Swallowing aspiration risk prediction method

Abstract

The invention discloses a swallowing aspiration risk prediction method, and belongs to the technical field of data prediction. The method comprises the following steps: collecting multi-dimensional clinical data of a patient, and combining Logistic regression and a convolutional neural network technology to construct a mistaken inhalation risk prediction model; the method specifically comprises six steps of data acquisition, variable screening, model construction, verification and optimization, risk classification and dynamic demarcation point optimization. According to the method, key features related to aspiration are extracted through a standardized evaluation scale, and efficient feature extraction and nonlinear analysis are realized by using fusion of deep learning and a traditional statistical technology; a dynamic demarcation point optimization technology is adopted, a classification threshold value can be adjusted in real time according to the change of the condition of a patient, the prediction sensitivity, specificity and generalization ability are remarkably improved, and a new technical path is provided for accurate prediction and personalized intervention of swallowing aspiration.

Description

Technical Field

[0001] The present invention belongs to the technical field of data prediction, and in particular relates to a method for predicting swallowing aspiration risk. Background Art

[0002] Dysphagia is a common complication in patients with stroke, neurodegenerative diseases and other neurological diseases, among which aspiration is one of the most serious risks. Aspiration not only leads to complications such as lung infection and suffocation, significantly increasing the mortality rate, but also significantly increases the patient's hospitalization time and medical expenses. At present, clinical aspiration risk assessment is mainly based on dysphagia scores, imaging examinations or the experience judgment of medical staff. These methods are highly subjective and usually rely on a single indicator and cannot fully integrate the patient's multidimensional data characteristics. This leads to limitations in the sensitivity and specificity of risk assessment, making it difficult to accurately predict high-risk patients. In addition, most of the existing methods are static assessments, which are difficult to meet the needs of rapid changes in patients' conditions in dynamic clinical environments, affecting the accuracy and timeliness of aspiration risk prediction.

[0003] In recent years, the application of artificial intelligence technology in the medical field has developed rapidly, and prediction methods based on deep learning have gradually become the main means to solve complex medical problems; in particular, convolutional neural networks, due to their excellent multi-dimensional feature extraction capabilities, have shown significant advantages in medical image analysis, complex data processing, etc.; in the prediction of aspiration risk, convolutional neural networks can extract potential high-risk features from complex clinical data, and achieve high-precision predictions through nonlinear modeling; however, relying solely on deep learning technology has certain limitations, such as the "black box" nature of the model, which makes its clinical interpretability weak, and at the same time has high requirements for data quality; in addition, deep learning methods are usually difficult to directly combine with traditional medical statistical models, and cannot fully utilize the ability of traditional methods to screen important features.

[0004] At present, aspiration risk prediction technology still faces many challenges in terms of dynamic adaptability and generalization ability. Traditional prediction models often use fixed classification thresholds. This static method is difficult to reflect the dynamic changes in the patient's condition, affecting the balance between sensitivity and specificity of risk classification. In addition, different patient groups have significant differences in risk characteristics, and the generalization ability of the model is often insufficient when applied across scenarios. In solving these problems, combining traditional statistical methods with artificial intelligence technology and improving the adaptability of the model through dynamic classification threshold optimization technology has become one of the main research directions. This method can not only achieve efficient feature screening, but also combine the powerful modeling capabilities of deep learning, which helps to comprehensively improve the accuracy and clinical applicability of risk prediction. Summary of the invention

[0005] The present invention provides a swallowing aspiration risk prediction method, which aims to integrate multidimensional data, extract key features and dynamically optimize thresholds, to achieve efficient and accurate risk prediction and real-time monitoring, and provide reliable support for clinical aspiration management.

