Sleep apnea curative effect prediction method based on upper airway muscle group training

Abstract

The invention belongs to the field of curative effect prediction, and particularly relates to a sleep apnea curative effect prediction method based on upper airway muscle group training. Comprising the following steps: 1) respectively collecting pressure data, electrical stimulation data and myoelectricity activity data generated during training of an upper airway muscle group, and preprocessing the data to construct a training data set; 2) extracting feature information after each time of training is finished through a multi-modal feature extractor; 3) constructing a curative effect prediction model based on a graph neural network, taking the extracted feature information as input, and training the model; 4) optimizing the curative effect prediction model by using a verification set in the training data set; and 5) inputting test set data in the training data set into the optimized curative effect prediction model to obtain a current treatment state, a later-stage curative effect and parameter adjustment in a treatment process. According to the method, multi-modal data features can be fused, deep understanding and efficient integration of complex data in different modals are realized, and the generalization ability of the model is improved.

Description

Technical Field

[0001] The invention belongs to the field of therapeutic effect prediction, and in particular is a method for predicting the therapeutic effect of sleep apnea based on upper airway muscle group training. Background Art

[0002] Upper airway muscle training is a treatment method that performs specific exercises on the respiratory muscles to restore damaged muscle function.

[0003] Although upper airway muscle training is more acceptable than CPAP and other treatments, there are requirements for training time, and the treatment effect of training on sleep apnea is not intuitive, and long-term compliance is difficult to ensure, which affects the efficacy and limits the clinical application of upper airway muscle training. If the data during the training process is collected and analyzed, combined with the results of polysomnography, a sleep apnea efficacy prediction model based on upper airway muscle training is established, it can effectively feedback the current treatment status and predict the later efficacy, so that the trainees can feel the treatment effect more intuitively and actively. Doctors can also optimize the treatment mode according to the above model, reduce training time, improve training efficiency, and improve patient compliance while ensuring efficacy. Summary of the invention

[0004] The present invention provides a method for predicting the efficacy of sleep apnea based on upper airway muscle group training. Based on the biomechanical characteristics of the upper airway muscle group and related training data, a prediction model for the efficacy of submental electrical stimulation combined with upper airway muscle group training is established through machine learning. The efficacy data and myoelectric activity during the training combined with electrical stimulation are synchronously and dynamically recorded. The efficacy prediction model can effectively feedback the current treatment status and predict the later efficacy, so that the trainee can feel the treatment effect more intuitively and actively, and at the same time, timely feedback is given to the doctor to optimize the treatment plan, so as to achieve long-term follow-up of the treatment of OSA patients.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0006] A method for predicting the efficacy of sleep apnea based on upper airway muscle group training, comprising the following steps:

[0007] 1) Collect pressure data, electrical stimulation data, and electromyographic activity data generated during upper airway muscle group training, preprocess them, and construct a training data set;

[0008] 2) Extract feature information after each training through a multimodal feature extractor;

[0009] 3) Build a therapeutic effect prediction model based on graph neural network, use the extracted feature information as input, and train the model;

[0010] 4) Use the validation set in the training data set to optimize the efficacy prediction model;

[0011] 5) Input the test set data in the training data set into the optimized efficacy prediction model to obtain the current treatment status, later efficacy, and parameter adjustment during the treatment process.

[0012] The step 1) comprises the following steps:

[0013] 1.1) Integrate the pressure data, electrical stimulation data and electromyographic activity data, remove outliers, supplement missing values ​​using interpolation, and normalize them as training data;

[0014] 1.2) The training data is divided into different time periods, and the latter time period is used as the label of the previous time period to predict the later efficacy;

[0015] 1.3) Collecting polysomnography results and processing them into time-segment input data, using the polysomnography results as labels for the current treatment status, and the expert's adjustment of parameters during data acquisition as labels for strategy adjustments in the later diagnosis and treatment process;

[0016] 1.4) Shuffle all data and divide them into training set, validation set and test set.

[0017] The multimodal feature extractor includes three channels, which respectively extract pressure data, electrical stimulation data and electromyographic activity data. Each channel consists of a unimodal feature extractor and an adaptive attention mechanism. The unimodal feature extractor fuses an adaptive residual convolutional network and a multi-head long short-term memory network to extract local features and global dependencies of the data, and calculates the attention score through a multi-layer perceptron in the adaptive attention mechanism.

