Self-adaptive control method for jet frequency of high-frequency jet breathing machine

Through the adaptive control method of high-frequency jet ventilator jet frequency, the injection frequency is dynamically adjusted by using multi-parameter real-time acquisition and multi-modal fusion model, which solves the carbon dioxide retention problem caused by excessive ventilation frequency of high-frequency jet ventilator, and improves gas Exchange efficiency and oxygenation effect.

CN118987425APending Publication Date: 2024-11-22LIDE HEALTH TECH (SHANDONG) CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411277270.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-12
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

High-frequency jet ventilators cause carbon dioxide retention when the ventilation frequency is too high, affecting the normal metabolism of cells and gas exchange, causing harm.

Method used

Adaptive control method for jet frequency of high-frequency jet ventilator is adopted, and parameters such as airway carbon dioxide concentration, oxygen concentration, blood oxygen saturation, heart rate are comprehensively collected through multi-parameter real-time acquisition system, multi-modal fusion model and Internet jet frequency expert system. , dynamically adjust the injection frequency to form the optimal injection frequency and adaptive adjustment scheme.

Benefits of technology

It has achieved the reduction of carbon dioxide retention, improved gas exchange efficiency, and reduced harm to patients while ensuring the oxygenation effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118987425A_ABST
    Figure CN118987425A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of medical instruments, in particular to a high-frequency jet breathing machine jet frequency self-adaptive control method which comprises the following steps: S1) establishing a multi-parameter real-time acquisition system: based on a high-frequency breathing machine, acquiring parameters including but not limited to airway carbon dioxide concentration, oxygen concentration, oxyhemoglobin saturation and heart rate; s2) performing two-dimensional parameter coupling on the output frequency and the output breathing time ratio; s3) extracting and preprocessing the parameters collected in the step S1, screening data to construct a multi-parent-fetal data set, and constructing and training a multi-modal fusion model; and S4) applying the multi-modal fusion model in the step S3 to form the Internet injection frequency expert system. According to the method, parameter characteristics of the high-frequency breathing machine are integrated, appropriate data are collected, a data set is established, and the optimal injection frequency and the self-adaptive adjustment scheme of the injection frequency are obtained through calculation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of medical devices, in particular to a method for adaptively controlling the jet frequency of a high-frequency jet ventilator. Background Art

[0002] High-frequency jet ventilator is a respiratory device that uses jet technology. Different from normal-frequency ventilator, its ventilation frequency can reach a very high level. High-frequency ventilator can achieve a high number of breaths per minute to improve oxygenation and carbon dioxide excretion and improve blood gas status. This ventilator is designed to provide high-frequency intermittent jet ventilation to improve the blood pressure fluctuation problem that may be caused by traditional positive pressure artificial ventilation. High-frequency jet ventilator sprays high-pressure oxygen into the airway through a tiny spray needle, and with the help of the principle of Venturi effect, the air at the spray needle mouth is sucked into the airway, thereby achieving high-frequency ventilation.

[0003] The jet frequency range of high-frequency jet ventilators is relatively large, generally ranging from 60 times / minute to 600 times / minute. Such high-frequency ventilation helps improve the patient's oxygen supply, especially when oxygenation needs to be improved quickly. The application of this type of ventilator has not only been extensively studied in animal experiments, but has also shown its unique advantages in clinical practice, especially when it is necessary to quickly adjust the patient's oxygen supply. In addition, the design of high-frequency jet ventilators also takes into account portability and efficiency, which makes this device also have a wide range of application prospects in emergency medical rescue and field first aid.

[0004] Traditional high-frequency jet ventilators deliver oxygen to patients at a frequency of 60 to 200 times per minute, which is suitable for ventilation of patients with open lung trauma and severe lung leakage. However, excessively high-frequency oxygenation will lead to carbon dioxide retention, affect normal cell metabolism and gas exchange, and produce a series of hazards. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a method for adaptively controlling the jet frequency of a high-frequency jet ventilator, which integrates specific indication parameters to obtain the optimal jet frequency and its adaptive adjustment scheme, and realizes the variable frequency mode. The technical solution adopted by the present invention is as follows: A method for adaptively controlling the jet frequency of a high-frequency jet ventilator comprises the following steps: S1) Establishing a multi-parameter real-time acquisition system: Based on a high-frequency ventilator, the acquisition parameters include but are not limited to airway carbon dioxide concentration, oxygen concentration, blood oxygen saturation, and heart rate; S2) performing two-dimensional parameter coupling of the output frequency and the output breathing time ratio; S3) extracting and preprocessing the parameters collected in step S1, screening the data to construct a multi-maternal-fetal data set, and building and training a multimodal fusion model; S4) Apply the multimodal fusion model in step S3 to form an Internet injection frequency expert system.

