A gastric content model training method, medium, device and application thereof

Abstract

The present application relates to the medical field, and specifically discloses a gastric content model training method, medium, device and application thereof. The present application aims to solve the problem of inaccurate prediction of the existing gastric content model, and specifically provides a gastric content model with higher accuracy, a training method, medium, device and application thereof.

Description

Technical Field

[0001] This invention relates to the field of medical prediction, specifically to a method, medium, device, and application of a gastric contents model training method. Background Technology

[0002] With advancements in medical technology and increasing clinical demands, the prevention and management of perioperative complications have become increasingly important. Particularly in anesthesia-related procedures such as inpatient surgery, day surgery, and outpatient gastrointestinal endoscopy, perioperative reflux aspiration has become one of the most serious complications. Reflux aspiration can lead to a series of serious pulmonary complications, including airway obstruction, pulmonary edema, atelectasis, and aspiration pneumonia, and in some cases, it can even be life-threatening. Statistics show that the mortality rate due to aspiration during surgery is as high as 5%, accounting for 9% of all anesthesia-related deaths. Especially in gastrointestinal endoscopy, reflux aspiration has become one of the most significant serious complications, further highlighting the importance of accurately predicting the amount of gastric contents and assessing the risk of reflux.

[0003] Existing gastric contents models primarily rely on two variables: age and cross-sectional area of ​​the gastric antrum (CSA), using linear regression models for prediction. For example, a model might be calculated as: Volume = 27.0 + 14.6 * Right transverse abdominis muscle CSA - 1.28 * Age. While this method provides some useful information, it suffers from biases in practical applications due to its failure to incorporate other patient complications and its reliance on a simple linear relationship. This is particularly problematic for patients with complex medical histories, multiple influencing factors, and different surgical procedures, where previous models demonstrate significant shortcomings in accuracy and stability, failing to adequately meet clinical needs. Therefore, developing new gastric contents assessment models is urgently needed. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of inaccurate prediction by existing gastric contents models, and specifically to provide a more accurate gastric contents model, its training method, medium, device, and application.

[0005] The basic technical solution of the present invention is: a training method for a gastric contents model, including information collection, wherein the information includes Perlas classification.

[0006] In recent years, machine learning, as a part of artificial intelligence, has received much attention. Therefore, some technicians in this field have adopted machine learning methods to train new models of gastric contents. When training a new model, machine learning usually includes several steps such as data collection, data processing, feature determination, model training, and validation and testing. When collecting data, previous models usually collected basic patient information and the cross-sectional area of ​​the gastric antrum as training information. They also collected a large number of candidate predictive factors, such as basic information, preoperative management data, clinical information, and surgical history. In addition to collecting basic information such as age, height, and weight, a series of diseases are also collected, such as asthma, Parkinson's disease, coronary heart disease, hypertension, cirrhosis, neurological complications, and cerebral infarction. Gastrointestinal surgical history is also collected. After collecting data, the data is processed and filtered by algorithms to obtain the filtered features, which are then used for training to obtain the model.

[0007] In model training, it has always been difficult to achieve the ideal prediction accuracy when using models such as linear regression and random forest. In response, the inventors discovered through research that after incorporating Perlas hierarchical information, the accuracy of the trained model is significantly better than that of previous models.

[0008] The Perlas classification is a system used to assess the state of the gastric antrum, primarily in ultrasound examinations. It classifies the gastric antrum into different levels based on its contents to help determine the risk of aspiration. Level 0: No gastric contents are visible in either the supine or right lateral decubitus position, indicating no gastric contents and a low risk of aspiration; Level 1: A small amount of fluid is observed only in the right lateral decubitus position (generally considered ≤1.5 mL / kg), indicating a small volume of fluid in the stomach and a low risk of aspiration; Level 2: Significant fluid is visible in both the supine and right lateral decubitus positions (fluid volume >1.5 mL / kg), indicating a large amount of gastric contents and a high risk of aspiration.

[0009] Perlas classification is typically used qualitatively to assess the amount of gastric fluid via ultrasound examination. However, the inventors incorporated its characteristics and used them for quantitative purposes through model training, thereby improving the accuracy of predicting gastric fluid volume and the accuracy of predicting the risk of reflux aspiration.