[0018] The unimodal feature extractor f(x) is specifically:

[0019]

[0020] Attention(x)=IVF(MLP(x))⊙x

[0021] Among them, x is the input data, Attention is the adaptive attention mechanism, ResConv is the residual convolution layer, Multi_LSTM is the multi-head long short-term memory network, is the feature concatenation operation, LSTM is the long short-term memory network, and w Critic It is a weight assignment method based on contrast intensity and conflict, measured by the Pearson correlation coefficient, IVF is the information weight, MLP is the multilayer perceptron, and ⊙ is the element-by-element multiplication.

[0022] The step 3) is specifically as follows:

[0023] The feature information of different modalities is used as the model input, and the feature information of different modalities is fused through the topological structure of the graph. The relevance and importance of the feature information of different modalities are captured through the position information attention mechanism to obtain the relative importance of the feature information, where:

[0024]

[0025] Among them, h i ′ is the updated node information, N i is the neighboring node of node i, w is the weight matrix, h j is the feature information of node j, b is the bias matrix, a ij is the attention weight, e ij is the attention score, p ij is the position weight information, σ is the activation function, and k is the number of adjacent nodes.

[0026] In the efficacy prediction model, four types of regression loss functions L are set to measure the losses of the three channels and the graph neural network fusion loss, specifically:

[0027]

[0028] Among them, Loss i For different channel loss functions, EMM (Loss i ) measures the information entropy of each loss, w i is the weight matrix trained using a multilayer perceptron, y j is the model prediction value, is the true value, and n is the number of samples.

[0029] The back propagation and Adam optimization algorithms were used to optimize the efficacy prediction model.

[0030] The present invention has the following beneficial effects and advantages:

[0031] 1. The present invention constructs an auxiliary efficacy prediction platform for structured, standardized collection and prediction of full-cycle data parameters for tracking individual patients before treatment, during training, and after treatment, thereby achieving the connection between training evaluation and clinical decision management.

[0032] 2. The present invention can integrate multimodal data features, achieve in-depth understanding and efficient integration of complex data under different modalities, and improve the generalization ability of the model.

[0033] 3. The present invention introduces graph neural network and position attention mechanism to perform feature fusion and extract topological information, making the model more interpretable. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A diagram showing a method for preparing a data set for a therapeutic efficacy prediction model of the present invention;

[0035] Figure 2 Schematic diagram of the efficacy prediction model of the present invention. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0037] A data-driven efficacy prediction model for submental electrical stimulation combined with upper airway muscle training includes the following steps:

[0038] Step 1: If Figure 1 As shown, it is a schematic diagram of the data set construction of the present invention. According to the biomechanical characteristics of the upper airway muscles and related training data, the corresponding pressure data, efficacy data and electromyographic activity data are integrated, the outliers are eliminated, the missing values ​​are supplemented by interpolation, and normalization is performed. The input data is divided into different time periods, and the latter time period is used as the label of the previous time period to predict the later efficacy. The polysomnography results are used as the current treatment status label, and the expert adjustment of the parameters during the data acquisition process is used as a label to obtain the efficacy prediction model data set, which is divided into a training set, a validation set and a test set in a ratio of 8:1:1;

[0039] Step 2: Use the training set data as model input to train the model, use the grid search method to optimize the hyperparameters for the validation set data, select the training model with the highest prediction accuracy as the final prediction model, and fix the model;

[0040] Step 3: Perform a performance test on the final model obtained in step 2 on the test set data.

[0041] like Figure 2 As shown, the structure diagram of the efficacy prediction model based on data-driven submental electrical stimulation combined with upper airway muscle training.

[0042] The multimodal feature extractor constructed as described is: the present invention uses residual convolutional neural network and long short-term memory network as the main feature extractors. The residual convolutional neural network alleviates the gradient vanishing problem in the training process by introducing residual blocks and jump connections. The stacked convolutional layers and pooling operations can effectively capture the local features of the input data from low-level to high-level. The long short-term memory network can effectively capture the long-term dependencies in the sequence data, aggregate the extracted local feature information, and the adaptive attention mechanism enables the model to focus on important feature information, reduce the redundancy of features, and effectively model global features.

[0043] The multimodal feature extractor constructed as described above includes three channels, each of which consists of a unimodal feature extractor and an adaptive attention mechanism:

[0044] The unimodal feature extractor integrates adaptive residual convolutional networks and multi-head long short-term memory networks to extract local features and global dependencies of data;

[0045]

[0046] Among them, x is the input data, Attention is the adaptive attention mechanism, ResConv is the residual convolution layer, Multi_LSTM is the multi-head long short-term memory network, is the feature concatenation operation, LSTM is the long short-term memory network, and w Critic It is a weight assignment method based on contrast intensity and conflict, measured by Pearson correlation coefficient.