[0006] The method for constructing a multimodal dataset is as follows: S2.1) The output parameters of the high-frequency ventilator are classified into {J, K} coordinates for retrieval in a two-dimensional graph, where J is the highest injection frequency and K is the time proportion of the highest injection frequency in the entire injection cycle; S2.2) Training data acquisition: Using the patient's medical records as the data source, record the hospitalization information of each patient in each medical record, and extract the high-frequency ventilator parameters of these patients to construct a multimodal dataset; S2.3) The multimodal fusion model consists of four parts: the first part is the feature extractor F1, which is composed of a deep convolutional neural network, thereby realizing the detailed feature extraction of the data; the second part is the report form and biochemical examination data feature extractor F2, which is composed of an artificial neural network and realizes the feature extraction of the report form and biochemical examination information; the third part is the feature fusion device F, which is used to fuse the feature vectors obtained from the feature extractor F1 and the feature extractor F2, and use the fused feature vector as the input of the fourth part; the fourth part is the classifier C, which is a multi-layer perceptron composed of a Dropout layer and three hidden layers, and finally realizes the diagnostic classification of the output parameters.

[0007] In the above step S2.2, the data is normalized and Min-Max normalization is used to complete the data normalization work. The specific formula is: ; In the formula, x i For the current need to standardize the value, min 1≤j≤n(xj) for x1, x2, …, x n The minimum value in max 1≤j≤n(xj) for x1, x2, …, x n The maximum value in y i For x i The standardized values ​​for x1, x2, ..., x n After standardization, we get y1, y2, ..., y n And y1, y2, …, y n ∈[0,1].

[0008] The random forest model is used to calculate the information gain of each attribute of the data. First, the information entropy is calculated: ; In the formula, p(c) represents the proportion of samples of category c in the data set; Then calculate the conditional entropy of data set D under feature F: ; In the formula, |D i | represents subset D i , |D| represents the number of samples in the data set D; Finally, the information gain of feature F is calculated, which indicates the degree to which the uncertainty of data set D is reduced after splitting using feature F. The information gain is calculated as follows: ; In the formula, Informationgain (D, F) represents the information gain of feature F in data set D, entropy(D) represents the initial information entropy of the data set D. conditionalentropy(D,F) represents the conditional entropy of data set D under feature F.

[0009] The above step S2.3 includes the following steps: S2.3.1) Feature extractor F1 The data matrix i1 is taken as input, the feature vector h1 after extracting the data features is taken as the output of this part, and the feature vector h1 is taken as the input of the feature fusion device F; First, the input feature vector h1 passes through a normal convolution layer, then an activation function layer, and finally a batch normalization layer to obtain the feature map i′: ; In the formula, h is the output channel of the convolutional layer, cin is the input channel of the convolutional layer, p is the kernel size and step size, BN is the batch normalization layer, and the specific calculation method is: ; In the formula, û B is the sample mean, σ B is the sample standard deviation of the mini-batch B, γ is the stretch parameter, β is the offset parameter, and the shapes of γ and β are the same as λ; The feature map i′1 after the batch normalization layer is input into N ConvMixer blocks consisting of depthwise convolutional layers and pointwise convolutional layers. The specific calculation method is: ; ; In the formula, ConvDepthwise() is the depth convolution layer, ConvDepthwise () is the point-by-point convolution layer, and the specific calculation methods are: ; In the formula, O k,i,j Represents the value of the kth channel of the output feature map O at position (i, j), I c,i+p-1,j+p-1 represents the value of the cth channel of the input feature map I at position (i+p-1,j+p-1), W k,c,p,q represents the weight of the cth channel of the kth convolution kernel at position (p, q), b k Represents the bias term of the kth convolution kernel; ; In the formula, ω i,j represents the weight value between the i-th output channel and the j-th input channel of the convolution kernel, x j Represents the value of the input tensor on the jth channel, y i Represents the value of the output tensor on the i-th channel; The feature map i after passing through N ConvMixer blocks` l+1 Input to the adaptive average pooling layer and a Dropout layer to output the feature vector. The specific calculation method is: ; In the formula, i` l+1 As the feature map output by the ConvMixer block and as the input of the Dropout layer, the probability of the Dropout operation is p, and h1 is the feature vector of the output of this part.

[0010] The above step S2.3 further includes the following steps: S2.3.2) Feature extractor F2 The report form and biochemical test data are converted into vector i2 as the input of the feature extractor. After being input into the network structure, it first passes through two Linear layers, then a Dropout layer, and finally a Linear layer to obtain the feature vector h2 as the output of this part; the value of the neuron is calculated through the following process: ① Weighted sum: Each neuron has a set of weights multiplied by the input, plus a bias term; Activation function: The net input is passed into an activation function to generate the output of the neuron; The specific calculation method of the artificial neural network is: ; Where i2 is the biochemical examination data and echocardiography report form as the input of this part, W is the weight, b is the bias, f is the activation function, and h2 is the output feature vector.