[0006] The present invention provides a method for predicting the risk of swallowing and aspiration, which specifically comprises the following steps: S1. Data collection: Collect clinical data of patients and build a database containing key indicators; S2. Variable screening and evaluation: extracting features related to aspiration through standardized assessment scales; S3. Model construction: Construct an aspiration risk prediction model based on logistic regression and convolutional neural network technology; S4. Model validation and optimization: The robustness of the model is verified through multi-dimensional embedding and high-dimensional projection; S5. Risk classification and dynamic cut-off point optimization: Adjust the classification threshold based on a complex optimization algorithm based on sensitivity and specificity; S6. Data prediction: Input real-time clinical data into the optimized classification detection model to predict the risk of swallowing and aspiration.

[0007] Preferably, the data collection in step S1 includes the following contents: General information includes the patient's age, gender, number of days in hospital, and underlying diseases; clinical indicators include dysphagia score, cough reflex score, NIHSS score, and consciousness disorder score; physiological characteristics include oral hygiene score, nutrition method, and mechanical ventilation status; collecting these required data can ensure the comprehensiveness and diversity of the model training data, laying the foundation for building an efficient prediction model.

[0008] Preferably, extracting features related to aspiration by using a standardized assessment scale in step S2 comprises the following steps: First, the continuous variables are normalized. The specific calculation formula is: Among them, x norm is the normalized variable value; x is the original variable value; μ x is the mean of the original variable; σ x is the standard deviation of the original variable.

[0009] Then the missing data is interpolated. The specific calculation formula is: Among them, x imp is the variable value after interpolation; X is the set of other variables related to the missing variable; It is the conditional expectation, which means the expected value of variable x when other related variables X are known.

[0010] Preferably, the model building process described in step S3 includes the following steps: Step S31, Logistic regression model construction, using Logistic regression to screen independent risk factors for aspiration can enhance the model's interpretability and feature importance analysis capabilities; the specific calculation formula is: Among them, Logit(P) is the logarithmic probability of the target variable P; P is the probability of swallowing aspiration, p is the total number of characteristic variables; X i is the i-th characteristic variable; βi is the model regression coefficient.

[0011] Preferably, step S32, convolutional neural network feature extraction, uses a convolutional neural network to perform nonlinear feature extraction, and adopts multi-layer convolution and pooling operations, which can extract deep features from complex data and improve the nonlinear analysis ability of the model. The specific calculation formula is: in, is the output feature value of the (i, j)th position after the lth convolution operation; l is the sequence number of the neural network convolution layer; k is the convolution kernel size; is the weight parameter of the l-th convolution kernel at the (p, q)th position; is the input feature matrix, i.e. the output of the previous layer; b (l) is the bias term of the lth layer, indicating the offset adjustment value of the convolution kernel; σ is the activation function.

[0012] Preferably, in step S33, the fully connected layer fusion operation is performed, the Logistic regression output and the CNN feature are fused through the fully connected layer to obtain the final feature information, and the final aspiration risk probability is output. By fusing the output of the Logistic regression and the convolutional neural network, the comprehensive judgment ability of the model on the aspiration risk can be improved; the specific calculation formula is: P final =Softmax(W f ·[P Logistic , P CNN ]+b f ); Among them, P final is the final predicted risk probability, which is a multi-category probability distribution after Softmax processing; Softmax(x) is the normalization function; W f is the weight matrix of the fully connected layer; P Logistic is the aspiration risk probability output by the Logistic regression model; P CNN b is the feature probability output by the convolutional neural network model; fis the bias term of the fully connected layer.

[0013] Preferably, step S34, the multi-task learning expansion stage, supports the simultaneous prediction of other related complications, improves the multi-task processing capability and clinical applicability of the model, and introduces a multi-task loss function, and the specific calculation formula is: in, is the total loss function of multi-task learning; M is the total number of tasks; α j is the weight of the jth task; is the cross entropy loss function; y j is the true label of the jth task; is the predicted value of the jth task.