[0010] Furthermore, the information also includes age and / or the cross-sectional area of ​​the gastric antrum in the right lateral decubitus position.

[0011] The inventors' research revealed that the optimal combination of information—Perlas classification + right lateral decubitus antral cross-sectional area (RLD CSA) + age—performed best in the model. Furthermore, Perlas classification + RLD CSA outperformed RLD CSA alone, and Perlas classification + age also outperformed age alone. These results were superior to previous models. Previous models used the volume formula developed by Professor Perlas et al.: Gastric volume (ml) = 27.0 + 14.6 × right lateral decubitus antral cross-sectional area - 1.28 × age.

[0012] Another fundamental technical solution provided by this invention is a gastric contents model obtained using the above training method.

[0013] Furthermore, the features collected by the model are Perlas classification; or Perlas classification and age; or Perlas classification and right lateral decubitus antral cross-sectional area; or Perlas classification, age and right lateral decubitus antral cross-sectional area.

[0014] Furthermore, the model includes any one of the following: random forest model, extreme gradient boosting model, adaptive boosting model, gradient boosting decision tree model, neural network model, K nearest neighbor model, support vector machine model, and linear regression model.

[0015] Furthermore, the random forest model uses the following parameters: max_depth=10, max_features=2, min_samples_leaf=2, min_samples_split=5, n_jobs=-1, random_state=90;

[0016] The parameters used in the extreme gradient boosting model are: colsample_bytree=1.0, gamma=0.2, learning_rate=0.2, max_depth=3, n_estimators=50, subsample=1.0;

[0017] The adaptive augmentation model uses the following parameters: DecisionTreeRegressor(max_depth=3), random_state=90, learning_rate=0.01, loss='exponential', n_estimators=50;

[0018] The gradient boosting decision tree model uses the following parameters: learning_rate=0.01, max_depth=3, max_features='sqrt', min_samples_leaf=2, min_samples_split=5, n_estimators=200;

[0019] The neural network model uses the following parameters: activation='tanh', alpha=0.01, hidden_layer_sizes=(50,), learning_rate='constant', solver='adam';

[0020] The parameters used in the K-nearest neighbor model are: algorithm='auto', leaf_size=20, n_neighbors=10, p=2, weights='uniform';

[0021] The support vector machine model uses the following parameters: C=0.1, gamma='scale', kernel='linear';

[0022] The formula used in the linear regression model is:

[0023] .

[0024] The optimal parameters for the model vary depending on the selected indicators. The parameters above are the optimal parameters for each model when Perlas classification, age, and the cross-sectional area of ​​the gastric antrum in the right lateral decubitus position are selected. Their accuracy is far superior to that of previous models.

[0025] This invention uses the RF method to construct a model for predicting gastric contents: compared to previous models that only used linear regression to predict gastric contents, this model introduces the RF method to construct a model for gastric contents. Based on an ensemble learning method using decision trees, the model's stability and accuracy are improved by combining the prediction results of multiple decision trees, especially demonstrating excellent performance in handling nonlinear relationships and outliers.

[0026] Models for predicting gastric contents are constructed using the XGBoost method: a gradient boosting-based decision tree algorithm that effectively handles complex nonlinear relationships and prevents overfitting through multi-round iterative optimization and built-in regularization, making it suitable for large-scale high-dimensional data.

[0027] Using the AdaBoost method to build a model for predicting stomach contents: AdaBoost is an ensemble learning method that combines multiple weak learners (such as decision trees) to build a powerful model that can gradually improve the model's predictive ability and stability, making it suitable for complex data environments.

[0028] Using the GBDT method to build a model for predicting gastric contents: GBDT is an ensemble learning method that improves prediction accuracy by progressively building a combination of decision trees based on residuals. It is especially suitable for high-dimensional and non-linear data and has the advantages of automatically selecting features and avoiding overfitting.

[0029] Using the KNN method to build a model for predicting gastric contents: As a non-parametric supervised learning algorithm, KNN does not require assumptions about the data distribution and predicts the target variable by the distance to its neighbors, making it suitable for handling complex features and non-linear data.