[0011] The above step S2.3 further includes the following steps: S2.3.3) Feature Fusion and Classifier The feature fusion device takes the feature vector h1 output by the feature extractor F1 and the feature vector h2 output by the report form and biochemical test data feature extractor F2 as input, horizontally concatenates the two feature vectors to generate a new feature vector p, and inputs the fused feature vector into the classifier. After passing through the multi-layer perceptron, the prediction result y is finally output; The horizontal stitching method is used for feature stitching. The specific calculation method is: ; Where, h1 is the feature vector i1 output by the feature extractor F1, h2 is the feature vector i2 output by the report form and biochemical examination data feature extractor F2, p is the fused feature, and f is the mapping relationship; The input layer contains n neurons, where the size of n is equal to the feature vector p output by the feature fuser. Given the input p, the predicted value y can be calculated through forward propagation by using uniformly distributed random initialization weights and biases. The specific calculation method is: ; Where W o is the classifier model weight, b o is the classifier model bias, y is the output result of the classifier, the softmax function can map the output value to the range of (0,1), and the sum of the probability values ​​of all outputs is 1.

[0012] The above step S2.3 further includes the following steps: S2.3.4) Loss Function and Optimizer The specific calculation method of the cross entropy loss function is: ; In the formula, N represents the total number of samples, C represents the number of categories, and y i,c Indicates that sample i belongs to the true label of category C, p i,c represents the predicted probability that sample i belongs to category C; The optimizer estimates the gradient of the loss function by randomly selecting a training sample in each iteration and uses the gradient to update the model parameters. The specific calculation method is: ; In the formula, θ t represents the parameter at iteration t, J(θ t ;x (i) ,y (i) ) is the loss function, x (i) and (i)is the input and output of the i-th training sample, η▽ θt J(θ t ;x (i) ,y (i) ) is the loss function with respect to the parameter θ t is the gradient of , and η is the learning rate.

[0013] 9. The method for adaptively controlling the jet frequency of a high-frequency jet ventilator according to claim 2 is characterized in that in the above step S2.2, during the data extraction process, when there are multiple examinations for the same patient, the first diagnosis of each patient in time sequence is selected to construct a data set. The beneficial effects of the present invention are: comprehensively considering the parameter characteristics of the high-frequency ventilator, collecting appropriate data and establishing a data set, obtaining the optimal jet frequency and the adaptive adjustment scheme of the jet frequency through calculation, and saving them online in the system as experience values ​​to guide similar or similar treatment scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic diagram of the hardware structure of an embodiment of the present invention; Figure 2 A schematic diagram of two-dimensional coupling of output frequency and I / E ratio according to an embodiment of the present invention; Figure 3 A parameter classification result retrieval diagram of an embodiment of the present invention; Figure 4 Schematic diagram of the structure of the feature extractor F1 according to an embodiment of the present invention; Figure 5 A schematic diagram of a network structure of a report table and biochemical examination data according to an embodiment of the present invention; Figure 6 A schematic diagram of the network structure of a classifier according to an embodiment of the present invention; Figure 7 The figure is a schematic diagram of a network topology structure according to an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The technical content of the present invention is described in detail below in conjunction with the accompanying drawings. This embodiment is a method for adaptively controlling the jet frequency of a high-frequency jet ventilator, which specifically includes the following steps: S1) Establish a multi-parameter real-time acquisition system In order to achieve an improvement in oxygenation effect and a maximum reduction in carbon dioxide retention effect, based on an ordinary high-frequency ventilator, this embodiment expands a multi-parameter real-time acquisition system, and the main acquisition parameters are airway carbon dioxide concentration, oxygen concentration, blood oxygen saturation, heart rate and other indicator parameters. By detecting the above parameters, the patient's current state can be known in real time and fed back to the core processor of the high-frequency ventilator for fusion calculation. Devices that can achieve the above parameter acquisition goals include but are not limited to Figure 1As shown. The traditional high-frequency ventilator is upgraded with wired / wireless finger-clip heart rate and blood oxygen saturation meters, internal carbon dioxide and oxygen concentration sensors (or interfaces), and communication modules (Wi-Fi, 4G or other modules) that can access cloud services. It can connect to remote cloud services and upload parameters as expert system records. In addition to being able to manage equipment online (such as geographic location, treatment records, current working status, validity period, etc.), its data can also guide artificial intelligence learning.

[0016] S2) Coupling the output parameters of the high-frequency ventilator The output frequency (times / minute) and the output breathing time ratio (i.e., I / E ratio, 1:1, 1:1.5, 1:2, 1:3, 1:4, etc.) are coupled as two-dimensional parameters. The principle is that for the same working condition, the higher the output frequency, the better the oxygenation effect in theory, but there will be a higher risk of carbon dioxide retention. The higher the output I / E ratio (higher the ratio of inhalation time, 1:4 is defined as the lowest and 1:1 as the highest), the larger the intake volume and the greater the airway pressure. Although it can bring better oxygenation effect, if the frequency is too high, it will irritate or damage the airway. Therefore, this embodiment couples the above parameters in two dimensions, such as Figure 2 Generally speaking, the higher the output frequency, the lower the I / E ratio needs to be to avoid damage caused by increased airway pressure due to excessive frequency. Figure 2 As shown by the slash in the figure, under the same working conditions, the parameters corresponding to the slash have similar working effects. However, although a lower I / E ratio can avoid the above problems and minimize carbon dioxide retention when the output frequency is high, it may affect the oxygenation effect due to the decrease in overall oxygen content. Therefore, under the premise of ensuring better oxygenation effect, it is necessary to adjust the output frequency and I / E ratio according to the dotted curve below the slash to improve the oxygenation effect.