[0014] Preferably, the model verification and optimization described in step S4 includes the following steps: Step S41, feature embedding and dimensionality reduction, using a neural network to embed high-dimensional features into a low-dimensional space, and through the embedding layer dimensionality reduction processing, the computational efficiency and feature extraction capability of high-dimensional data can be improved; the specific calculation formula is: H=φ(X·W e +b e ); Among them, H is the low-dimensional embedding feature matrix; X is the high-dimensional feature matrix of the input; W e is the weight matrix of the embedding layer; b e is the bias term of the embedding layer; φ is the activation function.

[0015] Preferably, step S42, projection space verification, verifies the consistency of feature distribution through orthogonal projection, which can ensure the stability and prediction accuracy of the model in multiple scenarios; the specific calculation formula is: Among them, T is the interaction matrix of verification features; H is the low-dimensional embedding feature matrix; Q is a symmetric positive definite matrix; H is the transposed matrix of H.

[0016] Preferably, step S43, verifying the consistency index, and using the contrast loss function to evaluate the prediction consistency can further improve the reliability of the model; the specific calculation formula is: in, is the loss function for verifying consistency; n is the number of samples; S ij is the similarity matrix of samples i and j; ||H i -H j || 2 is the square of the distance between samples i and j in the embedding space.

[0017] Preferably, in step S44, multi-scenario verification, the model is verified in different scenarios, and the comprehensive ROC curve area (AUC) is calculated. Through cross-scenario ROC curve analysis, the generalization ability of the model can be evaluated to ensure its applicability in different medical environments; the specific calculation formula is: Among them, AUC combined is the area under the comprehensive ROC curve AUC value; K is the total number of verification scenarios; AUC k is the AUC value in the kth scenario.

[0018] Preferably, the risk classification and dynamic cutoff point optimization in step S5 includes the following steps: Step S51: construct a classification loss function, define a loss function combining sensitivity and specificity, and dynamically adjust the classification threshold to improve the accuracy and adaptability of classification; the specific calculation formula is: L(τ)=λ1·(1-TPR(τ))+λ2·(1-TNR(τ)); Among them, L(τ) is the classification loss function; τ is the classification threshold; λ1, λ2 are weight parameters; TPR(τ) is the true positive rate; TNR(τ) is the true negative rate.

[0019] Preferably, in step S52, dynamic demarcation point optimization, the optimal classification threshold is determined using the minimum value of the loss function, which can ensure the dynamic adaptability of the optimal classification threshold; the specific calculation formula is: Among them, τ * is the optimal classification threshold; L(τ) is the classification loss function; τ is the classification threshold; TPR(τ) is the true positive rate; TNR(τ) is the true negative rate.

[0020] Preferably, step S53, adjusting the risk probability, recalculating the adjusted risk probability based on the optimized classification threshold, and adjusting the risk probability by the optimized classification threshold can improve the classification accuracy of the model; the specific calculation formula is: Among them, P adjusted is the adjusted risk probability; TPR(τ * ) is the optimized threshold τ * True positive rate under * ) is the optimized threshold τ * The true negative rate.

[0021] Preferably, step S54 introduces regularized classification optimization, and introduces a regularized loss function based on the dynamic cutoff point to prevent overfitting of the model, which can enhance the adaptability and robustness to actual clinical data; the specific calculation formula is: in, Total loss function optimized for classification; is the binary cross entropy loss; is the L2 norm square of the model weight; γ is the weight coefficient of the cross entropy loss; δ is the weight coefficient of the regularization term.

[0022] Compared with the prior art, the traditional aspiration risk assessment method of the present invention mostly relies on a single indicator or manual experience, is highly subjective and difficult to integrate multidimensional data, and has low prediction accuracy; the present invention combines Logistic regression and convolutional neural network technology, and through in-depth analysis of patients' multidimensional clinical data, not only achieves accurate screening of independent risk factors, but also can extract deep nonlinear features from complex data, significantly improving the sensitivity and specificity of prediction; in addition, the dynamic cutoff point optimization technology is adopted, and the classification threshold can be adjusted in real time according to changes in the patient's condition, solving the problem of poor adaptability of the traditional static threshold method; the present invention also enhances the cross-scenario generalization ability and clinical applicability of the model through multi-scenario verification and multi-task expansion, providing a new technical path for accurate prediction and personalized intervention of swallowing aspiration. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The figure is an overall flow chart of a swallowing aspiration risk prediction method.