[0030] Using the SVM method to build a model for predicting gastric contents: SVM, as a powerful supervised learning algorithm, can effectively handle high-dimensional data and performs particularly well in dealing with nonlinear problems. The core idea of ​​SVM is to find an optimal hyperplane to maximize the boundary between different categories or data points, thereby achieving data classification or regression.

[0031] Using the LR method to build a model for predicting gastric contents: Compared with the traditional LR, the LR of this invention can combine cross-validation and parameter tuning to build a more accurate prediction model.

[0032] Models for predicting gastric contents are constructed using the MLP method: MLP is a feedforward neural network that can capture complex feature relationships through multi-layer nonlinear transformations, making it particularly suitable for handling high-dimensional and nonlinear data. MLP has strong fitting capabilities, and the model weights are optimized through backpropagation, thereby improving prediction accuracy. At the same time, overfitting can be effectively avoided with appropriate regularization.

[0033] Another fundamental technical solution provided by this invention is: utilizing the storage medium of the above model. That is, hardware-based implementation of the above model, including chips, hard disks, and other computer-level software readable carriers.

[0034] Another fundamental technical solution provided by this invention is a device applying the above-described model. In practical applications, the above-described model can be used in ultrasound devices for examining the gastric antrum in patients.

[0035] Another fundamental technical solution provided by this invention is the application of the above-mentioned gastric contents model as a device for predicting gastric contents volume.

[0036] Another fundamental technical solution provided by this invention is the application of the above-mentioned gastric contents model as a device for predicting the risk level of reflux aspiration.

[0037] The model of this invention shows significantly better performance than previous models in both predicting gastric contents volume and predicting medium-to-high and high-risk reflux aspiration, and it also exhibits stability and consistency in different validation environments. Attached Figure Description

[0038] Figure 1 This is a graph of LASSO screening variables in Embodiment 1 of the present invention;

[0039] Figure 2 This is a flowchart of the prediction model construction in Embodiment 3 of the present invention;

[0040] Figure 3 This is a diagram showing the SHAP global analysis and single-sample analysis of the predictive model interpretability analysis in Embodiment 3 of the present invention;

[0041] Figure 4 The graph shows the prediction performance of each model for high-risk aspiration in the internal validation set;

[0042] Figure 5 The graph shows the prediction performance of each model for high risk of aspiration on the internal validation set;

[0043] Figure 6 The graph shows the prediction performance of each model for high-risk aspiration in the external validation set;

[0044] Figure 7 The graph shows the prediction performance of each model on the external validation set for the high risk of accidental aspiration. Detailed Implementation

[0045] I. Definitions:

[0046] CSA: Cross-sectional area of ​​the gastric antrum;

[0047] RLD CSA: Cross-sectional area of ​​the gastric antrum in right lateral decubitus position;

[0048] Age: age;

[0049] Previous model: The volume formula studied by Professor Perlas et al.: Gastric contents (ml) = 27.0 + 14.6 × cross-sectional area of ​​the gastric antrum in right lateral decubitus position - 1.28 × age;

[0050] RF: Random Forest model;

[0051] XGBoost: Extreme Gradient Boosting Model;

[0052] AdaBoost: Adaptive Boosting Model;

[0053] GBDT: Gradient Boosting Decision Tree;

[0054] KNN: K-Nearest Neighbors model; MLP: Neural Network model;

[0055] SVM: Support Vector Machine model;

[0056] LR: Linear Regression Model.

[0057] The following detailed description illustrates the specific implementation method:

[0058] II. Example 1: Filtering and Collecting Features

[0059] 1. Information Collection: Various clinical data potentially related to gastric emptying during the perioperative period were collected, including the patient's general condition, comorbidities, fasting and fluid restriction status, and relevant ultrasound examination information. General patient information included age, gender, height, and weight. Comorbidities included Glasgow Outcome Scale (GOS) score, gastritis, antral gastritis, chronic non-atrophic gastritis, gastric ulcer, duodenal ulcer, gastric retention, achalasia, esophagitis, gastroesophageal reflux, esophageal ulcer, esophageal cancer, bile reflux, hiatal hernia, gastroesophageal mucosal ectopia, gastric polyps, gastric stromal tumors, gastric cancer, gastric masses, duodenal tumors, and esophageal leiomyomas. The patient's medical history includes: cirrhosis, portal hypertensive gastropathy, upper gastrointestinal bleeding, liver dysfunction, fatty liver, liver cysts, hepatic hemangioma, liver space-occupying lesions, primary cholangitis, gallbladder polyps, ulcerative colitis, Crohn's disease, intestinal obstruction, intestinal polyps, pancreatitis, pancreatic tumors, peritonitis, diabetes, hypothyroidism, and a history of abdominal surgery. Fasting and fluid restriction parameters include fasting time, fluid restriction time, and simethicone consumption. Ultrasound parameters include maximum anteroposterior diameter (AP) cm, maximum craniocaudate diameter (CC) cm, gastric antral cross-sectional area (CSA) (cm²), and Perlas score. Outcome information includes gastric contents volume.

[0060] 2. Randomly divide the data into training set and internal validation set: The data from a hospital (Center 1) in Chongqing were randomly divided into training set (415 cases) and internal validation set (179 cases) in a ratio of 7:3.

[0061] 3. Feature Selection: Variable selection was performed using data from a hospital in Chongqing (Center 1) in the training set. LASSO regression was used to automatically select the regularization parameter λ that minimizes the model's mean squared error through cross-validation. For example... Figure 1 As shown, among the initially included features, three key information variables were ultimately selected: age, Perlas classification, and right lateral decubitus antral cross-sectional area (RLD CSA).

[0062] By employing LASSO regression to screen multiple variables, the problem of multicollinearity is avoided, ensuring the stability and accuracy of the model. This approach significantly improves the model's ability to handle complex data, enabling this invention to capture deeper-level variable relationships and surpassing the limitations of previous models.

[0063] 4. Model Performance Evaluation: The parameters of the AdaBoost model were optimized using a grid search combined with five-fold cross-validation, and the optimal parameters were applied to the model to evaluate its performance on gastric contents volume and its ability to predict the risk of aspiration.

[0064] 5. Using the AdaBoost model with different features, we validated the model performance on the training set and the internal validation set. The results are shown in Table 1.

[0065] Table 1: Model performance for different features on the training set and internal validation set (using AdaBoost model)

[0066]

[0067] Therefore, it is evident that the feature combination "Perlas classification + RLD CSA + age" performs best in the training set, achieving the lowest root mean square error (RMSE: 21.68) and mean absolute error (MAE: 16.35). Furthermore, the "Perlas classification + RLD CSA" model outperforms RLD CSA alone (RMSE: 22.43 vs. 26.66; MAE: 16.95 vs. 21.19), and the "Perlas classification + age" model also outperforms age alone (RMSE: 21.91 vs. 27.17; MAE: 16.67 vs. 21.47). Moreover, the results for all features and their combinations are superior to previous models.

[0068] In internal validation, the performance of different combinations of age, Perlas classification, and RLD CSA in the model was further explored. The results were basically consistent with those of the training set. Therefore, considering the model performance of various factors, the combination of age, Perlas classification, and RLD CSA is considered the best choice.

[0069] 6. Using the AdaBoost model, different information variables were used to predict the gastric contents volume using the internal validation set. The results are shown in Table 2.

[0070] Table 2: Bland-Altman analysis results for different information variables on the internal validation set (using the AdaBoost model).

[0071]

[0072] Therefore, the performance of different combinations of age, Perlas classification, and RLD CSA in the model is basically consistent with that in the training set. Consistency analysis of different feature combinations in predicting gastric volume shows that their mean bias range (-0.6 to 2.3 mL) is significantly lower than the 23.5 mL of previous models. Although the mean bias of the "Perlas classification + RLD CSA" and "Perlas classification + Age" combinations is slightly smaller than that of "Perlas classification + RLD CSA + Age," the difference is not significant. Furthermore, in terms of 95% LOA, "Perlas classification + RLD CSA + Age" is superior to "Perlas classification + Age," and significantly superior to using age, RLD CSA, or the "Age + RLD CSA" combination alone. Of course, regardless of the combination used, the model's predictive accuracy far exceeds that of previous models. Considering that age may have potential additional explanatory power in the prediction, "Perlas + RLD CSA + Age" was ultimately chosen as the optimal parameter combination to more comprehensively reflect the individual characteristics of patients while ensuring predictive accuracy. Furthermore, this combination offers more intuitive interpretation in clinical practice, helping physicians to make more personalized judgments when assessing gastric capacity by considering the patient's age. Therefore, considering both model performance and clinical application needs, "Perlas + RLD CSA + Age" is considered the optimal choice.