[0017] S3) Build and train a multimodal fusion model A multimodal fusion algorithm for dynamic adjustment of high-frequency injection frequency is used through airway carbon dioxide concentration, oxygen concentration, blood oxygen saturation, heart rate and other parameters, and a multimodal fusion model is built and trained. The model can learn the characteristics of different modal data during the training process to improve the performance of the model when adjusting the injection frequency. In order to train the multimodal fusion model, a multimodal data set is constructed, and parameters such as airway carbon dioxide concentration, oxygen concentration, blood oxygen saturation, heart rate, etc. are extracted and preprocessed from the original data file, and then the data required for this embodiment is screened out to complete the construction of the multimodal data set. Based on the above multimodal data, a multimodal fusion model consisting of a feature extractor, a feature fusion device and a classifier is proposed.

[0018] S4) Forming Internet jet frequency expert system The multimodal fusion model and data set construction method in the previous step are applied to realize an Internet jet frequency expert system. The system is built based on the C / S architecture and realizes the function of obtaining real-time data for patients who need high-frequency ventilator treatment. In actual use, users upload sensor data by connecting to various sensors and using this embodiment, and can obtain high-frequency jet data and dynamic adjustment solutions online.

[0019] In the above step S2, the method for constructing the multimodal dataset is: S2.1) Classification of high-frequency ventilator output parameters According to the therapeutic effect of high-frequency ventilator, it is mainly considered under what conditions of patients are suitable for high-frequency ventilator and how to adaptively adjust the injection frequency to meet the treatment needs in different situations. Its output classification is divided into {J, K} coordinates for retrieval in two-dimensional graphs. The two-dimensional graph is as follows Figure 3 shown. Figure 3 The classification can be classified into one category within a small range of parameters. The classification only determines the range of the coordinates of {J, K}, where J is the highest injection frequency and K is the time proportion of the highest injection frequency in the entire injection cycle. The injection cycle can be manually adjusted on the ventilator. The transition of the injection frequency from the highest to the normal one can be done according to the linear method.

[0020] S2.2) Training data acquisition The patient's medical record is used as the data source. Each medical record records each patient's condition, hospitalization, past medical history, examination records, doctor's diagnosis, patient course records, examination result indicators, etc. In order to ensure data balance, it is necessary to evenly screen the medical records of multiple different patients and extract the airway carbon dioxide concentration, oxygen concentration, blood oxygen saturation, heart rate and other data of these patients to construct a multimodal data set.

[0021] In order to ensure the accuracy and reliability of the data, during the data extraction process, when the same patient has multiple examinations, this embodiment selects the first diagnosis of each patient in chronological order to construct the data set of this embodiment. During the data set construction process, the confidentiality and privacy of the data also need special attention. In the data extraction stage, this embodiment desensitizes the personal information in the medical records of each patient.

[0022] In the data set construction process of this embodiment, the patient's report form and biochemical examination data are extracted from the archives and saved in a table file. A unique identification code is assigned to each piece of data, and the corresponding disease type is marked. Then the data is extracted from each patient's medical record file, and "label_identification code" is used as the table file name to complete the data alignment work. Finally, it is divided into a training set, a validation set, and a test set.

[0023] In the data extraction stage, because the units of various examination indicators saved in the patient's medical records are inconsistent, it is necessary to first convert the units of the report form and biochemical examination data to achieve the unification of the units of various examination indicators. Then the data is normalized and the Min-Max standardization is used to complete the data normalization work, thereby eliminating the dimensional differences between different features and improving the convergence speed and stability of the model. The specific formula is: ; In the formula, x i For the current need to standardize the value, min 1≤j≤n(xj) for x1, x2, …, x n The minimum value in max 1≤j≤n(xj) for x1, x2, …, x n The maximum value in y i For x i The standardized values ​​for x1, x2, ..., x n After standardization, we get y1, y2, ..., y n And y1, y2, …, y n ∈[0,1].

[0024] The random forest model is used to calculate the information gain of each attribute. In the process of calculating the information gain, it is necessary to first calculate the initial information entropy of the data set, which represents the degree of chaos in the data set. The calculation method of information entropy is: ; In the formula, p(c) represents the proportion of samples of category c in the data set; Then calculate the conditional entropy of data set D under feature F: ; In the formula, |D i | represents subset D i , |D| represents the number of samples in the data set D; Finally, the information gain of feature F is calculated, which indicates the degree to which the uncertainty of data set D is reduced after splitting using feature F. The information gain is calculated as follows: ; In the formula, Informationgain (D, F) represents the information gain of feature F in data set D, entropy(D) represents the initial information entropy of the data set D. conditionalentropy(D,F) represents the conditional entropy of data set D under feature F. .