[0047] The output of the unimodal feature extractor passes through an adaptive attention mechanism and calculates the attention score through a multi-layer perceptron to enhance the model's ability to express multimodal input data;

[0048] Attention(x)=IVF(MLP(x))⊙x

[0049] Among them, IVF is the information weight, and the weight is evaluated according to the amount of information carried by the feature. MLP is a multi-layer perceptron, and ⊙ is element-by-element multiplication.

[0050] The therapeutic efficacy prediction model constructed by the graph neural network is as follows: the present invention takes into account the fusion of data features of different modalities, uses graph neural networks to construct modal feature graphs for information dissemination and fusion, captures the correlation information between different modalities, makes full use of the complementarity between different modalities, better understands the complex interactions and dependencies between different modalities, improves the accuracy of prediction, and enhances the robustness and generalization ability of the model. Position weight information is introduced into the attention mechanism to capture direction and position sensitive information, dynamically adjust the attention to different parts, effectively capture and utilize long-distance information, enhance the representation ability of the model, and improve the interpretability of the model.

[0051] In the therapeutic efficacy prediction model of the constructed graph neural network, the extracted feature information of different modalities is used as input, the feature information of different modalities is fused through the topological structure of the graph, and the position information attention mechanism is introduced to capture the relevance and importance of different modal features. The feature information of different modalities is aggregated, the direction and position sensitive information is captured, and the representation ability of the model is enhanced.

[0052]

[0053] Among them, h i ′ is the updated node information, N i is the neighboring node of node i, w is the weight matrix, h jis the feature information of node j, b is the bias matrix, a ij is the attention weight, e ij is the attention score, p ij is the position weight information.

[0054] In the efficacy prediction model of constructing a graph neural network, four types of regression loss functions are set to measure the losses of the three channels and the graph neural network fusion loss;

[0055]

[0056] Among them, EMM (Loss i ) measures the information entropy of each loss, w i is the weight matrix trained using a multilayer perceptron, y j is the model prediction value, is the true value, and n is the number of samples.

[0057] The global parameters in the process of training the model are set as follows: learning rate is 0.0005, batch size is 128, and the number of training times is 10,000.

[0058] The efficacy prediction model of subchin electrical stimulation combined with upper airway muscle training based on data-driven elimination in the present invention consists of two parts: a multimodal feature extractor and a multimodal feature fusion based on a graph neural network; the multimodal feature extractor extracts single-modal feature information by fusing a residual convolutional network, a long short-term memory network and an adaptive attention mechanism, models the global dependency of single-modal data, and reduces the redundancy of features; the efficacy prediction model based on a graph neural network utilizes the topological structure of the graph to capture the correlation information and complementarity between different modalities, updates the representation of nodes through information transfer and aggregation between nodes, better understands the complex interactions and dependencies between different modalities, performs multimodal feature fusion, and improves the prediction performance of the model, so that the model can effectively feedback the current treatment status and predict the later efficacy, thereby improving patient compliance.

[0059] The specific steps are as follows:

[0060] Step 1: Data acquisition and preprocessing.

[0061] According to the biomechanical characteristics of the upper airway muscles and related training data, the corresponding pressure data, efficacy data and electromyographic activity data are integrated and normalized. The input data is divided according to the treatment course to make labels for predicting the efficacy of different treatment courses. The polysomnography results are used as the treatment status label of the current treatment course, and the adjustment of the parameters by the experts during the treatment process is used as the feedback label to obtain the efficacy prediction model data set, which is divided into training set, validation set and test set in a ratio of 8:1:1;

[0062] Step 2: Perform unimodal feature extraction through the optimized feature extractor.

[0063] Taking into account that the collected data belongs to different modalities, the present invention utilizes residual convolutional neural networks, long short-term memory networks and adaptive attention mechanisms to extract local features and global dependencies of input under a single modality. Local features at different depths are extracted by stacking residual convolutional neural networks, global dependencies are modeled by long short-term memory networks, and the adaptive attention mechanism enables the model to pay more attention to important features and reduce feature redundancy.

[0064] Step 3: Predict through the optimized graph neural network-based efficacy prediction model.

[0065] During the training process, in order to ensure the effective fusion of data features of different modalities, the present invention uses the topological structure of the graph to capture the correlation information and complementarity between different modalities, and fuses data features under different modalities by converging the features of nodes and adjacent nodes; introduces an attention mechanism with position information to capture direction and position sensitive information, and dynamically adjusts the attention of different modalities. In view of the loss of feature information that occurs when features of different modalities are fused, the output of each channel is spliced, the loss function is calculated, and the weight of the loss function is dynamically adjusted according to the importance of different channels, the model parameters are effectively updated, and the final prediction result is obtained after multiple rounds of iterative updates of feature information.