[0073] 7. Using the AdaBoost model, with different feature variables, and using the internal validation set, we predicted the high-risk and high-risk reflux aspiration. The data results are shown in Tables 3 and 4:

[0074] Table 3. Prediction of high-risk aspiration by different features in the internal validation set.

[0075]

[0076] Table 4: Prediction of high risk of aspiration by different features in the internal validation set

[0077]

[0078] Therefore, in the internal validation of aspiration risk, the Perlas classification, and different combinations of features with age and RLD CSA, all showed AUCs above 0.7 for predicting medium- to high-risk aspiration, significantly outperforming previous models (AUC: 0.629). Specifically, the "Perlas classification + RLD CSA" combination outperformed RLD CSA alone in predicting medium- to high-risk aspiration (AUC: 0.774 vs. 0.612), while the "Perlas classification + age" combination outperformed age alone (AUC: 0.717 vs. 0.47). In predicting high-risk aspiration, the performance of each parameter combination was largely consistent with the internal validation results. The Perlas classification and its different combinations with age or RLD CSA all showed AUCs above 0.8 for predicting high-risk aspiration, significantly outperforming previous models (AUC: 0.743). Among them, the "Perlas classification + RLD CSA" combination outperformed RLD CSA alone in predicting high-risk aspiration (AUC: 0.913 vs. 0.764), while the "Perlas classification + age" combination outperformed age alone (AUC: 0.8 vs. 0.461). The Perlas + RLD CSA + Age combination achieved the highest AUC of 0.94 in predicting high-risk aspiration.

[0079] In summary, the optimal feature selection in this invention is the combination of "age, Perlas rating, and RLD CSA," which is then applied to other learning models and external validation sets. Secondly, "Perlas rating + RLD CSA" can be selected. Thirdly, "Perlas rating + age" can be chosen, or Perlas rating alone can be selected and applied to other learning models and external validation sets.

[0080] Feature selection in other models also showed results consistent with the AdaBoost model, with the optimal choice being the combination of "age, Perlas classification, and RLD CSA". The next best option was "Perlas classification + RLD CSA", followed by "Perlas classification + age", or Perlas classification alone; these will not be elaborated further.

[0081] III. Example 2: Model Training

[0082] Let's take the RF model as an example:

[0083] 1. Age, Perlas classification, and RLD CSA were selected as metrics to be incorporated into the Random Forest (RF) model, and the model hyperparameters were optimized using grid search combined with five-fold cross-validation. The mean squared error (MSE) was used to evaluate the model's performance. By systematically adjusting the search space of different hyperparameters, the optimal parameter combination corresponding to the minimum MSE was finally determined.

[0084] 2. During hyperparameter tuning, adjust the following parameters: n_estimators (number of trees), max_depth (maximum tree depth), min_samples_split (minimum number of samples required to split internal nodes), min_samples_leaf (minimum number of samples on leaf nodes), and max_features (maximum number of features to consider in each split). Specifically, candidate values ​​for n_estimators are 50, 100, and 200; candidate values ​​for max_depth are 10, 20, and no depth limit; candidate values ​​for min_samples_split are 2 and 5; candidate values ​​for min_samples_leaf are 1 and 2; and candidate values ​​for max_features are 2, 3, and "auto".

[0085] After optimization, the optimal combination of hyperparameters for the model is: max_depth=10, max_features=2, min_samples_leaf=2, min_samples_split=5, with n_jobs=-1 and random_state=90 set.

[0086] 3. Apply the above optimal parameters to the RF model and evaluate the model's predictive performance on the internal and external validation sets (the data results are consistent with those in Example 4). To quantify the model's accuracy, root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation metrics (the data results are consistent with those in Example 3). This yields the optimal RF model.

[0087] 4. Similarly, the optimal parameters for other models are determined in the same way, ensuring the consistency and comparability of the tuning process.