[0025] The greater the information gain, the greater the reduction in overall entropy after using the feature for classification, which means that the feature contributes more to the classification task. Since there are 92 attributes in total, the data dimension is high and the information gain variance between attributes is large. It is necessary to perform dimensionality reduction on the attribute data, and finally obtain the attributes with the highest information gain.

[0026] S2.3) Multimodal fusion model and its algorithm The structure of the multimodal fusion model is a core module of the multimodal fusion algorithm. How to build an effective multimodal fusion model structure, how to extract data features and extract report form and biochemical examination data features, and effectively fuse these features are the keys to improving model classification performance.

[0027] The multimodal fusion method used in this embodiment belongs to feature-level fusion, and the data of different modes are extracted and then fused. The data of different modes are respectively passed into the corresponding feature extractors for processing, and their respective feature representations are extracted, and then these features are fused to obtain the final comprehensive feature representation. The multimodal fusion model proposed in this embodiment consists of four parts: the first part is the feature extractor F1, which is composed of a deep convolutional neural network, so as to realize the detailed content of the feature extraction of the data; the second part is the report form and biochemical examination data feature extractor F2, which is composed of an artificial neural network, and realizes the feature extraction of the report form and biochemical examination information; the third part is the feature fusion F, which is used to fuse the feature vectors obtained from the feature extractor F1 and the feature extractor F2, and use the fused feature vector as the input of the fourth part; the fourth part is the classifier C, which is a multilayer perceptron composed of a Dropout layer and three hidden layers, and finally realizes the diagnostic classification of the output parameters. The specific algorithm of the multimodal fusion model proposed in this embodiment is shown in Table 1.

[0028] S2.3.1) Feature extractor F1 The overall structure of the feature extractor F1 is a deep convolutional neural network, which takes the data matrix i1 as input, extracts the feature vector h1 from the data as output, and takes the feature vector h1 as input to the feature fusion device F. The feature extractor F1 in this embodiment has made some improvements based on this model. The network structure of this part is as follows: Figure 4 As shown in the figure, this part of the network structure includes convolution layer (C), depth convolution layer (DW), point-by-point convolution layer (PW), shortcut layer, batch normalization layer (BN), Gaussian error linear unit (GELU), and adaptive average pooling layer (AAP).

[0029] The deep convolutional neural network used in this embodiment extracts the feature information of the image layer by layer through multi-layer convolution and pooling operations. The first layer is the convolution layer, and the input of this layer is the original image. The convolution layer performs a convolution operation on the input image through a set of convolution kernels to obtain a new feature map. The convolution kernel can be regarded as a locally connected weight matrix, which extracts different features of the image by sliding on the input image. A nonlinear activation function is added after the convolution layer. Next, the ConvMixer block composed of a deep convolution layer and a point-by-point convolution layer is used to further extract the features of the image. Finally, an adaptive average pooling layer is added to reduce the size of the feature map and reduce the amount of calculation. Thereby achieving feature extraction.

[0030] First, the input feature vector h1 passes through a normal convolution layer, then an activation function layer, and finally a batch normalization layer to obtain the feature map i′: ; In the formula, h is the output channel of the convolutional layer, cin is the input channel of the convolutional layer, p is the kernel size and step size, BN is the batch normalization layer, and the specific calculation method is: ; In the formula, û B is the sample mean, σ B is the sample standard deviation of the mini-batch B, γ is the stretch parameter, and β is the offset parameter. The shapes of γ and β are the same as λ.

[0031] The feature map i`1 after the batch normalization layer is input into N ConvMixer blocks consisting of depthwise convolutional layers and pointwise convolutional layers. The specific calculation method is: ; ; In the formula, ConvDepthwise() is the depth convolution layer, ConvDepthwise() is the point-by-point convolution layer, and the specific calculation methods are: ; In the formula, O k,i,j Represents the value of the kth channel of the output feature map O at position (i, j), I c,i+p-1,j+p-1 represents the value of the cth channel of the input feature map I at position (i+p-1,j+p-1), W k,c,p,q represents the weight of the cth channel of the kth convolution kernel at position (p, q), b k Represents the bias term of the kth convolution kernel; ; In the formula, ω i,jrepresents the weight value between the i-th output channel and the j-th input channel of the convolution kernel, x j Represents the value of the input tensor on the jth channel, y i Represents the value of the output tensor on the i-th channel; The feature map i after passing through N ConvMixer blocks` l+1 Input to the adaptive average pooling layer and a Dropout layer to output the feature vector. The specific calculation method is: ; In the formula, i` l+1 As the feature map output by the ConvMixer block and as the input of the Dropout layer, the probability of the Dropout operation is p, and h1 is the feature vector of the output of this part.