[0024] Figure 2 This is a list diagram of the data that needs to be collected.

[0025] Figure 3 Flowchart of model construction for step S3.

[0026] Figure 4 This is the model verification and optimization flow chart for step S4.

[0027] Figure 5 Flowchart for risk classification and dynamic cutoff point optimization in step S5. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0029] Please see attached Figure 1 -Attached Figure 4 The present invention provides a method for predicting the risk of swallowing and aspiration.

[0030] As attached Figure 1 The process shown in the figure is a swallowing aspiration risk prediction method proposed by the present invention, which specifically includes the following steps: S1, data collection: collecting clinical data of patients and building a database containing key indicators; S2. Variable screening and evaluation: extracting features related to aspiration through standardized assessment scales; S3. Model construction: Construct an aspiration risk prediction model based on logistic regression and convolutional neural network technology; S4. Model validation and optimization: The robustness of the model is verified through multi-dimensional embedding and high-dimensional projection; S5. Risk classification and dynamic cut-off point optimization: Adjust the classification threshold based on a complex optimization algorithm based on sensitivity and specificity; S6. Data prediction: Input real-time clinical data into the optimized classification detection model to predict the risk of swallowing and aspiration.

[0031] Furthermore, in step S1, the data collection extracts the multidimensional clinical data of the patient through the electronic medical record system as shown in the attached Figure 2 As shown, specifically including the following: General information includes the patient's age, gender, number of days in hospital, and underlying diseases; clinical indicators include dysphagia score, cough reflex score, NIHSS score, and consciousness disorder score; physiological characteristics include oral hygiene score, nutrition method, and mechanical ventilation status; collecting these required data can ensure the comprehensiveness and diversity of the model training data, laying the foundation for building an efficient prediction model.

[0032] Further, extracting features related to aspiration by using a standardized assessment scale as described in step S2 includes the following steps: First, normalize the continuous variables using Python's sklearn.preprocessing.StandardScaler method, with the mean set to 0 and the standard deviation set to 1. The specific calculation formula is: Among them, x norm is the normalized variable value; x is the original variable value; μ x is the mean of the original variable; σ x is the standard deviation of the original variable.

[0033] Furthermore, missing data interpolation is performed using Python's sklearn.impute.IterativeImputer method to interpolate missing data to ensure data integrity; the specific calculation formula is: Among them, x imp is the variable value after interpolation; X is the set of other variables related to the missing variable; It is the conditional expectation, which means the expected value of variable x when other related variables X are known.

[0034] Further, as attached Figure 3 As shown, the model building process described in step S3 includes the following steps: Step S31 is to construct a Logistic regression model, and use Logistic regression to screen the independent risk factors of aspiration using the data to be predicted. Specifically, Python's sklearn.linear_model.LogisticRegression is used for modeling; the specific calculation formula is: Among them, Logit(P) is the logarithmic probability of the target variable P; P is the probability of swallowing aspiration, p is the total number of characteristic variables; X i is the i-th characteristic variable; β i is the model regression coefficient.

[0035] Furthermore, step S32 is convolutional neural network feature extraction, which uses a convolutional neural network to extract nonlinear features of the data to be predicted, and adopts multi-layer convolution and pooling operations. Specifically, the CNN network is built using PyTorch's torch.nn.Conv2d and torch.nn.MaxPool2d, the convolution kernel size is set to 3×3, and the activation function is ReLU; the specific calculation formula is: in, is the output feature value of the (i, j)th position after the lth convolution operation; l is the sequence number of the neural network convolution layer; k is the convolution kernel size; is the weight parameter of the l-th convolution kernel at the (p, q)th position; is the input feature matrix, i.e. the output of the previous layer; b (l) is the bias term of the lth layer, indicating the offset adjustment value of the convolution kernel; σ is the activation function.