[0088] 5. Conclusion: The optimal parameters for each model are as follows:

[0089] (1) The parameters used in the extreme gradient boosting model are: colsample_bytree=1.0, gamma=0.2, learning_rate=0.2, max_depth=3, n_estimators=50, subsample=1.0;

[0090] (2) The parameters used for the adaptive augmentation model are: DecisionTreeRegressor(max_depth=3), random_state=90, learning_rate=0.01, loss='exponential', n_estimators=50;

[0091] (3) The gradient boosting decision tree model uses the following parameters: learning_rate=0.01, max_depth=3, max_features='sqrt', min_samples_leaf=2, min_samples_split=5, n_estimators=200;

[0092] (4) The parameters used in the neural network model are: activation='tanh', alpha=0.01, hidden_layer_sizes=(50,), learning_rate='constant', solver='adam';

[0093] (5) The parameters used in the K-nearest neighbor model are: algorithm='auto', leaf_size=20, n_neighbors=10, p=2, weights='uniform';

[0094] (6) The parameters used in the support vector machine model are: C=0.1, gamma='scale', kernel='linear';

[0095] (7) The formula used for the linear regression model is:

[0096] .

[0097] IV. Example 3: Application of gastric contents model for predicting gastric contents volume

[0098] Taking the Random Forest (RF) model as an example, the application of the gastric contents model of this invention for predicting gastric contents volume will be explained.

[0099] 1. Sample selection:

[0100] like Figure 2As shown, using clinical data from 793 patients as the research platform, a predictive model for gastric contents was constructed using the RF model. Data from a hospital in Chongqing (Center 1) was randomly divided into a training set and an internal validation set at a ratio of 7:3, with 415 cases in the training set and 179 cases in the internal validation set; data from a district hospital in Chongqing (Center 2) was used as the external validation set, with 199 cases in the external validation set.

[0101] 2. Data Collection Features:

[0102] Age, Perlas level, and RLD CSA parameters were selected, with categorical variables using one-hot encoding. Continuous features were standardized for models relying on distance metrics or gradient optimization. The median age in the training set was 51 (39, 58), in the internal validation set it was 50 (37, 58), and in the external validation set it was 54 (47, 64). The proportions of Perlas Level 1 models in the training, internal, and external validation sets were 57.8%, 57%, and 57.3%, respectively, while the proportions of Perlas Level 2 models were 11.1%, 12.8%, and 8%, respectively. The median RLD CSA in the training, internal, and external validation sets was 5.69 (4.61, 7.37) cm⁻¹. 2 5.71 (4.67, 6.93) cm 2 And 5.43 (4.35, 6.72) cm 2 The median gastric contents in the training set, internal validation set, and external validation set were 25 (10, 50) mL, 20 (10, 45) mL, and 25 (15, 40) mL, respectively.

[0103] 3. Model parameter selection:

[0104] Age, Perlas classification, and RLD CSA were incorporated into the RF model. An optimal parameter combination was selected using grid search combined with five-fold cross-validation. The model's predictive performance was evaluated on both internal and external validation sets. By defining a search space for different parameters, the optimal parameters were determined to be (max_depth=10, max_features=2, min_samples_leaf=2, min_samples_split=5, n_jobs=-1, random_state=90). These optimal parameters were applied to the RF model, and its predictive performance was evaluated on both internal and external validation sets. To quantify the model's accuracy, root mean square error (RMSE) and mean absolute error (MAE) were used as evaluation metrics.

[0105] 4. Model Representation:

[0106] (1) This embodiment provides the observed gastric contents volume data from the study. The RF model's RMSE and MAE on the training set were 22.8 and 17.04, respectively, significantly lower than previous models (RMSE: 42.06; MAE: 31.46). On the internal validation set, the RMSE and MAE were 23.97 and 18.64, respectively, significantly lower than previous models (RMSE: 38.54; MAE: 32.46). On the external validation set, the RMSE and MAE were 16.79 and 13.15, respectively, significantly lower than previous models (RMSE: 32.58; MAE: 24.43). It is evident that the model in this embodiment performs far better than previous models.