[0032] S2.3.2) Feature extractor F2 The overall structure of the report form and biochemical test data feature extractor is an artificial neural network. This part converts the report form and biochemical test data into vector i2 as the input of the feature extractor. After input into the network structure, it first passes through two Linear layers, then passes through a Dropout layer, and finally passes through a Linear layer to obtain the feature vector ℎ2 as the output of this part. The network structure of this part is as follows Figure 5 shown.

[0033] The artificial neural network used in this embodiment realizes the extraction of feature information of the echocardiogram report form and biochemical examination data layer by layer through multiple hidden layers and Dropout layers. The input layer is the first layer of the neural network, and the number of neurons is equal to the number of input features. The input layer inputs the echocardiogram report form and biochemical examination data into the neural network, and passes the value of each feature to the next layer. The hidden layer receives the feature values ​​transmitted by the input layer, and the number of neurons can be arbitrarily selected, and uses weights and activation functions to calculate a new set of feature values, and passes them to the next layer.

[0034] The Dropout layer receives the new feature values ​​transmitted by the hidden layer and randomly turns off some neurons to reduce the dependence on certain features. This improves the generalization ability of the model and prevents overfitting. The output layer is the last layer of the neural network. Since the function of the neural network is to extract the features of the report form and biochemical examination data, the number of its neurons needs to meet the requirements of the feature fusion input vector, and the feature value is used as the output of this part.

[0035] In each hidden layer and output layer, the value of the neuron is calculated by the following process: ① Weighted sum: Each neuron has a set of weights that are multiplied by the input, plus a bias term. The sum of these products is the net input to the neuron.

[0036] ②Activation function: The net input is passed to an activation function to generate the output of the neuron. The activation function can be linear or nonlinear, such as sigmoid, ReLU, etc. The specific calculation method of this artificial neural network is: ; Where i2 is the biochemical examination data and echocardiography report form as the input of this part, W is the weight, b is the bias, f is the activation function, and h2 is the output feature vector.

[0037] S2.3.3) Feature Fusion and Classifier The feature fusion device takes the feature vector h1 output by the feature extractor F1 and the feature vector h2 output by the report form and biochemical test data feature extractor F2 as input, horizontally splices the two feature vectors to generate a new feature vector p, and inputs the fused feature vector into the classifier. After passing through the multi-layer perceptron, it finally outputs the prediction result y. The network structure of this part is shown in the figure. Figure 6 shown.

[0038] Feature splicing refers to a method of splicing features from different sources and different feature dimensions together to form a new feature representation. Common feature splicing methods include horizontal splicing and vertical splicing. Horizontal splicing refers to connecting features from different feature dimensions together in columns to form a new feature vector. In tasks such as image recognition and target detection, horizontal splicing is usually used to splice feature maps of different scales and positions together to obtain a more comprehensive and accurate feature representation. Vertical splicing refers to connecting features from different samples or different time periods together in rows to form a new feature matrix. In tasks such as time series prediction and speech recognition, vertical splicing is usually used to splice feature vectors of different time periods together in chronological order to obtain feature representations of longer time windows. The feature splicing method used in this embodiment is horizontal splicing. The specific calculation method is: ; Wherein, h1 is the feature vector i1 output by the feature extractor F1, h2 is the feature vector i2 output by the report form and biochemical examination data feature extractor F2, p is the fused feature, and f is the mapping relationship.

[0039] This embodiment uses a multi-layer perceptron as the classifier part of the multimodal fusion model. The input layer contains n neurons, where the size of n is equal to the feature vector p output by the feature fuser. The output layer contains three neurons for outputting the probability values ​​of the three classifications. In the multi-layer perceptron, weights and biases are parameters that need to be optimized. The weights and biases are randomly initialized by using a uniform distribution. Given the input p, the predicted value y can be calculated by forward propagation. The specific calculation process multiplies the input p by the weight W, adds the bias b, and inputs the result into the softmax function to obtain the output y. The specific calculation method is: ; Where W o is the classifier model weight, b o is the classifier model bias, y is the output result of the classifier, the softmax function can map the output value to the range of (0,1), and the sum of the probability values ​​of all outputs is 1.

[0040] S2.3.4) Loss Function and Optimizer The cross entropy loss function is a loss function used to measure the difference between the predicted probability distribution and the true probability distribution, and is usually used in classification tasks. The specific calculation method of the cross entropy loss function is: ; In the formula, N represents the total number of samples, C represents the number of categories, and y i,c Indicates that sample i belongs to the true label of category C, p i,c Represents the predicted probability that sample i belongs to category C. The cross entropy loss function makes the predicted probability distribution as close to the true probability distribution as possible by reducing the difference between the predicted probability distribution and the true probability distribution.

[0041] The SGD optimizer is a commonly used optimization algorithm, especially suitable for large-scale machine learning and deep learning. The SGD optimizer estimates the gradient of the loss function by randomly selecting a training sample in each iteration and uses the gradient to update the model parameters. The specific calculation method is: ; In the formula, θ t represents the parameter at iteration t, J(θ t ;x (i) ,y (i) ) is the loss function, x (i) and (i) is the input and output of the i-th training sample, η▽ θt J(θ t ;x (i) ,y (i) ) is the loss function with respect to the parameter θ tis the gradient of , and η is the learning rate.