[0036] Furthermore, step S33 is a fully connected layer fusion operation, in which the Logistic regression output and the CNN features are fused through the fully connected layer to obtain the final feature information, and the final aspiration risk probability is output. Specifically, the Logistic regression and CNN features are spliced ​​through PyTorch's torch.cat, and the output activation function of the fully connected layer is set to Softmax; the specific calculation formula is: P final =Softmax(W f ·[P Logistic , P CNN ]+b f ); Among them, P final is the final predicted risk probability, which is a multi-category probability distribution after Softmax processing; Softmax(x) is the normalization function; W f is the weight matrix of the fully connected layer; P Logistic is the aspiration risk probability output by the Logistic regression model; P CNN b is the feature probability output by the convolutional neural network model; f is the bias term of the fully connected layer.

[0037] Furthermore, step S34 is a multi-task learning expansion stage, which supports the simultaneous prediction of other related complications. Specifically, the weighted cross entropy loss function is set using torch.nn.CrossEntropyLoss of PyTorch, and the task weight parameter weight = [0.7, 0.3]; the specific calculation formula is: in, is the total loss function of multi-task learning; M is the total number of tasks; α j is the weight of the jth task; is the cross entropy loss function; y j is the true label of the jth task; is the predicted value of the jth task.

[0038] Further, as attached Figure 4 As shown, the model verification and optimization described in step S4 includes the following steps: Step S41 is feature embedding dimensionality reduction, which uses a neural network to embed the high-dimensional feature prediction model into a low-dimensional space. Specifically, PyTorch's torch.nn.Embedding is used for dimensionality reduction, and the embedding dimension is set to 128. The specific calculation formula is: H=φ(X·W e +b e ); Among them, H is the low-dimensional embedding feature matrix; X is the high-dimensional feature matrix of the input; W e is the weight matrix of the embedding layer; b e is the bias term of the embedding layer; φ is the activation function.

[0039] Furthermore, step S42 is projection space verification, which verifies the consistency of feature distribution through orthogonal projection, specifically by calculating the feature projection matrix through Python's numpy.dot; the specific calculation formula is: Among them, T is the interaction matrix of verification features; H is the low-dimensional embedding feature matrix; Q is a symmetric positive definite matrix; H is the transposed matrix of H.

[0040] Furthermore, step S43 verifies the consistency index and uses the contrast loss function to evaluate the prediction consistency. Specifically, torch.nn.CosineEmbeddingLoss of PyTorch is used to evaluate the consistency between features. The specific calculation formula is: in, is the loss function for verifying consistency; n is the number of samples; S ij is the similarity matrix of samples i and j; ||H i -H j || 2 is the square of the distance between samples i and j in the embedding space.

[0041] Further, step S44, multi-scenario verification, verifies the model in different scenarios, calculates the comprehensive ROC curve area (AUC), specifically uses Python's sklearn.metrics.roc_auc_score to calculate the comprehensive AUC value and evaluate the cross-scenario performance of the model; the specific calculation formula is: Among them, AUC combined is the area under the comprehensive ROC curve AUC value; K is the total number of verification scenarios; AUC k is the AUC value in the kth scenario.

[0042] Further, as attached Figure 5 As shown, the risk classification and dynamic cut-off point optimization described in step S5 include the following steps: Step S51 is to construct a classification loss function, define a loss function combining sensitivity and specificity, so as to dynamically adjust the classification threshold, specifically a custom loss function combining sensitivity and specificity, and calculate it through a Python optimizer; the specific calculation formula is: L(τ)=λ1·(1-TPR(τ))+λ2·(1-TNR(τ)); Among them, L(τ) is the classification loss function; τ is the classification threshold; λ1, λ2 are weight parameters; TPR(τ) is the true positive rate; TNR(τ) is the true negative rate.