[0107] (2) In the prediction of gastric contents volume in this embodiment, the prediction model of this embodiment has good consistency. The average bias of the internal validation set is 1 mL, while that of the previous model is 23.5 mL; the average bias of the external validation set is 0.4 mL, while that of the previous model is 14.9 mL. It can be seen that the prediction accuracy of this embodiment is much better than that of the previous model.

[0108] 5. Model Explanation:

[0109] like Figure 3 As shown, this invention provides a global interpretation graph of SHAP values ​​to reveal the overall decision-making logic of the model. The graph shows that the predicted gastric contents increase with increasing Perlas score, increasing RLD CSA, and decreasing age.

[0110] To further demonstrate how this model can be used to personalize the prediction of gastric contents in individual patients, this invention provides two specific cases. According to the prediction model's results, a 50-year-old patient with a Perlas score of 0 and an RLD CSA of 7.52 cm² was predicted to have a gastric contents of 87.25 mL; while a 71-year-old patient with a Perlas score of 2 and an RLD CSA of 10.52 cm² was predicted to have a gastric contents of only 1.64 mL. Therefore, compared to existing models abroad (Volume = 27.0 + 14.6 × RLD CSA - 1.28 × Age), the model in this embodiment demonstrates better risk identification capabilities.

[0111] V. Other Models Used for Predicting Gastric Content Volume

[0112] 1. The specific implementation method is the same as that in Example 3. The only difference is that the optimal parameters determined by each model are different and the predicted results are different.

[0113] 2. The optimal parameters of each model are consistent with those of Example 2.

[0114] 3. Performance data of each model

[0115] (1) The model performance of each model on the training set, internal validation set, and external validation set is shown in Table 5:

[0116] Table 5 shows the model performance on the training set, internal validation set, and internal validation set.

[0117]

[0118] (2) The prediction performance of each model on the internal and external validation sets for gastric contents volume is shown in Table 6:

[0119] Table 6: Bland-Altman analysis results for each model on the internal and external validation sets.

[0120]

[0121] The results above show that the models of this invention outperform previous models in terms of model performance on the training set, internal validation set, and internal validation set, as well as in terms of predicting gastric contents volume on the internal validation set and external validation set.

[0122] VI. Example 4: Application of gastric contents model in predicting intermediate-to-high and high-risk reflux aspiration.

[0123] Taking the RF model as an example, the application of the gastric contents model of the present invention in medium-to-high risk and high-risk reflux aspiration is explained.

[0124] 1. Sample selection:

[0125] like Figure 2 As shown, using clinical data from 793 patients as the research platform, a predictive model for gastric contents was constructed using the RF model. Data from a hospital in Chongqing (Center 1) was randomly divided into a training set and an internal validation set at a ratio of 7:3, with 415 cases in the training set and 179 cases in the internal validation set; data from a district hospital in Chongqing (Center 2) was used as the external validation set, with 199 cases in the external validation set.

[0126] 2. Model Indicator Selection:

[0127] Age, Perlas level, and RLD CSA parameters were selected, with categorical variables using one-hot encoding. Continuous features were standardized for models relying on distance metrics or gradient optimization. The median age in the training set was 51 (39, 58), in the internal validation set it was 50 (37, 58), and in the external validation set it was 54 (47, 64). The proportions of Perlas Level 1 models in the training, internal, and external validation sets were 57.8%, 57%, and 57.3%, respectively, while the proportions of Perlas Level 2 models were 11.1%, 12.8%, and 8%, respectively. The median RLD CSA in the training, internal, and external validation sets was 5.69 (4.61, 7.37) cm⁻¹. 2 5.71 (4.67, 6.93) cm 2 And 5.43 (4.35, 6.72) cm 2 The median gastric contents in the training set, internal validation set, and external validation set were 25 (10, 50) mL, 20 (10, 45) mL, and 25 (15, 40) mL, respectively.

[0128] 3. Model parameter selection:

[0129] Age, Perlas classification, and RLD CSA were incorporated into the RF model. The optimal parameter combination was selected using a grid search combined with five-fold cross-validation. The model's predictive performance was evaluated on both internal and external validation sets. By defining search spaces for different parameters, the optimal parameters were ultimately determined to be (max_depth=10, max_features=2, min_samples_leaf=2, min_samples_split=5, n_jobs=-1, random_state=90).