[0042] In the above step S3, the classification results are saved in the cloud and transmitted bidirectionally. The classification results, i.e., the working status of the high-frequency jet working frequency and time index after the comprehensive algorithm based on the input parameters of F1 and F2, can be saved to the cloud service platform. There are two key points: First, after the high-frequency jet ventilator expands the Internet module, combined with multiple sensors, the working parameters can be uploaded to the cloud platform, and the adjustment and correction of the multimodal learning algorithm can be guided; especially when multiple high-frequency jet ventilators are connected to the cloud platform at the same time, different working conditions bring different data, forming massive big data, and the cloud platform also becomes a carrier for collecting data; Second, the working parameters of the high-frequency jet ventilator no longer need to be set by doctors based on experience. Instead, the optimal working parameters, i.e., {J, K} coordinates, are obtained from the cloud platform based on sensor values ​​(current and historical values) and previous multimodal deep learning results. The network structure of the cloud platform and multiple high-frequency jet ventilators is shown in the figure. Figure 7 shown.

Claims

1. A method for adaptively controlling the jet frequency of a high-frequency jet ventilator, characterized in that: The following steps are involved: S1) Establishing a multi-parameter real-time acquisition system: Based on a high-frequency ventilator, the acquisition parameters include but are not limited to airway carbon dioxide concentration, oxygen concentration, blood oxygen saturation, and heart rate; S2) performing two-dimensional parameter coupling of the output frequency and the output breathing time ratio; S3) extracting and preprocessing the parameters collected in step S1, screening the data to construct a multi-maternal-fetal data set, and building and training a multimodal fusion model; S4) Apply the multimodal fusion model in step S3 to form an Internet injection frequency expert system.

2. The method for adaptively controlling the jet frequency of a high-frequency jet ventilator according to claim 1, characterized in that: In the above step S2, the method for constructing the multimodal dataset is: S2.1) The output parameters of the high-frequency ventilator are classified into {J, K} coordinates for retrieval in a two-dimensional graph, where J is the highest injection frequency and K is the time proportion of the highest injection frequency in the entire injection cycle; S2.2) Training data acquisition: Using the patient's medical records as the data source, record the hospitalization information of each patient in each medical record, and extract the high-frequency ventilator parameters of these patients to construct a multimodal dataset; S2.3) The multimodal fusion model consists of four parts: the first part is the feature extractor F1, which is composed of a deep convolutional neural network, thereby realizing the detailed feature extraction of the data; the second part is the report form and biochemical examination data feature extractor F2, which is composed of an artificial neural network and realizes the feature extraction of the report form and biochemical examination information; the third part is the feature fusion device F, which is used to fuse the feature vectors obtained from the feature extractor F1 and the feature extractor F2, and use the fused feature vector as the input of the fourth part; the fourth part is the classifier C, which is a multi-layer perceptron composed of a Dropout layer and three hidden layers, and finally realizes the diagnostic classification of the output parameters.

3. The method for adaptively controlling the jet frequency of a high-frequency jet ventilator according to claim 1, characterized in that: In the above step S2.2, the data is normalized and Min-Max normalization is used to complete the data normalization work. The specific formula is: ; In the formula, x i For the current need to standardize the value, min 1≤j≤n(xj) for x1, x2, …, x n The minimum value in max 1≤j≤n(xj) for x1, x2, …, x n The maximum value in y i For x i The standardized values ​​for x1, x2, ..., x n After standardization, we get y1, y2, ..., y n And y1, y2, …, y n ∈[0,1].

4. The method for adaptively controlling the jet frequency of a high-frequency jet ventilator according to claim 3, characterized in that: In the above step S2.2, the random forest model is used to calculate the information gain of each attribute of the data. First, the information entropy is calculated: ; In the formula, p(c) represents the proportion of samples of category c in the data set; Then calculate the conditional entropy of data set D under feature F: ; In the formula, |D i | represents subset D i , |D| represents the number of samples in the data set D; Finally, the information gain of feature F is calculated, which indicates the degree to which the uncertainty of data set D is reduced after splitting using feature F. The information gain is calculated as follows: ; In the formula, Informationgain (D, F) represents the information gain of feature F in data set D, entropy(D) represents the initial information entropy of the data set D. conditionalentropy(D,F) represents the conditional entropy of data set D under feature F.

