes an encoder and a decoder connected to each other, the encoder algorithm incorporates the learning of correlation rules between clinical features and hypoxia risk, and is used to encode input data to obtain a feature vector containing the correlation rule, and the decoder is used to perform risk category label mapping according to the feature vector; The improved interpretable model is trained by using the historical sample data set and the negative sample data set to obtain a hypoxia risk assessment model; Wherein, the training method of the XGBoost model includes: Segmenting the historical sample data set to obtain corresponding training sets and test sets; Perform hyperparameter adjustment and optimization on the XGBoost model on the training set and test set to obtain the optimal parameter combination; The XGBoost model is set according to the optimal parameter combination to obtain a trained XGBoost model; Wherein, constructing a negative sample data set based on the historical sample data set includes: Clinical features that were positively correlated with the risk of hypoxic events were selected as target features; Searching for sample data containing the target feature in the historical sample data set; The target feature value is increased in the found sample data to obtain new sample data.

7. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the interpretable painless digestive endoscopy hypoxic event risk assessment method as described in any one of claims 1 to 5 is implemented.

Images