[0043] Further, step S52 is a dynamic cutoff point optimization, which uses the minimum value of the loss function to determine the optimal classification threshold, specifically using the scipy.optimize.minimize method in Python to optimize the classification threshold to ensure the convergence of the optimal point; the specific calculation formula is: Among them, τ * is the optimal classification threshold; L(τ) is the classification loss function; τ is the classification threshold; TPR(τ) is the true positive rate; TNR(τ) is the true negative rate.

[0044] Further, step S53 is to adjust the risk probability, and recalculate the adjusted risk probability based on the optimized classification threshold, specifically recalculating the risk probability according to the dynamic threshold, and using PyTorch to implement probability mapping; the specific calculation formula is: Among them, P adjusted is the adjusted risk probability; TPR(τ * ) is the optimized threshold τ * True positive rate under * ) is the optimized threshold τ * The true negative rate.

[0045] Furthermore, step S54 introduces regularized classification optimization. On the basis of the dynamic cutoff point, a regularized loss function is introduced to prevent the model from overfitting. An L2 regularization term is introduced into the classification loss, and the regularization parameter is set to δ=0.01. The specific calculation formula is: in, Total loss function optimized for classification; is the binary cross entropy loss; is the L2 norm square of the model weight; γ is the weight coefficient of the cross entropy loss; δ is the weight coefficient of the regularization term.

Claims

1. A method for predicting swallowing aspiration risk, characterized in that: The following steps are involved: S1. Data collection: Collect clinical data of patients and build a database containing key indicators; S2. Variable screening and evaluation: extracting features related to aspiration through standardized assessment scales; S3. Model construction: Construct an aspiration risk prediction model based on logistic regression and convolutional neural network technology; S4. Model validation and optimization: The robustness of the model is verified through multi-dimensional embedding and high-dimensional projection; S5. Risk classification and dynamic cut-off point optimization: Adjust the classification threshold based on a complex optimization algorithm based on sensitivity and specificity; S6. Data prediction: Input real-time clinical data into the optimized classification detection model to predict the risk of swallowing and aspiration.

2. A swallowing aspiration risk prediction method according to claim 1, characterized in that: The data collection described in step S1 includes the following contents: General information includes the patient's age, gender, length of hospital stay, and underlying diseases; clinical indicators include dysphagia score, cough reflex score, NIHSS score, and consciousness disorder score; Physiological characteristics included oral hygiene scores, nutrition methods, and mechanical ventilation status; Extracting features related to aspiration using a standardized assessment scale as described in step S2 includes the following steps: First, the continuous variables are normalized. The specific calculation formula is: Among them, x norm is the normalized variable value; x is the original variable value; μ x is the mean of the original variable; σ x is the standard deviation of the original variable; Then the missing data is interpolated. The specific calculation formula is: Among them, x imp is the variable value after interpolation; X is the set of other variables related to the missing variable; It is the conditional expectation, which means the expected value of variable x when other related variables X are known.