5. The method for adaptively controlling the jet frequency of a high-frequency jet ventilator according to claim 1, characterized in that: The above step S2.3 includes the following steps: S2.3.1) Feature extractor F1 The data matrix i1 is taken as input, the feature vector h1 after extracting the data features is taken as the output of this part, and the feature vector h1 is taken as the input of the feature fusion device F; First, the input feature vector h1 passes through a normal convolution layer, then an activation function layer, and finally a batch normalization layer to obtain the feature map i′: ; In the formula, h is the output channel of the convolutional layer, cin is the input channel of the convolutional layer, p is the kernel size and step size, BN is the batch normalization layer, and the specific calculation method is: ; In the formula, û B is the sample mean, σ B is the sample standard deviation of the mini-batch B, γ is the stretch parameter, β is the offset parameter, and the shapes of γ and β are the same as λ; The feature map i′1 after the batch normalization layer is input into N ConvMixer blocks consisting of depthwise convolutional layers and pointwise convolutional layers. The specific calculation method is: ; ; In the formula, ConvDepthwise() is the depth convolution layer, ConvDepthwise() is the point-by-point convolution layer, and the specific calculation methods are: ; In the formula, O k,i,j Represents the value of the kth channel of the output feature map O at position (i, j), I c,i+p-1,j+p-1 represents the value of the cth channel of the input feature map I at position (i+p-1,j+p-1), W k,c,p,q represents the weight of the cth channel of the kth convolution kernel at position (p, q), b k Represents the bias term of the kth convolution kernel; ; In the formula, ω i,j represents the weight value between the i-th output channel and the j-th input channel of the convolution kernel, x j Represents the value of the input tensor on the jth channel, y i Represents the value of the output tensor on the i-th channel; The feature map i after N ConvMixer blocks is l+1 Input to the adaptive average pooling layer and a Dropout layer to output the feature vector. The specific calculation method is: ; In the formula, i` l+1 As the feature map output by the ConvMixer block and as the input of the Dropout layer, the probability of the Dropout operation is p, and h1 is the feature vector of the output of this part.

6. The method for adaptively controlling the jet frequency of a high-frequency jet ventilator according to claim 5, characterized in that: The above step S2.3 further includes the following steps: S2.3.2) Feature extractor F2 The report form and biochemical test data are converted into vector i2 as the input of the feature extractor. After being input into the network structure, it first passes through two Linear layers, then a Dropout layer, and finally a Linear layer to obtain the feature vector h2 as the output of this part; the value of the neuron is calculated through the following process: ① Weighted sum: Each neuron has a set of weights multiplied by the input, plus a bias term; Activation function: The net input is passed into an activation function to generate the output of the neuron; The specific calculation method of the artificial neural network is: ; Where i2 is the biochemical examination data and echocardiography report form as the input of this part, W is the weight, b is the bias, f is the activation function, and h2 is the output feature vector.

7. The method for adaptively controlling the jet frequency of a high-frequency jet ventilator according to claim 6, characterized in that: The above step S2.3 further includes the following steps: S2.3.3) Feature Fusion and Classifier The feature fusion device takes the feature vector h1 output by the feature extractor F1 and the feature vector h2 output by the report form and biochemical test data feature extractor F2 as input, horizontally concatenates the two feature vectors to generate a new feature vector p, and inputs the fused feature vector into the classifier. After passing through the multi-layer perceptron, the prediction result y is finally output; The horizontal stitching method is used for feature stitching. The specific calculation method is: ; Where, h1 is the feature vector i1 output by the feature extractor F1, h2 is the feature vector i2 output by the report form and biochemical examination data feature extractor F2, p is the fused feature, and f is the mapping relationship; The input layer contains n neurons, where the size of n is equal to the feature vector p output by the feature fuser. Given the input p, the predicted value y can be calculated through forward propagation by using uniformly distributed random initialization weights and biases. The specific calculation method is: ; Where W o is the classifier model weight, b o is the classifier model bias, y is the output result of the classifier, the softmax function can map the output value to the range of (0,1), and the sum of the probability values ​​of all outputs is 1.

8. The method for adaptively controlling the jet frequency of a high-frequency jet ventilator according to claim 7, characterized in that: The above step S2.3 further includes the following steps: S2.3.4) Loss Function and Optimizer The specific calculation method of the cross entropy loss function is: ; In the formula, N represents the total number of samples, C represents the number of categories, and y i,c Indicates that sample i belongs to the true label of category C, p i,c represents the predicted probability that sample i belongs to category C; The optimizer estimates the gradient of the loss function by randomly selecting a training sample in each iteration and uses the gradient to update the model parameters. The specific calculation method is: ; In the formula, θ t represents the parameter at iteration t, J(θ t ;x (i) ,y (i) ) is the loss function, x (i) and (i) is the input and output of the i-th training sample, η▽ θt J(θ t ;x (i) ,y (i) ) is the loss function with respect to the parameter θ t is the gradient of , and η is the learning rate.

9. The method for adaptively controlling the jet frequency of a high-frequency jet ventilator according to claim 2, characterized in that: In the above step S2.2, during the data extraction process, when there are multiple examinations for the same patient, the first diagnosis of each patient in time sequence is selected to construct a data set.

Citation Information

Cited By

  • Intelligent control method and system for intermittent jet ventilation nasal oxygen inhalation catheter

    CN120771403A

  • Intelligent control method and system for intermittent jet ventilation oxygen inhalation catheter

    CN120771403B