[0066] Step 4: Loss function setting.

[0067] The present invention sets four types of regression loss functions to measure the losses of three channels and the graph neural network fusion loss;

[0068]

[0069] Among them, EMM (Loss i ) measures the information entropy of each loss, w i is the weight matrix trained using a multilayer perceptron, y j is the model prediction value, is the true value, and n is the number of samples.

[0070] Step 5: Optimize policy settings.

[0071] The back propagation and Adam optimization algorithms are used to optimize the efficacy prediction model of the graph neural network, that is, to minimize the total loss function.

[0072] Step 6: Model performance test.

[0073] The grid search method is used to find the optimal hyper-hyperparameter combination, and the training model is fixed under the optimal hyper-hyperparameter combination. The efficacy prediction is performed on the test set to test the model performance.

Claims

1. A method for predicting the efficacy of sleep apnea based on upper airway muscle training, characterized in that: The following steps are involved: 1) Collect pressure data, electrical stimulation data, and electromyographic activity data generated during upper airway muscle group training, preprocess them, and construct a training data set; 2) Extract feature information after each training through a multimodal feature extractor; 3) Build a therapeutic effect prediction model based on graph neural network, use the extracted feature information as input, and train the model; 4) Use the validation set in the training data set to optimize the efficacy prediction model; 5) Input the test set data in the training data set into the optimized efficacy prediction model to obtain the current treatment status, later efficacy, and parameter adjustment during the treatment process.

2. The method for predicting the efficacy of sleep apnea based on upper airway muscle training according to claim 1, characterized in that: The step 1) comprises the following steps: 1.1) Integrate the pressure data, electrical stimulation data and electromyographic activity data, remove outliers, supplement missing values ​​using interpolation, and normalize them as training data; 1.2) The training data is divided into different time periods, and the latter time period is used as the label of the previous time period to predict the later efficacy; 1.3) Collecting polysomnography results and processing them into time-segment input data, using the polysomnography results as labels for the current treatment status, and the expert's adjustment of parameters during data acquisition as labels for strategy adjustments in the later diagnosis and treatment process; 1.4) Shuffle all data and divide them into training set, validation set and test set.

3. The method for predicting the efficacy of sleep apnea based on upper airway muscle training according to claim 1, characterized in that: The multimodal feature extractor includes three channels, which respectively extract pressure data, electrical stimulation data and electromyographic activity data. Each channel consists of a unimodal feature extractor and an adaptive attention mechanism. The unimodal feature extractor fuses an adaptive residual convolutional network and a multi-head long short-term memory network to extract local features and global dependencies of the data, and calculates the attention score through a multi-layer perceptron in the adaptive attention mechanism.

4. The method for predicting the efficacy of sleep apnea based on upper airway muscle training according to claim 3, characterized in that: The unimodal feature extractor f(x) is specifically: Attention(x)=IVF(MLP(x))⊙x Among them, x is the input data, Attention is the adaptive attention mechanism, ResConv is the residual convolution layer, Multi_LSTM is the multi-head long short-term memory network, is the feature concatenation operation, LSTM is the long short-term memory network, and w Critic It is a weight assignment method based on contrast intensity and conflict, measured by the Pearson correlation coefficient, IVF is the information weight, MLP is the multilayer perceptron, and ⊙ is the element-by-element multiplication.

5. The method for predicting the efficacy of sleep apnea based on upper airway muscle training according to claim 1, characterized in that: The step 3) is specifically as follows: The feature information of different modalities is used as the model input, and the feature information of different modalities is fused through the topological structure of the graph. The relevance and importance of the feature information of different modalities are captured through the position information attention mechanism to obtain the relative importance of the feature information, where: Among them, h i ′ is the updated node information, N i is the neighboring node of node i, w is the weight matrix, h j is the feature information of node j, b is the bias matrix, a ij is the attention weight, e ij is the attention score, p ij is the position weight information, σ is the activation function, and k is the number of adjacent nodes.

6. The method for predicting the efficacy of sleep apnea based on upper airway muscle training according to claim 5, characterized in that: In the efficacy prediction model, four types of regression loss functions L are set to measure the losses of the three channels and the graph neural network fusion loss, specifically: Among them, Loss i For different channel loss functions, EMM (Loss i ) measures the information entropy of each loss, w i is the weight matrix trained using a multilayer perceptron, y j is the model prediction value, is the true value, and n is the number of samples.

7. The method for predicting the efficacy of sleep apnea based on upper airway muscle training according to claim 1, characterized in that: The back propagation and Adam optimization algorithms were used to optimize the efficacy prediction model.

Images