3. A swallowing aspiration risk prediction method according to claim 1, characterized in that: The model building process described in step S3 includes the following steps: S31. Logistic regression model construction. Logistic regression was used to screen the independent risk factors for aspiration. The specific calculation formula was: Among them, Logit(P) is the logarithmic probability of the target variable P; P is the probability of swallowing aspiration, p is the total number of characteristic variables; X i is the i-th characteristic variable; β i is the model regression coefficient; S32, convolutional neural network feature extraction, using convolutional neural network for nonlinear feature extraction, using multi-layer convolution and pooling operations, the specific calculation formula is: in, is the output feature value of the (i, j)th position after the lth convolution operation; l is the sequence number of the neural network convolution layer; k is the convolution kernel size; is the weight parameter of the l-th convolution kernel at the (p,q)th position; is the input feature matrix, i.e. the output of the previous layer; b (l) is the bias term of the lth layer, indicating the offset adjustment value of the convolution kernel; σ is the activation function; S33, the fully connected layer fusion operation, the Logistic regression output and the CNN feature are fused through the fully connected layer to obtain the final feature information, and the final aspiration risk probability is output. The specific calculation formula is: P final =Softmax(W f ·[P Logistic ,P CNN ]+b f ); Among them, P final is the final predicted risk probability, which is a multi-category probability distribution after Softmax processing; Softmax(x) is the normalization function; W f is the weight matrix of the fully connected layer; P Logistic is the aspiration risk probability output by the Logistic regression model; P CNN b is the feature probability output by the convolutional neural network model; f is the bias term of the fully connected layer; S34, multi-task learning expansion stage, in order to simultaneously predict other related complications, a multi-task loss function is introduced, and the specific calculation formula is: in, is the total loss function of multi-task learning; M is the total number of tasks; α j is the weight of the jth task; is the cross entropy loss function; y j is the true label of the jth task; is the predicted value of the jth task.

4. A method for predicting swallowing aspiration risk according to claim 1, characterized in that: The model verification and optimization described in step S4 includes the following steps: S41, feature embedding and dimensionality reduction, using neural network to embed high-dimensional features into low-dimensional space. The specific calculation formula is: H=φ(X·W e +b e ); Among them, H is the low-dimensional embedding feature matrix; X is the high-dimensional feature matrix of the input; W e is the weight matrix of the embedding layer; b e is the bias term of the embedding layer; φ is the activation function; S42, projection space verification, verify the consistency of feature distribution through orthogonal projection, the specific calculation formula is: Among them, T is the interaction matrix of verification features; H is the low-dimensional embedding feature matrix; Q is a symmetric positive definite matrix; is the transposed matrix of H; S43, verify the consistency index, and use the contrast loss function to evaluate the prediction consistency. The specific calculation formula is: in, is the loss function for verifying consistency; n is the number of samples; S ij is the similarity matrix of samples i and j; ||H i -H j || 2 is the square of the distance between samples i and j in the embedding space; S44, multi-scenario verification, verify the model in different scenarios, calculate the comprehensive ROC curve area (AUC), the specific calculation formula is: Among them, AUC combined is the area under the comprehensive ROC curve AUC value; K is the total number of verification scenarios; AUC k is the AUC value in the kth scenario.

5. The method for predicting swallowing aspiration risk according to claim 1, characterized in that: The risk classification and dynamic cutoff point optimization described in step S5 includes the following steps: S51. Construct a classification loss function and define a loss function combining sensitivity and specificity to dynamically adjust the classification threshold. The specific calculation formula is: L(τ)=λ1·(1-TPR(τ))+λ2·(1-TNR(τ)); Among them, L(τ) is the classification loss function; τ is the classification threshold; λ1, λ2 are weight parameters; TPR(τ) is the true positive rate; TNR(τ) is the true negative rate; S52, dynamic cut-off point optimization, using the minimum value of the loss function to determine the optimal classification threshold, the specific calculation formula is: Among them, τ * is the optimal classification threshold; L(τ) is the classification loss function; τ is the classification threshold; TPR(τ) is the true positive rate; TNR(τ) is the true negative rate; S53, adjusting the risk probability, and recalculating the adjusted risk probability based on the optimized classification threshold. The specific calculation formula is: Among them, P adjusted is the adjusted risk probability; TPR(τ * ) is the optimized threshold τ * True positive rate under * ) is the optimized threshold τ * True negative rate under S54, introduce regularized classification optimization. Based on the dynamic cutoff point, introduce a regularized loss function to prevent the model from overfitting. The specific calculation formula is: in, Total loss function optimized for classification; is the binary cross entropy loss; is the L2 norm square of the model weight; γ is the weight coefficient of the cross entropy loss; δ is the weight coefficient of the regularization term.

Images