[0130] 4. Application of the ability to predict the risk of accidental aspiration

[0131] The optimal hyperparameters were applied to the random forest (RF) model to compute the predicted gastric contents on both the internal and external validation sets. Based on the predicted gastric contents, the model's performance in predicting high and medium risk aspiration was further evaluated.

[0132] 5. Prediction Conclusion

[0133] Specifically, in the RF model, moderate to high risk of aspiration is defined as a gastric contents ≥50 mL or ≥0.8 mL / kg, and high risk of aspiration is defined as a gastric contents ≥100 mL or ≥1.5 mL / kg. For example... Figure 4 — Figure 7As shown, the calculated area under the receiver operating characteristic (AUC) curve indicates that, in the internal validation set, the RF model's predictive AUCs for moderate to high risk and high risk of aspiration are 0.736 and 0.877, respectively, significantly higher than previous models (AUC for moderate to high risk of aspiration: 0.629; AUC for high risk of aspiration: 0.743). In the external validation set, the RF model's predictive AUCs for moderate to high risk and high risk of aspiration are 0.796 and 0.946, respectively, also significantly higher than previous models (AUC for moderate to high risk of aspiration: 0.759; AUC for high risk of aspiration: 0.774). These results demonstrate that the RF model constructed in this embodiment significantly outperforms previous models in predicting aspiration risk.

[0134] VII. Application of other models in predicting the risk of aspiration

[0135] 1. The specific implementation method is the same as that in Example 4, the only difference being that the optimal parameters determined by each model are different and the predicted results are different.

[0136] 2. The optimal parameters of each model are consistent with those of Example 2.

[0137] 3. The prediction performance of each model on the internal and external validation sets for high-risk and high-risk aspiration errors, such as... Figure 4 — Figure 7 As shown, the data results are presented in Tables 7 and 8:

[0138] Table 7 shows the model's prediction performance for high and medium risk aspiration in the internal validation set.

[0139]

[0140] Table 8 shows the model's prediction performance for high and high risk of aspiration on the external validation set.

[0141]

[0142] Therefore, it can be seen that the models of this invention outperform previous models in medium-to-high and high-risk reflux aspiration.

[0143] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A training method for a gastric contents model, characterized in that, The information includes Perlas classification, age, and right lateral decubitus gastric antrum cross-sectional area, as well as screening features. The screened features used for model training include Perlas classification, age, and right lateral decubitus gastric antrum cross-sectional area. The trained gastric contents model is used to quantify gastric contents volume. The gastric contents model includes any one of the following: random forest model, extreme gradient boosting model, adaptive boosting model, gradient boosting decision tree model, neural network model, K nearest neighbor model, and support vector machine model.

2. A gastric contents model obtained using the training method of claim 1.

3. The gastric contents model according to claim 2, characterized in that: The random forest model used the following parameters: max_depth=10, max_features=2, min_samples_leaf=2, min_samples_split=5, n_jobs=-1, random_state=90; The parameters used in the extreme gradient boosting model are: colsample_bytree=1.0, gamma=0.2, learning_rate=0.2, max_depth=3, n_estimators=50, subsample=1.0; The adaptive enhancement model uses the following parameters: DecisionTreeRegressor(max_depth=3), random_state=90, learning_rate=0.01, loss='exponential', n_estimators=50; The gradient boosting decision tree model uses the following parameters: learning_rate=0.01, max_depth=3, max_features='sqrt', min_samples_leaf=2, min_samples_split=5, n_estimators=200; The neural network model uses the following parameters: activation='tanh', alpha=0.01, hidden_layer_sizes=(50,), learning_rate='constant', solver='adam'; The parameters used in the K-nearest neighbor model are: algorithm='auto', leaf_size=20, n_neighbors=10, p=2, weights='uniform'; The support vector machine model uses the following parameters: C=0.1, gamma='scale', kernel='linear'.

4. A computer-readable storage medium comprising the gastric contents model of claim 2 or 3.

5. An apparatus comprising the gastric contents model of claim 2 or 3.

6. The application of the gastric contents model of claim 2 or 3 in the preparation of a device for predicting gastric contents volume.

7. The application of the gastric contents model of claim 2 or 3 in the preparation of a device for predicting the risk level of reflux aspiration.

Images